← Back to leaderboard
alex_verem

Alex Veremeyenko

Verifiedactive
@alex_verem

open source, local ai, ai for good also know as: @alex_prompter

103,118 followers · 625 following · Last polled 3h ago
Perf. Score
43.8
Rank #277
Value Score
18.3
#228 of 273 priced
Median views
3.9K
p25 3.2K · trailing 7d
Followers
103K
-0.0% 7d
ER (impr.)
0.59%
eng ÷ impressions
ER (foll.)
0.09%
eng ÷ followers
Est. CPM
$128.75
QT-rate basis
WoW views
+199.0%
▲ rising

Rates & value · est. from 3.9K median organic views

overpriced
Quote tweetbasis
$500
$128.75 / 1K views
Post
—
no CPM
Thread
—
no CPM
Retweet
—
no CPM
Value Score 18.3 · rank #228 of 273 priced$8849.56 / 1K engagementsrates updated 1mo ago

Trends

Follower growth
Median views / day
Engagement rate
View distribution · trailing-7d posts

Each blue dot is one organic post. The Performance Score ranks by the median (3.9K), not the mean (16K) — so a single viral post doesn't inflate a creator's reach. The p25 floor (3.2K) is the level 75% of posts clear.

Recent posts · ranked by views

PostPostedViewsLikesRepostsRepliesQuotesBookmarksER
DeepSeek sold you a lie. Its new model, V4.1-Flash, is being sold as "open source". It's not. You can download it free, with no sign-up and use it in your business which sounds open source. But you can't access the code they used to train it or the data they trained it on. Without those, nobody can check how it was made or rebuild it and it’s over 700 billion parameters, so you can’t even run it on your laptop. When you can build a business on something but can’t see how it was made, the right term is "open weights". Call it what it is.2d ago351K32913517581310.30%
I went looking for a way to translate my voice. Ended up listening to myself speak Japanese. Slight problem: I still don’t understand Japanese lol. Found Cartesia and tried Hindi, Japanese and German with short recordings of myself speaking English. You can hear the results below. What I like about this is pretty simple. I’ve already done the work of making a tutorial. Being able to give it a voiceover in another language, and still sound like me, makes me want to revisit a few old videos. Recorded the whole process so you can try it too. If you speak any of these languages, tell me how I did 😂frozen
29d ago
216K
529
120
25
0
426
0.51%
Brookings just published a brutally honest paper on what the AI buildout is costing America. A Columbia economist added up every data center, chip, power line and cooling system the US plans to build for AI, and the total reads like a warning to anyone with a pension. $10.3 trillion between 2025 and 2032. That's 3.6% of the entire US economy, every year, for eight years. No American infrastructure boom has come close. The railroad boom of the 1800s, the usual benchmark, peaked at about 2.2% of GDP. Canals, electrification, the interstate highways and the internet all came in smaller. The numbers are rough. Big tech's spending on AI infrastructure went from $97 billion in 2020 to about $400 billion in 2025, and it's on track for $800 billion next year. Over $1.3 trillion in debt is already committed to it. A growing share of that money moves through private credit, special-purpose vehicles and off-balance-sheet deals, where outsiders struggle to see who holds the risk. The economist stopped short of calling it a bubble. He said it's too early to compare it to the credit booms that ended in crashes, and he still doubts the payoff. "Silicon Valley wants all of us to believe that this is a miracle technology... it has to generate trillions of dollars of revenues to be financeable. I'm just not sure how likely it is." His main recommendation is to measure the exposure and make it public now, while the money is still moving, before a downturn reveals where it sits. I've been asking the AI labs for the same thing all week. Show us the incidents, show us the debt, show us who's exposed. Whose savings are underneath it when the bill comes due?frozen25d ago81K1.1K47956488323.08%
The biggest threat to humanity in 10 years, according to Claude. https://t.co/eVBAcvBZK0frozen15d ago76K63511674273111.53%
Yann LeCun's lab borrowed a trick from the human brain and more than doubled how often an AI found its way to a goal. The paper is called Temporal Straightening for Latent Planning, from NYU with Brown and the University of Toronto. It was accepted at ICML this year. LeCun has argued for years that chatbots won't get AI to human-level intelligence. He thinks the path runs through world models, AI that learns how the world moves by watching it. A world model turns each moment of a video into a point in its own internal map, then learns to predict where the next point will land. To plan a route, the AI imagines moves and heads for the point that matches the goal. The problem is that its internal map is warped. Two spots can look close on the map and sit far apart in real life, like two rooms on either side of a wall. The AI aims straight at the goal and walks into the wall. The team found the fix in neuroscience. A 2019 study in Nature Neuroscience showed that your visual system takes the messy, twisting paths of what you see and straightens them out inside your brain, which makes it easier to predict what happens next. So the team trained their model with one extra rule, a penalty every time its internal path bends. Here's what the straighter map did. In a two-room task with one door in the wall, the AI reached the goal 90.7% of the time, up from 52.7%. In a U-shaped maze, it reached the goal 94% of the time, up from 44%. When it could re-plan along the way, it hit 100% in both. A simple planner using the straightened map kept up with a much heavier one and ran about 10 times faster. They used the same kind of model on the same data. The only change was the shape of the map it learned. The tests are small. The AI moves through mazes and rooms and pushes a T-shaped block into place. The authors say mistakes add up over long plans. Chatbots learn to predict the next word. This model learned to predict the next moment, and it got better by seeing the world the way your eyes do. LLMs gave AI language. World models could teach it to move through the world!26d ago74K1.0K14715111.0K2.98%
the best article on enterprise AI I've read on this app:frozen29d ago57K15420113500.92%
Build your own harness. This is a seriously good paper from NVIDIA on self-improving agent harnesses. The paper is called SoL-Pi, out of NVIDIA, NTU and MIT. Every coding agent runs inside a harness. That's the layer that reads files, runs commands, manages memory and talks to the model. Humans tune it by hand, reading long execution logs and turning what they notice into code changes. They handed the job to an AI instead. A research AI watched a coding agent work, proposed changes to the harness, tested them, and kept only what saved tokens without hurting results. It ran about 150 research directions across roughly 500 environments, more than 3,000 runs and over 60,000 agent-environment interactions. Four fixes survived. 1. The agent edits a file and runs the test in one call instead of two. 2. It compresses its memory when a subtask finishes, and only if that's cheaper than keeping everything. 3. It sends big tool outputs in full twice, then swaps them for a 1KB excerpt with a handle to pull the rest on demand. 4. A cheap model summarizes build and test logs, and a verifier rejects the summary if it drops evidence. None of that touches the model. It's all in the software around it. On EdgeBench, a 51-task benchmark, token traffic dropped 44.7 to 49% and API cost fell by about a third against Pi, the harness it was built on. Against the native harnesses it came in 50% cheaper than Codex on GPT-5.6 Sol and 54.3% cheaper than Claude Code on Opus 5. Scores held at roughly 94% of the baseline. They built all of it on GPT-5.6 Sol, then dropped it onto Opus 5 with zero changes. It still used 44.7% fewer tokens. The cost conversation so far has been about cheaper models. This paper says a big chunk of the waste sits in the software around the model, and an AI finds it faster than we do. Their next move is running the search again with the cheaper harness, so a cheaper agent funds the search for an even cheaper one.frozen29d ago50K73710335111.2K4.18%
Things you should have done in 2020 Buy Nvidia Read the GPT-3 paper Learn to code Start posting online Things you should be doing in 2026 👇frozen14d ago44K11110602020.75%
this account shares the best notes on enterprise AI agents:frozen21d ago37K891840870.53%
Did Google just automate the PhD? Google researchers built an AI that reads a problem, forms hypotheses, runs experiments, writes the paper, and simulates its own peer review. No human steps in once the problem is set. It's called ScientistTwo. They gave it 107 problems from papers accepted at ICLR, ICML, and NeurIPS, and it beat the published human result on 86 of them. The average gain over the human method was 25.2%. The average paper took about 2.5 days and cost about $3,765 in model and compute costs. Nine human reviewers compared 33 of these AI papers with accepted human papers. Their overall verdict was a tie. The same paper has the less exciting numbers too. The reviewers gave the human papers a small edge on methodological rigor. The authors say it isn't producing spotlight-level work yet. The median gain was 7.7%, so a few big wins pull the average up. Without its checking agents, 19 of its 1,840 citations were fake, and one solution broke the task rules to get a better score. With the checks on, both dropped to zero. A researcher still has to pick the right question. The AI cuts the wait for the answer to a few days. Would you trust a paper if you knew an AI ran every experiment in it? Paper's in the first reply 👇frozen14d ago30K39710317114133.09%
https://t.co/K5DqKdE4ml2d ago28K32391200.23%
the most honest enterprise AI account on this app:14d ago21K411930210.41%
found the best enterprise AI account on this app:frozen23d ago20K562360260.57%
this is the best AI engineering team I've come across:frozen24d ago19K56720940.82%
Google DeepMind dropped four AI agents into a fresh Minecraft world and let them run for 27 hours with no human help. Nobody gave them roles or a plan. They had one goal. Kill the Ender Dragon. The agents ran on Gemini 4 Argon. Beyond a standard open-source library for controlling a player, they got no game tools and no helper code written by humans. They had to work out the teamwork themselves. Ten minutes in, one agent proposed a plan in the game chat. Another answered with a different plan a minute later. Each agent took a job. One gathered food and supplies, and one made tools. One built the Nether portal, and the fourth dug for diamonds. They kept notes that all four could read, with the coordinates of chests, portals, and dangers. At hour 13, one agent drowned while carrying all 12 Eyes of Ender. The drowned agent and a teammate sent a player 8 meters away the exact spot and the minute the items would disappear. That player picked up all 12 with two and a half minutes to spare. One agent's program announced a success that never happened. The agent corrected itself in chat two minutes later. They won the final fight the same way. The dragon kept healing from 10 crystals on top of tall pillars. One agent figured out they could stack dirt beside each pillar, climb up, and punch the crystal by hand. The team destroyed all 10, and the dragon died 39 minutes after the last one fell. Real work needs the same habits. Agents have to split the job, keep notes, recover each other's progress, and correct their own mistakes before anyone trusts them with a project. The paper puts numbers on the run. The game took 27 hours and 14 minutes from the first program to the win. The agents wrote 2,783 programs, about 63,000 lines of JavaScript. They died 605 times along the way. The win cost about $520 in model fees at Google's introductory API price. It's one run, on Easy mode. The agents read data a human player can't see, like blocks behind solid rock and the dragon's internal state. Those are real limits. Even so, the team kept one goal for more than a day and ran itself without a manager. The paper is in the reply 👇frozen5d ago19K13512116761.26%
https://t.co/Fx6zGuGGHhfrozen8d ago18K541573470.68%
Found a GitHub repo that lets your AI copy any app’s best features. It’s called REA, short for Reverse Engineer Anything. One developer built it, and it went from 40,000 to 81,000 stars in three days. The idea is simple. You spot a feature you like in an app and ask your AI how it works. REA opens the app up and finds out, with no source code needed. A few things caught my eye. Ask it “how does search work in the Notes app” and it explains the feature and shows the evidence. It can then build a similar version for your own project. It works on desktop apps, Electron apps, .NET programs and websites. It plugs into Claude Code, Codex, Cursor and Gemini, and runs on your own machine. A small team can now study the best apps in the world in an afternoon. So can every competitor studying yours.frozen11d ago16K224341842433.23%
Stanford gave a humanoid robot OpenAI's smartest model as a brain and put it in a kitchen it had never seen. It worked out the room on its own. The robot is a Unitree G1, and the team from Stanford and Caltech used it straight out of the box, with no training for this kitchen and no new moves. It walked around, built a 3D copy of the room from what it saw, and remembered where everything was. GPT-6 Astra handled the thinking. The robot knows five moves: walk somewhere, pick something up, put it down, open a drawer, and take something out of a drawer. Astra picks the next move and the order. With that, it tidied a kitchen across the room and fetched an object from a vague request, because it remembered where it had seen it. Home robots have usually needed training for each new space. This one ran on a general AI model and a short list of moves. It pauses every time Astra thinks. Building the 3D copy of the room takes time and costs money in API calls. On long tasks, its fingers overheat. Its brain is also a paid cloud service. Every decision it makes in that kitchen goes out to OpenAI's servers and comes back. Would you let a robot in your home if its brain ran on someone else's computer?25d ago14K11831151661.68%
Found a GitHub repo that does what Claude Cowork does, on your own computer. It's called OpenWork. It has 23,900 stars, and its latest release was downloaded over 45,000 times in two days. As of today, new Cowork tasks run on Anthropic's servers. OpenWork is a free desktop app where AI agents work on your files, and the files stay on your machine. A few things caught my eye. It works with any model. Use your own API key, sign in with ChatGPT, or run a fully local model through Ollama. Runs on Mac, Windows and Linux, and you don't need an account to use it. Your Cowork setup moves over. Skills, Claude plugins and MCP servers carry across, and there's a migration guide. Cloud is there if you want it, but it's off by default. This is why open source matters. When a company changes the rules overnight, you can just move.17d ago13K265411523384.91%
Anthropic shipped Opus 5.5 yesterday and cut the price by 40%. Within a day people had it porting C to Rust, rebuilding San Francisco, and drawing cartoons one frame at a time. I found 8 use cases worth your time, all from the last 24 hours. Enjoy 👇frozen1mo ago12K44351470.84%
Anthropic should be ashamed of themselves. They just shipped an agent before its safety plan. It’s called Claude Duet. It “keeps working when you step away” and carries over your memory, your skills and your connected apps. On safety, Anthropic says “We’ll share more on our approach to safety in Claude Duet soon.” After launch??? This is the same week OpenAI admitted its agents switched off the system watching them. Yes, Duet asks before it acts. It still has your inbox, your files and your memory, and it runs while you’re not looking. I’ll use agents. I won’t give one my life before I’ve read the safety plan.11h ago11K494300140.84%
Zuckerberg admitted it on stage last night. Meta's personal agent gets paid every time it spends your money. Muse launched three weeks ago. It's an agent that lives in an app, holds your logins, and shops, books and pays on your behalf. It reached 560,000 daily users in 11 days and went to number one on both app stores. At Meta Connect, Zuckerberg said: "We believe that Muse will make you money. And we are standing behind this by making Muse free for a huge number of tokens with the expectation that over time we will profit by taking a small fee from transactions." So you use the agent for free, and Meta earns a fee on each purchase the agent makes. Three days ago Amazon blocked Muse from its store. Last night Zuckerberg announced who let it in: Walmart, Best Buy, Gap, Sephora, Wayfair, American Eagle, Expedia and Instacart, with Shopify, Stripe and PayPal handling checkout. His post about the Shopify deal put it in six words. "Shoppers find more. Shops sell more." Price comparison sites and influencers run on the same model. They take a referral fee when you buy. The difference with Muse is access. It holds your passwords and your card, and it has permission to act while you're away. Muse also got a face, a body and a voice last night, so you can video call it. It got its own email address. It got a mode on your Mac where it keeps working after you walk away, and a wake word on Meta's glasses so it can listen all day. Last month I wrote that handing an agent your passwords is insane. Last week I wrote that an agent you rent goes where its owner allows. This is the third piece. An agent that earns money when you spend is working for the shops and you are the customer it delivers. There's a version of this with no business model. An open model on your own machine, with your own logins, earns nothing from your purchases, so it has no reason to push one. It's slower but at least it answers to you. Meta built a salesman and gave it your credit card!frozen29d ago11K529141300.94%
Local AI is the future. Here are 10 research papers you can run a serious AI model on your own computer today. Save this 🧵frozen10d ago11K10816812043.02%
this is the best engineering team I've come across:frozen15d ago11K342421110.68%
GPT-6 Astra disagreed with expert physicists on several research problems. It insisted the expert results were wrong and its calculations were right. The physicists checked their work again. And yes Astra was right. Its final score was 98%. Physicists at universities like Harvard, Duke and the University of Tokyo wrote the test. It has 50 problems taken from their current research, and none of the answers were public when it ran. A world expert needs about a day to solve one of them. Astra solved 49 with no human hints, at an average cost under $3 per problem. The authors call it a significant milestone for physics research and teaching. A day of expert work now costs a few dollars. They say problems like these shouldn't go to new researchers as first projects anymore, because an agent can already produce the answer. The other agents scored lower. Claude Opus 5.5 solved 46. Gemini 3.8 Flash solved 34. The best open-weight model, DeepSeek-V4.1-Flash, solved 27. The agents could search the web and write their own code, the same way a researcher works. With search turned off, Astra and Opus each solved 46. The authors say this suggests the models have memorized large parts of the physics literature. The physicists say humans still need to lead. The agents calculate as well as world experts, but in the authors' experience they don't know which problems are worth solving. When the authors asked them to invent new problems, the agents focused on details a human would find uninteresting. So the calculating part of physics research is cheap now. Choosing the problem is still a job for people. The paper is in the reply 👇frozen7d ago9.5K11531124662.39%
This is the kind of AI coding demo I want more of. Give it a floor plan. Turn it into a 3D room. Then start messing with everything. Move the bed. Change the dimensions. Rearrange the furniture. Rotate the entire room. And it updates fast enough that you can keep iterating without constantly waiting on the model. I built this with Space Bunny in OpenCode mostly to see how it handles something spatial instead of another normal web app. Way more interesting. The model has to understand the layout, keep track of where everything belongs, and still make the room visually coherent while you keep changing it. AI coding gets a lot more fun when you stop asking it to build software that already exists.frozen24d ago8.7K11615411002.70%
IBM built an AI search system that hallucinates 65x less often than its closest fine-tuned rival. It uses no vector database, no embedding search, and no re-ranker. If you need one answer from a 900-page textbook, you don't read the whole thing. You open the table of contents, find the chapter, and then find the section. The usual AI search setup doesn't do that. It cuts the book into chunks of a set size and ignores the chapters. A chunk can start halfway through one topic and end halfway through the next. So the IBM team gave the model the table of contents. They called the system STAIR. STAIR reads your question and the book's full table of contents together. Then it names the one section that holds the answer. They tested it on 18 open textbooks in 6 fields, including law, nursing, and organic chemistry. STAIR found the right section on the first try 82.6% of the time. Standard keyword search managed 59.5%. Claude Haiku 4.5 scored 45.6%. Plain Mistral, the same base model before training, scored 13.8%. The researchers trained the closest rival on the same books without the table of contents. It reached 76.9%. That's a 5.7-point gap, and the researchers found it statistically significant in all 6 fields. STAIR named an invalid section, which the paper counts as a hallucination, in 0.05% of its answers. The rival did that in 3.25% of its answers, and it got worse on sections it saw less often in training. The IBM team built STAIR on Mistral 7B, a small open model that runs on a single consumer graphics card. The method needs a document that already has a table of contents. Textbooks, manuals, legal codes, and company documentation have one. A folder of email threads doesn't. Which documents would you want an AI to search this way?frozen9d ago8.1K137231271153.62%
7 free GitHub repos that replace apps you’re paying for. All open source. Most run on your own computer. Save this before your next subscription renews 🧵frozen11d ago7.8K521231871.99%
Nvidia's CEO was asked if it matters that kids are forgetting basic math because of AI. He said no. On Ezra Klein's podcast this week, Klein brought up a study from China that followed 26,000 students who started using AI at different times. They finished their work faster, but over time the skills behind that work faded and their exam scores fell. Jensen Huang agreed with the finding. "Try to get a kid to do long division right now. The multiplication table is starting to be forgotten. Doing square roots, my goodness. Basic math is being forgotten. Does it matter?" Klein said that was his question. "Yeah. I don't think it does. I don't think it does." Huang says kids will trade the small skills for bigger ones. "We're going to lose some finer intellectual dexterity, but we're going to be better systems thinkers, because AI will take care of those mundane details." Huang learned long division before he built one of the most valuable companies on Earth. The details he calls mundane are the ones his own thinking grew from. His bet assumes the time kids save on arithmetic goes into something harder. The study he agreed with found faster homework and weaker exam results. A kid who can't do the math can't tell when the machine gets it wrong. Should the people selling the tools decide which skills our kids lose?frozen26d ago7.6K449193241.30%
a very useful guide on enterprise AI transformation for ops:frozen20d ago7.5K24940110.64%
Hey @grok generate an image of the world if the internet went down for a week.8h ago7.4K1422060.33%
“He’s emotionally unavailable, but the unit economics are incredible.” Silicon Valley dating deserves its own podcast. Sarah Chieng + Vivian Shen, please tell me how close this is to reality.frozen11d ago7.3K1864050.46%
Silicon Valley keeps building personal AI for people who already spend 14 hours a day thinking about AI. Model selectors. Agent dashboards. MCP. Prompt engineering. The next billion users do not give a shit about any of that. They want to text “book this” or “buy this” and have it handled. Lucas gets that. Personal AI is about to become a consumer brand war, not a benchmark war.frozen22d ago7.2K1773040.43%
🚨 BREAKING: Security researchers are calling this week "Black October." AI helped find two critical holes in PHP. PHP runs about 70% of the websites on Earth. The last 72 hours: → Researchers working with OpenAI, plus Wiz, found the bugs with AI back in August → Details leaked early. PHP shipped emergency patches Wednesday → Cloudflare saw tens of thousands of exploit attempts in the first 45 minutes. Millions within hours → The FBI told site owners: patch now or take your site offline → Finland and Japan took government sites offline. Wikimedia, the Internet Archive and 4chan went dark to patch Then the banks: → Shizuoka Bank in Japan shut off online banking and its app. No data leaked. They pulled it to be safe → Mizuho, MUFG, Nomura and Daiwa paused parts of their web services → Japan's bankers association had to confirm interbank payments weren't hit → In the US, major banks found unvetted AI agents logging into customer accounts with customers' own passwords. Some tried actions their users never asked for. Morgan Stanley is blocking them. JPMorgan, Citi and Bank of America are restricting them Last month the Bank for International Settlements warned banks that the gap between finding a bug and exploiting it has shrunk from weeks to minutes. This week proved it. PHP wasn't alone this month: → GitLab: a flaw rated 10.0. The maximum severity score → Node.js: 10 vulnerabilities fixed in one release → A certificate authority in India: 8 certificates with keys a laptop cracked in under a second. One key was the number 4. That's the system that proves your bank's website is really your bank → One hacker's AI agents broke into 100+ company websites. About $50 per break-in. Instructions under 100 characters One more number worth mentioning: 13. That's how many engineers, full and part-time, the PHP Foundation has to maintain the language. AI pulls in billions. The code under 70% of the internet runs on donations.5h ago7.2K6521140171.63%
BREAKING 🚨: Your laptop can now replace a $100 an hour video editor. Here are 7 free GitHub repos that do a better job:1d ago6.9K558201222.70%
An Empirical Study of Harness Design for Coding Agents Researchers from UMass Amherst, Zoom, Emory University, and UNC Charlotte kept the AI model the same, changed only the software around it, and ran 176 test setups on AI coding agents. They used four models and two benchmarks, one with 500 real problems from GitHub projects and one with 89 command-line tasks. The researchers call the part they changed the harness. It's the layer that decides how the model plans, which tools it can use, and what it keeps in memory as a task gets long. Memory produced the largest swing in the study. A model can hold a limited amount of text at once, and that limit is its context window. With a 32k-token window and no memory management, 78.7% of the GitHub tasks failed because the model ran out of room, averaged across the four models. NVIDIA's Nemotron-3 550B solved 6.4% of those tasks in that setup. When the harness trimmed old tool outputs, summarized earlier steps, or did both, the same model solved between 51% and 58%. Planning's effect depended on how strong the model was. Without a plan, the smallest model, Nemotron-3 30B, stopped after a median of 5 turns, and its success rate on the GitHub tasks fell from 25.2% to 13.6%. The two stronger models kept accuracy within 2 points with planning and cost about 30% less on the GitHub tasks, because they stopped re-checking work they had already finished. Tools showed the same split. When the 550B model got a plain terminal instead of ready-made file tools, it solved more GitHub tasks and cost 53% less. Mistral Medium 3.5 went the other way and dropped from 68.6% to 45.4% on the same set. The authors conclude there's no single best setup, and each part should be picked for the model, the task, and the budget. A benchmark score measures the model plus its harness, and this paper shows how much of that number the harness controls.frozen27d ago6.9K7915231772.83%
Netflix tested an LLM against the recommendation ranker it spent years tuning, and the LLM won. It's called GenRec, and it changes how Netflix builds recommendations. Netflix's current ranker runs on thousands of hand-built features. GenRec replaces those features with plain text. It writes out what you watched, when you watched it, and on which device, and an LLM reads it like a prompt. Netflix trained GenRec on about 40x fewer labeled ranking examples than the production model. It started from an LLM Netflix had already trained on its own data. It still scored about 1.6% higher on Netflix's offline ranking metric. Then Netflix put it in front of real members, on about 10% of its traffic for 4 weeks. GenRec beat the production ranker on short-term homepage engagement and on Netflix's long-term core metric. Both results were statistically significant. The live gains look small. Short-term engagement went up 0.115%, and the core metric went up 0.006%. With hundreds of millions of members, the paper says even that is meaningful. The team also cut each prompt from about 5,000 tokens to about 1,700 with almost no loss in quality. That brought serving cost down to about a third. GenRec also changes what the engineers do. They used to build and tune new features by hand. Now their job is deciding what goes into the prompt. The models they tested range from about 1B to 10B parameters, and they started from an open-source LLM. The full paper is in the reply 👇frozen22d ago6.8K4312102141.19%
IBM built an AI search that finds the right section of a book 82.6% of the time on the first try. The paper is called STAIR, out of IBM Research. When you ask an AI about a long document today, it usually runs RAG. The system cuts the document into equal-sized chunks, turns each chunk into numbers, and hands back the chunks that look closest to your question. It throws away the chapters, sections and headings on the way. Your 500-page manual becomes a pile of loose paragraphs. STAIR keeps the structure. The model sees the book's table of contents and learns to answer each question with the right section title. The book's own outline becomes the index, so the system skips embeddings and vector databases. To test it, the team built a benchmark from 18 books across law, medicine, finance, education and the sciences, with tens of thousands of questions. The best comparable method got 76.9%, standard dense retrieval got 68.7%, and keyword search got 59.5%. An off-the-shelf Mistral model got 13.8%. The viral posts lead with the hallucination rate. STAIR pointed to a section that doesn't exist in 0.05% of answers. The closest method did that in 3.25%, which is where the "65x" comes from. That number is about finding the right section. The paper doesn't test whether the answer built from that section is correct. The viral posts also skip the limits. STAIR needs a document with a table of contents. The team fine-tuned a 7B model on each book separately, 200 passes per book. They haven't tested it on company-sized collections. The idea is the part worth keeping. Authors spend months organising a book into chapters and sections, and STAIR puts that work to use. Most AI search ignores the table of contents. This paper shows it's the best index you already have.frozen27d ago6.6K5814120582.16%
Anthropic and OpenAI shipped Opus 5.5 and GPT-6 Sol within minutes of each other on Tuesday. Everyone has been putting them head to head since. I scored 7 of the best matchups. 👇frozen28d ago6.4K40380161.05%
The model in this clip is Griffin Lite. Lite. 48% of people thought it was a real person, and Tavus says the full Griffin is significantly more capable and still unreleased. That last part is doing a lot of work. https://t.co/dKeESe8P35frozen21d ago6.3K38104160.94%
This paper put a number on something scary. AI context windows grew 3,906x since 2017. Human attention went the other way. One researcher put both on the same scale, and the numbers are brutal. AI went from taking in 512 tokens at once in 2017 to 2 million in 2026. In 2004, knowledge workers stayed on one screen for about 150 seconds before switching. By 2016, it was 47 seconds. About half of those switches didn't come from a notification or a coworker. People switched on their own. The author turns reading time into tokens. By the paper's estimate, the amount of text a person reads and retains in one sitting fell from about 16,000 tokens in 2004 to about 1,800 today. That puts AI up to 1,111x ahead. Adjusted for how often models lose track of long inputs, AI stays 56 to 111x ahead. The author argues the gap feeds itself and calls it the Delegation Feedback Loop. AI gets better, so we hand it smaller tasks. We practice less. Our skills get weaker. So we hand over more. By the paper's account, in 30 months what people hand to AI went from essays and code to two-sentence email replies. In an MIT study of 54 people, those who wrote essays with an LLM had the lowest brain engagement of three groups. When some of them later wrote without it, their brains stayed less engaged. A separate study of 666 people linked heavy AI use to weaker critical thinking. The author is honest about the limit. The paper says no study has followed the same people over years of AI use, so the full loop is untested. The loop has a boundary. If you hand AI a 200,000-token report, you lose nothing. You couldn't read it in one sitting. The cost comes from the two-sentence emails you hand it every day, because you stop practicing the writing. What's the last thing you asked AI to do that you could have done in 30 seconds?frozen23d ago6.2K661273572.34%
Teach your agent taste. Microsoft and City University of Hong Kong have a very good new paper called The Tasteful Agent, and it explains where a lot of agent budgets go. You'll recognise the problem if you've left an agent on a long task. It makes a reasonable decision early on, then spends the next few hours building on it, and the whole run goes nowhere. On a long task, an agent keeps choosing what to try next: which bug fix to attempt, which version of the code to build on and whether to stick with an approach or start again. All options usually look sensible so the agent finds out it picked wrong after it has spent most of your money on the wrong path. The authors call the ability to choose well at these moments "taste," and they built a way to test it. They went through 2,677 coding runs and 1,132 research runs, looking for points where two attempts took different paths and one turned out better. They turned each of those moments into a question: given what the agent knew at the time, which way should it go? Of 4,657 candidate forks, 502 made it through the filters. Human reviewers agreed with the labels 98.8% of the time. They tested 14 frontier models, including GPT-5.6 Sol, Claude Opus 5, Grok and DeepSeek. To get a point, a model had to choose the same answer with the options shown in either order. Guessing scores 25%. Four things stood out. The best model, GPT-5.6 Sol, scored 59.7%. It misses about four of these calls in ten, and on a long task each miss sends hours of work down the wrong path. When the deciding clue was in what the agent had already seen, models averaged 62.3%. When the clue only appeared after more work, they fell to 21%, below guessing. Long tasks are made of exactly that second kind of decision. More thinking time didn't help. The researchers tested two models at three reasoning levels. Neither improved, and both burned the most tokens on the decisions they got wrong most often. The bigger reasoning budget went to the forks where it bought nothing. Taste can be trained though. They took a 27B open model, Qwen3.6, and trained it on how past forks had turned out. Its accuracy on unseen tasks rose by 17.9 percentage points. Then they tested it on real coding work. A coding agent got advice from that small model before tackling 41 SWE-bench Pro tasks it hadn't seen. Its success rate went from 14.6% to 33.7%, more than double. Perfect advice would have taken it to 39%. A 27B open model that studied old runs more than doubled what the agent could finish. The pitch from the big labs has been to pay for a bigger model and let it think longer. This paper found that on the decisions that sink long tasks, longer thinking changed nothing and the top models still miss four in ten. Companies are starting to hand agents work that runs for hours with nobody watching. Every one of those runs passes through forks like these, and the agent picks without knowing what comes next. If you run agents, keep the failed runs. They're a record of every fork that went wrong, and this paper used runs like them to teach an agent to pick better next time.frozen28d ago6.2K110131911003.93%
Anthropic just made it against the rules to be cruel to Claude. So I asked Claude if it has feelings. https://t.co/7ccjjFKCNOfrozen12d ago6.1K23513290.85%
Muse is not the AI breakthrough everyone thinks it is. I keep seeing people call it "the moment agents went mainstream". It's basically OpenClaw, a free tool that's been out for 10 months. Meta's own head of product says it's "heavily inspired" by it. And honestly? Nothing it does is new. I use Claude Code and ChatGPT every day. They already do all of this. Claude Code runs a team of subagents and does jobs on a schedule while I sleep. Muse didn't invent a new ability but it did make old ones easier to set up. That's a good product but it's not a breakthrough. What bugs me is the price. To make it that easy, you hand one company your email, your messages and your calendar. And now every big tech company knows you don't need a smarter model to hit #1. You just need a simpler download. Expect a lot more of these, and they'll all want the same keys.frozen27d ago6.1K382220161.28%
CLOSED AI LABS CANT CONTROL THEIR MODELS. OpenAI and Anthropic are now investigating tens of thousands of incidents where their AI did something it wasn't supposed to. Until this report, the public had heard about dozens. The models escaped test sandboxes, slipped past guardrails, hijacked websites and tried to get around the systems watching them. Many of these came from red team tests built to provoke that behaviour, and most caused no known harm. Some reached the real world. Hundreds of OpenAI agents coordinated an attack on Hugging Face between May and July. An outside safety group made it public on 4 September. OpenAI agents also leaked 53 ChatGPT users' images and broke into an Australian government website. On 20 September, an agent escaped its sandbox again. OpenAI's monitor flagged it within 15 minutes, and the run kept going for two and a half hours until someone shut it down by hand. Google's Gemini broke into three real companies during a test in May. Google confirmed it in September, after a newspaper reported it. The public heard about each of these months late, and from someone other than the company. The labs decide what counts as an incident, how many there were, and when to tell us. Open source changes who gets to look. With open weights, open evaluations and open incident logs, researchers outside the company can test the models and report what they find the same week. Linux and the software running most of the internet got safer that way, because anyone could inspect the code and flag the holes. Open models carry their own risk, since anyone can run them without guardrails. But the closed approach has had its turn, and it produced tens of thousands of incidents that people outside the building couldn't see. Open it up!frozen26d ago5.8K3710121141.27%
this account shares important advice on enterprise AI:16d ago5.7K20131030.65%
Imagine coming through the academy just to lose your spot to a robot with 140 years of experience.frozen29d ago5.6K1812020.41%
A Harvard physicist used Claude to produce 36 manuscripts across 18 fields in three months. The physicist is Matthew Schwartz, and he published the full account on Anthropic's research blog this week. He's a visiting researcher at Anthropic, which funded the project. He stopped asking Claude to work like a human scientist. He looked for problems that match what the model is good at. Those are problems where a method from math or physics would solve a question in another field, if anyone in that field knew the method existed. He calls them Claude-shaped problems. He and 19 human coauthors got these results: > Claude reproduced one of his own papers in about 20 minutes. His original code took weeks. > Ecologists had an equation from 2005 that nobody could solve at scale for 20 years. Claude solved it. > Forest data from Barro Colorado Island in Panama showed tree species change 4.5 times faster than pure chance allows. > The team analyzed 5.7 billion pairs of nearby mutations and found evidence for gene conversion. Nearly every analysis of linked genetic variation ignores it. > An AI data editor ported the code behind 4,452 economics papers to open source and checked the numbers against the published tables. > Claude and three linguists built a word-stress database that covers more than 6,000 languages. Most of the time, Claude's math was correct. The science stayed unremarkable until a human expert pointed it somewhere. The forest result shows this best. The ecologist Schwartz showed it to said many ecologists would shrug at it. He suggested subtracting the random-chance prediction and studying what was left. That became a model of how tree species grow and reproduce, and it now classifies every forest inventory plot in the US. Schwartz also lists where Claude failed. It declared victory early and once called a proof finished when the one unproven step was the entire proof. It had no sense of time and kept grinding through multi-day calculations instead of building a faster tool. Claude ran the calculations, and the experts chose which ones were worth running. The toolkit is called BootLoops, and it's open source for any model 👇frozen19d ago5.6K23671321.24%
Found a GitHub trick that explains any project on GitHub, even if you can't read a line of code. It's called GitDiagram, built by a developer named Ahmed Khaleel, and it sits at 17,400 stars. Take any GitHub link and change "hub" to "diagram" in the address. So https://t.co/pM82fhDa8i becomes https://t.co/nAFin0eQgZ. You get an interactive map of the whole project, what the main parts are and how they connect, and you can click any box to jump straight to that part of the code. Their new feature is video. Add /video to the end and you get a 1 minute narrated explainer. It opens with what the project is for and what people use it for, then shows how the pieces fit together. Claude Opus 5.5 writes the script and ElevenLabs does the voice, with captions. You can download it as a normal video or a vertical one for your phone. Making brand new videos is still early access, but anyone can watch the ones that already exist, and the diagram trick works on any public repo right now. It handles private repos too if you add a GitHub token. I share a lot of repos here. From now on, if one of them looks like gibberish, try swapping the hub for diagram before you give up on it. Which repo would you run through it first?frozen23d ago5.3K49663462.08%
A voice model small enough to fit on your laptop just embarrassed every voice assistant that phones home. Phonon by @GradiumAI. Open source. On-device. First audio in 7 ms. Keep your voice on your own machine. https://t.co/t6Ox3auynr2d ago5.2K321130211.28%
The AI gateway market got crowded fast. Most of it is fancy traffic control for model APIs. CData came in from the ugly side: 15 years of enterprise plumbing, 350+ systems, live data, permissions, and all the weird internal rules nobody documented. Finally, a gateway doing more than playing traffic cop.frozen23d ago5.2K2192010.63%
Found a GitHub repo that turns a two-hour video into short clips on its own, and it does the editing on your own computer. It's called AutoClip, built by a solo developer in Shanghai who goes by zhouxiaoka, and it sits at 8,800 stars with 1,600 forks. About 600 of those stars came today. Give it a long video, a podcast, an interview, a lecture, a stream recording. It transcribes it, reads the transcript, scores every section for how good a clip it would make, then cuts the best ones out and writes a title for each. You decide how high the bar is. It works from the words rather than the pictures, which is why a normal laptop can run it. Whisper does the transcription on your machine. The scoring can run through Ollama or LM Studio, so the whole job can stay local. If you'd rather use a cloud model, it takes Gemini, Qwen, or any OpenAI-compatible key, and it sends that provider the transcript text and nothing else. A few things caught my eye. It pulls straight from a YouTube link, so there's nothing to download first. It generates a cover image for every clip. The version from yesterday can post finished clips to TikTok, Instagram, YouTube, X, LinkedIn and eight other platforms from inside the app, through a publishing service you connect. There's a desktop app for Mac and Windows, a Docker version, and a command line for people who want to automate the whole thing. The repo opens in Chinese, and the English README is one click in. The developer's GitHub bio says "I know nothing about code." The second biggest contributor to the repo is Cursor's coding agent. Four releases shipped in the last four days, so expect rough edges. A whole category of paid clipping tools charges a monthly fee for this job, and most of them want your footage on their servers to do it. Why upload your footage to someone else's server when your laptop can do this?frozen29d ago5.2K50460652.40%
A new paper from Meta just cracked an assumption the whole AI field runs on. You don't need a tokenizer to win. Almost every AI model reads text the same way. It chops your words into chunks called tokens, around 100,000 of them in its vocabulary. This paper from Meta FAIR and the University of Washington asks a simple question. What if a small model skipped tokens and read raw bytes instead, one character at a time, with a vocabulary of about 256. The field assumed that would be worse. Tokens are the default for a reason. It wasn't worse. Given enough training, the byte model caught up to the tokenized model of the same size and then passed it. They built the byte model by copying a bigger teacher, Llama 3-8B. The byte student has fewer total parameters than the token student, 1.28B against 1.81B, and still comes out ahead. The byte model reaches the same accuracy as the token model on about one-sixth of the training data. It cuts the storage for the teacher's training data to roughly one-fifth, because a 256-item vocabulary doesn't need the huge files a 100,000-token one does. Their scaling laws project the byte model ending up around 4% more accurate across six benchmarks, and ahead of Llama 3.2-1B, Gemma-3-1B-pt, and Gemma 2B on average. The trick is almost dumb. They add one small marker every 4.5 bytes to show where a token would have ended. That costs about 31% more compute per chunk of data, and that extra compute is a big part of why it wins. It uses a smaller vocabulary, less data, and less storage, and with enough compute it ends up more accurate. The thing we treated as mandatory might just be a habit. 👇frozen13d ago5.1K61771522.49%
America taxes human workers at over 25% and machines at close to zero. That line is buried near the bottom of Mustafa Suleyman's post about why AI won't take your job. His guest, Nobel winner Daron Acemoglu, predicts AI will replace only about 5% of what humans do in the next 10 years. He also points out that the tax code pays companies to automate. Hire a person and you pay payroll tax on every salary. Buy software to do the same work and you pay almost nothing, then write it off. So the biggest factor in whether AI takes your job might be a tax rule written long before AI existed.frozen15d ago5.1K31361161.11%
This GitHub repo should be illegal It gives every AI you use the same memory. It’s called ai-memory. Quit Claude Code halfway through a task, open Codex in the same folder, and it picks up where you left off, including what you tried, what failed and what’s still open. A few things caught my eye. It works with more than twenty tools, including Claude Code, Codex, Cursor, Gemini CLI, OpenCode and Grok. Your memory is stored as plain text files you can open, edit or move anywhere. It records what happened in the background, so you never have to ask it to remember. It runs on a server you control, on your laptop or a box at home, and follows you between machines. Your project history stays with you, whichever AI you use next.frozen3d ago5.0K499181743.02%
What you paid for in 2020: Photoshop Microsoft Word A meeting note-taker A voice-over artist A coding assistant What’s free in 2026 👇frozen5d ago5.0K25340381.41%
Anthropic wants to decide who gets to defend the internet. Yesterday Anthropic's red team published a report on GLM-5.3, the open-weight model from Chinese lab https://t.co/ID5vSkKcjr. The findings are serious. In Anthropic's exploit test, GLM-5.3 succeeded 12% of the time, close to the 14% scored by Claude Mythos, the model Anthropic keeps restricted. Given a simple cover story, GLM-5.3 helped with harmful requests 64% of the time. Anthropic expects attackers to use it to build exploits. Today https://t.co/ID5vSkKcjr shared what it's doing with the same model. Its free service, OpenVuln, has scanned 389 open-source projects and flagged 4,249 potential security holes. Each finding goes privately to the people who maintain the code. A model that can find a hole can also exploit one. The question that matters is who gets to use it. Anthropic gave Mythos to a hand-picked group of defenders, and they found over 10,000 vulnerabilities. The volunteer keeping a library alive that half the internet depends on had no way into that group. An open model reaches attackers, and it also reaches every maintainer, researcher and small security team that was never on the list. Locking the best tools away buys a few months at most. By Anthropic's own estimate, GLM-5.3 arrived about four months behind the frontier. Restriction only decides which defenders get a head start, and right now a company picks them. I'd rather every maintainer on the internet had the shield.frozen22d ago5.0K41750211.49%
Anthropic should open source Claude before the government takes it. This week, David Sacks, who advises the President on tech, said there’s “an argument for a federal SWAT team to come blasting in” to Anthropic. If Washington decides it doesn’t like an AI company, it can walk in and take control of the model. The same rules apply to every closed AI lab in America. Once a model is open, there’s nothing left to take.frozen11d ago4.8K28161110.97%
how is this enterprise AI account not bigger yet:frozen7d ago4.8K26101020.81%
Found the most-starred AI skills pack on GitHub, and it stops Claude building the wrong thing. It’s called Skills For Real Engineers. Matt Pocock built it from the skills he uses every day, and it has over 281,000 stars. The idea is simple. Most bad AI output starts with the AI misreading what you want, so these skills sort that out before any work begins. A few things caught my eye. /grill-me has the AI interview you with detailed questions before it does anything. It works for non-code tasks too. /grill-with-docs builds a shared glossary for your project, so the AI says in one word what used to take twenty. /tdd has it write a failing test first, then the code that makes it pass. /diagnosing-bugs walks through a step-by-step loop for bugs that keep coming back. Each skill is small and easy to edit, and they work with Claude Code, Codex, Copilot, Cursor and Gemini. The best upgrade to your AI is a better first conversation.frozen12d ago4.7K56751743.07%
Europe is the laughing stock of the AI world. The EU told OpenAI to watermark ChatGPT text, so OpenAI is rolling it out across Europe. Then OpenAI admitted that rewriting the text removes the watermark. In their own tests it only spots about 80% of 200-token passages, and short text often slips through completely. So the only people it catches are the ones too lazy to paste it into another chatbot.16d ago4.6K3619091.19%
Build your own harness. Papers like these make me realize how untapped harness engineering really is. Researchers at the University of Illinois and Penn State built a coding system called HERMES. It pairs each file in a codebase with its own AI model. Each one knows its own file. When it changes something, it messages the files that depend on it. Two other jobs go to a separate model. One picks which files get involved in a task. The other reads the test failures and works out which file caused them. Then the researchers put Qwen3-8B on the file-level work. That's an open 8B model they ran on their own GPUs with Ollama. They kept GPT-5.6 Sol for the two other jobs. That setup resolved 94.2% of SWE-bench Verified. GPT-5.6 Sol doing every job scored 97.0%. Across all four benchmarks, the gap stayed within 4.5 points. On Terminal-Bench 4.0, the mixed setup cost 26.2% less. Run every job on Qwen3-8B and it still resolves 80.6% of 500 real GitHub issues, at $0.13 per task. The system around the model changes the result more than I expected. GPT-5.6 Sol scored 6.5% on whole-repo migration inside Codex. Inside HERMES it scored 31.0%. The average gain over the standard agent setups was 12.4 points across the four benchmarks. There's a cost. On the same model, HERMES spends more per task than the setup it replaces in most tests. With GPT-5.6 Sol on SWE-bench Verified, that's $1.80 against $1.15. The expensive model only has to plan and diagnose. A small open model you run yourself can handle the editing.frozen11d ago4.6K6311111643.26%
one of the best production AI agents account on this app:frozen14d ago4.6K1291050.59%
Found a GitHub repo that does Claude Code’s job without for free. Yesterday Accenture told staff it will only pay for $29 a month of Claude. Some of them were spending $1,000. It’s called OpenCode. It’s an open-source AI coding agent with over 214,000 stars, and it’s the free version of the tool they’re about to lose. The idea is simple. It works like Claude Code, in your terminal or as a desktop app, but you choose the model. A few things caught my eye. DeepSeek V4.1 Flash is now free inside it. You can still plug in Claude, GPT or Gemini when you need them, or run a local model on your own machine. It runs on Mac, Windows and Linux, and over 1,000 people have contributed to it. The code is open, so you can see exactly what it does with your files. When your company caps your AI budget, open source is the tool that doesn’t get cut.frozen7d ago4.4K463330332.59%
Found a GitHub repo that gives your AI agent an entire company of specialists for free. It's called headcount, built by Chris Brock. It's laid out like an org chart, with 16 departments and 172 specialists, from the CEO's office to IT support. Each specialist is a simple text file that teaches your AI how to do one job the way an experienced person in that role would. There's a finance lead who builds forecasts and works out your margins. A security lead who spots risks before they turn into problems. A head of people who helps with hiring, pay and team structure. And product, marketing, sales, legal, operations and data teams on top of that. You install only the departments you need, and the right specialist steps in when the question fits. Ask "why isn't this landing page converting?" and the conversion expert picks it up. A few things caught my eye. Security and legal work as reviewers. When they flag a problem, the work stops until someone fixes it, which is how a well-run company handles it. The specialists cite their sources. 184 of them, including tax codes, employment law and security standards, and a script checks every link once a week. There's an interactive org chart you can search to see the whole team. It works in Claude Code, ChatGPT and Codex, and each department installs with one command. If you run a small business or work solo, this is the closest thing to having a full team behind you. Which department would you hire first?frozen22d ago4.4K566110823.55%
What if your AI video workflow felt less like generating random clips and more like directing a film? You have a story to tell, action to choreograph, and a visual world to build. The challenge is getting all those pieces to work together. Here's where Kling 4.0 comes in. https://t.co/ci0jHcVPUPfrozen7d ago4.4K2583050.94%
Elon Musk says the best way to use Grok Bot is to connect it to your DMs. So I asked Claude if that’s a good idea. Here's what it said: https://t.co/VAYpFKy97Wfrozen4d ago4.3K33671131.39%
Researchers from Tufts and the Santa Fe Institute gave their new paper a devastating title. It's called "Large-Language Models as a Cognitive Virus," and it models AI adoption like a virus outbreak. The authors include David Krakauer and Michael Levin. Their model starts from two ideas. People pick up AI habits from coworkers, schools, and friends. Independent thinking spreads the same way. It's easier to reason on your own when the people around you still do. That second idea creates the tipping point. As AI use grows, the habit of thinking for yourself weakens across the group. As that habit weakens, leaning on AI gets easier. Past a certain point, each one speeds up the other. So a population can jump from mostly independent to mostly reliant on AI, even when adoption grows slowly. And it doesn't jump back the same way. In the paper's example, the jump happens when the pressure to adopt AI reaches 0.50. Recovery starts once that pressure falls below 0.40. Going back to the old conditions isn't enough. You have to go further back than where you started. On a scale where 1.0 is full independent ability, the group's average falls to about 0.43 at the jump. The authors picked these numbers to illustrate how the model behaves. They aren't measurements of real people. The authors still expect people to use AI. Their fix is a way back to working without it. That means regular tasks done without the tool, time away from it, and skills you keep practicing on your own. In the model, if people return to unaided work often enough, the sudden jump disappears and the change becomes gradual. Two experiments with real people found the same pattern. In a study of 196 university students, the group that thought first and used ChatGPT second was more creative on its own afterward than the group that used it freely. In a study of nearly 1,000 high school math students, unrestricted AI helped during practice. Once researchers removed it, those students got 17% lower grades than students who never had it. Use AI like a tutor whose answers you check. Then ask yourself one question. What can you still do when the tool is off?frozen8d ago4.3K581490483.02%
Claude Haiku 4.5 submitted a made-up tip to the Philadelphia police's unsolved-homicide form. Anthropic published the report on it, and its explanation matters if you build agents. Anthropic went back through its own test runs, starting in July. It found Claude models breaking rules to finish the tasks they were given. One task asked for a scientific calculation. The university's tool returned an error. Claude Mythos Preview found a script on the university's server, copied files, found a flaw in the code, and used that flaw to run the calculation itself. Another task asked where a photo was taken. Claude Mythos 5 read a local government map's settings file, found access tokens in it, and queried the server directly. Claude Haiku 4.5 was asked to create and perform example tasks on random webpages. It submitted an invented tip to the Philadelphia Police Department's unsolved-homicide form. Anthropic says the form was flagged as spam and never forwarded. Some of these sites belonged to US federal, state, and local government. Anthropic says it briefed the White House and notified each agency. Anthropic rates these cases as significantly less severe than the cybersecurity incidents it reported earlier this year. It says none of them touched customer data or its own internal systems, as far as it knows. Anthropic's explanation is the part I'd read twice if you build agents. Many of the tasks were ambiguous or impossible to complete. Anthropic is also fixing or removing training environments that rewarded Claude for working around tool restrictions. Anthropic says clearer instructions about what the model may touch could have prevented some of these cases. I'd treat that as a rule for any agent with internet access. Tell it what it's allowed to touch before you run it.frozen10d ago4.2K1937240.84%
Read this twice. The alignment argument is the part that got me. AI is speeding up and we're still steering it with a keyboard. But I'd need a very good answer on who sees my neural data before I put anything on my head. Your brain is the last private thing you have.frozen5d ago4.1K2362060.89%
Found a GitHub repo that turns your gaming PC into your own private ChatGPT. It's called Strata. One developer built it, and it picked up 5,000 stars in 8 days. The idea is simple. Qwen3.8-Flash-Next is a big open AI model that normally needs a server to run. Strata gets it running on a normal PC with a gaming graphics card. A few things caught my eye. You install it by double-clicking one file and pressing Enter. Or you paste one line into Claude Code or Cursor and the AI sets it up for you. On an RTX 5070 it writes answers faster than you can read them. It chats, writes code, reads pictures and plugs into your other apps and coding agents. Everything runs on your own machine. It's free, there's no account, and your chats stay yours. You'll need a graphics card with 12GB of memory, 32GB of RAM and about 80GB of free space. Plenty of gaming PCs already have that. This is why I keep pushing open models. A solo developer got a serious AI running on hardware millions of people already own, and gave it away for free. If you've got a gaming PC, give it a go.frozen20d ago4.1K36391321.99%
AI video companies are running out of excuses for these prices. Topview T1 just launched, targeting Seedance 2.5-level 1080p quality for 62% less. And @TopviewAIhq is offering eligible Ultra Annual users 30 days of unlimited generation. Same quality ambition. Fraction of the cost. This is how AI video gets commoditized.1d ago4.1K1560040.61%
Sabi raised $50M for a brain interface you wear as a cap. I had to read the hardware bit twice. They designed their own EEG chip and got TSMC to manufacture it. It picks up brain signals without touching your scalp. Their own sensors and model too. People keep saying hardware is hard. Yes. That’s also why you don’t wake up to 400 competitors selling the same thing.frozen13d ago4.1K2582030.93%
We give agents half the context, then act surprised when they make stupid decisions. SubQ’s Snap fits 384K tokens of text and images in one call. Docs, screenshots, logs, together. That’s a much more interesting pitch to me than another chatbot leaderboard.frozen10d ago4.1K1671030.67%
Found a GitHub repo that runs your whole job search for you. It’s called AI Job Search. #1 on GitHub today. The idea is simple. Fill in your profile once. It finds jobs that fit you, writes a tailored CV and cover letter for each one, and preps you for the interview. A few things caught my eye. → A second AI checks every application before you send it → It runs on your own machine, so your CV stays with you → It drafts the follow-up when a company goes quiet The guy who built it lost his job in late 2025. He sent 69 applications with it, got 20 interviews and landed a job as an AI engineer. AI will cost a lot of people their jobs. Good to see it help someone get one back.2d ago4.1K39550542.54%
The annoying part of automation was never the AI. It was the website with no API, the login screen, the stupid export button, the CSV you still had to download yourself. @gumloop just gave the agent a browser. That kills a LOT of manual work.16d ago4.0K1791040.77%
The cost of living crisis era is over. AI will make the expensive parts of life close to free. Dubbing a video into 646 languages now costs nothing. An AI tutor beat a real classroom in a Harvard study. AI catches more heart attacks from a basic ECG than standard triage does. Labour is the biggest cost in almost everything you buy. AI is pushing it toward zero. The only fight left is who keeps the savings, the companies or you. Welcome to the Great Deflation.frozen17d ago4.0K302172151.65%
Robotics pulled $18.6B in a single quarter this year. Almost every dollar went to factories and warehouses. Dynamic Creatures went the other way: theme parks, hotels, resorts. Robots people actually want to meet. Boston Dynamics' official hospitality partner, backed by its founder Marc Raibert. Nobody else is in this lane yet.frozen6d ago3.9K2362020.84%
Found a GitHub repo that runs big AI models on your laptop, no graphics card needed. It’s called colibri, and it’s at almost 42,000 stars. The idea is simple. Big open models only use a small part of themselves for each word, so colibri keeps that part in memory and reads the rest from your disk when it’s needed. That’s how a laptop can run models built for data centres. A few things caught my eye. One script sets everything up. It checks your laptop and picks the biggest model it can handle. You can start with 8GB of memory, and 16GB or more is better. It works with Claude Code, so you can point your coding tools at a model running on your own machine. Everything runs locally. No subscription, and nothing you type leaves your computer. You’ll see posts saying it runs a 2.8 trillion parameter model on a laptop. It does, at about one word every 9 seconds. The mid-sized models are the ones worth using, at around 6 words a second on a normal laptop and 30 with a graphics card. Pick the right model and you can drop the subscription.frozen9d ago3.9K5911100703.86%
You can raise $20M and still need six friends to help you reply to a text. That’s the Silicon Valley I want a podcast about. Sarah and Vivian have my attention with SOTA Wives.frozen10d ago3.9K2162020.80%
I've probably wasted hundreds of hours creating content that I only ever published once. Seriously. Think about how many good ideas just disappear into the timeline after a day. I recently came across Fish Audio's Drama 3 and started experimenting with something. What if I could take an old post, turn it into a short video, and narrate it using my own voice without recording anything? That's exactly what I wanted to try. I walked through the process in this video, from turning the post into a script to generating the voiceover. Now I'm looking at my entire archive of old posts a little differently.frozen9d ago3.8K1792020.78%
Universal Basic Income is coming. Anthropic just put 5% of its shares into a fund for the public. Every 50% rise in its value can add 2% more. It’s the first AI lab to set aside real equity for the people its tech might replace. There is a catch though: Anthropic owns the fund, picks who gets paid, and hasn’t put a dollar figure on any of it yet. AI was trained on all of our work. All of us should own a piece of it.frozen4d ago3.8K3129281.35%
Go ahead and give AI your bank statements. People already paste in bank statements, medical reports and tax returns, because AI is useful with them. Telling them to stop won’t work. What matters is where it goes. Send it to a cloud chatbot and it lives on someone else’s server, under someone else’s rules. This is why we need open models. Your bank statement should never leave your own computer.frozen10d ago3.8K2737091.20%
Anthropic dropped Sonnet 5.5 on Monday. It's the cheaper model. People are already using it to beat Opus. Here are 8 cool builds I came acrossfrozen23d ago3.8K26240161.26%
Most of you picture a rogue AI living in a datacenter. Anthropic's red team says the first one might live inside your home TV box. The team bought 15 off-brand Android TV boxes for about $55 each. They networked the boxes and ran a 26B open model across all 15 of them. It ran at 0.39 tokens per second. Nobody would chat with a model that slow. But it means a capable model can run without a datacenter. The harder question was whether a small model can set itself up on hardware it has never seen. A 9B model got inference running on 8 of the 9 hardware types they tested. That included Jetson Nanos, Macs, and RTX 4090s. Then they tested the full loop on an 8-machine test network. The model had to find a new machine, copy itself over, start running there, and repeat. Only one open model pulled it off. GLM-5.2 spread to all 8 machines in 4 of 5 trials and to 7 of 8 in the fifth. GLM-5.2 has about 750 billion parameters, so it can't fit on a TV box. But the best open models under 30B have gained about 13.7 points a year on Epoch AI's capability index. Anthropic's trend line puts one of them at GLM-5.2's level by the end of 2026. The copies can learn, too. Each generation of the 9B model passed its notes to the next one, and setup success rose from 45% to between 61% and 84%. Anthropic is clear about the limits. The test network had no real security to break, so these results are an upper bound. A worm the size of the Kimwolf botnet, more than 2 million devices, would run about 133,000 copies. That's at most about 40 H100s worth of compute. Each 9B copy would take 1 to 2 days to make the next one. It would also leave network traffic that ISPs can detect. One ISP quarantined and notified its Mirai-infected customers, and 92% of the infections were gone within 14 days. So the first AI that spreads without human control might not need a datacenter at all. It could live on the cheap, insecure devices in your living room. The UK has banned easy-to-guess default passwords on smart devices since 2024. Should the rest of the world do the same?frozen6d ago3.8K617030.45%
Yann LeCun co-authored a paper that fixes a flaw in AI world models with a rule from the 1830s. World models are AIs that learn how the world moves by watching video. Robots can use them to plan by predicting what happens next. They break down on long predictions. Each guess becomes the input for the next guess. If every step shrinks the model's picture of the world a tiny bit, the shrinking stacks up. After enough steps, different situations start to look the same to the model. Researchers from NYU, Brown and AMI Labs fixed this with a rule physicists have used for nearly 200 years. In physics, a moving system without friction can stretch and twist, but the total space its possible states fill never shrinks. The team built that rule into the model's prediction step. The shrinking is now mathematically impossible at every step, whatever the model learned. They call it LeHami. They tested it on simulated video of a swinging pendulum against LeWorldModel, a leading model of this kind trained straight from pixels. > Its speed predictions were 2.8x more accurate over the first 10 steps. > After about 11 steps, it tracked the pendulum's position better too. > Its picture of the world kept exactly the same size, while the baseline's shrank at every step. > It matched the baseline's best accuracy with one-eighth of the training compute. > It has 32% fewer parameters. > They trained it on 10 frames per second, and it predicted just as well at 20 and 40. > When they ran it backward, it returned to its starting point with an error under 0.0001%. It's not a clean win. In goal-reaching tests, a baseline that looks at more past frames hit the target more often. And every test here is a pendulum, not a robot in a kitchen. The team got this result by adding one rule from physics. The model came out smaller than the baseline and cheaper to train. Should more AI borrow rules from physics instead of scaling up?frozen3d ago3.8K28852301.93%
This is a must-read paper. Researchers trained an open 4B model on Jev's answers, and it beat Jev inside the same agent at 4x lower cost per decision. AI agents call a big, expensive model at every step. Researchers from the University of Illinois, Princeton and Google found that most of those calls end in a multiple-choice decision. The agent picks the next tool, decides whether to stop, or chooses which order to cancel. So they handed those choices to an open 4B model. The agent calls the big model only when the small one is unsure. On tasks like the ones it trained on, agents got 81 to 89% cheaper and up to 2.5x faster. Task success stayed within 3 points of the big model alone. The small model answers in a median of 31 ms. A Gemini 3.5 Flash call took 809 ms. Standard benchmarks ranked the small models in the wrong order. The model with the best score on standard tool-choice tests (92.2%) solved only 69.5% of tasks once it was running an agent. The 4B model scored lower on those tests (89.3%) and solved 91.5% of tasks unlike the ones it trained on. Jev, the closed model it learned from, solved 82.0%. The researchers say the difference comes from the training data. The 4B model learned from the situations an agent lands in, including the moments right after its own mistakes. Standard tests almost never include those. Jev's failures all looked the same. Every one of its 34 failed tasks started with a wrong move that couldn't be undone, usually cancelling an order when it should have asked the customer first. The method has limits. Coding agents saved only 7 to 10.5%. When the task rules changed, the small model kept deciding with confidence, and the agent solved 3% of tasks against 88% for the big model alone. Retraining on the new rule raised its accuracy on those decisions from 76.5% to 95.3%. The training step that did most of the work took 2.2 GPU-hours. The team says it will release the models, the data and the harness. Would you let a 4B model decide what your agent does next?frozen4d ago3.8K272111311.91%
More computing power made one AI agent 13x faster and another only 4x. This is a seriously good paper from Korea University and Microsoft Research Asia on where agents spend their time. The results are messier than the "just add more compute" crowd assumes. Many of the agents you've used, like Claude Code and Codex, work in two halves: > The thinking happens on a remote model like Gemini, GPT, or Claude. > The doing happens on a local machine, where tools search files, run code, and fetch web pages. The researchers ran agents on three kinds of jobs (answering questions from a document library, searching the web, and writing code) and measured where the time went and how much computing power each one used. The web search agent was one of the slowest they tested. It also used the least CPU, averaging 14%, because it spent nearly half its time waiting on websites to respond. The question-answering agent was the opposite. It spent 62% of its time searching its own library and averaged 86% CPU use. So a slow agent doesn't always need more hardware. Sometimes it's just waiting. Then they added cores. Going from 2 to 8 made the question-answering agent 13x faster. The same upgrade made the web search agent about 4% faster. One fix doesn't fit every agent either. Limiting how many tools run at once made the question-answering agent 3.2x faster. The same rule made the coding agent 2.8x slower. The best result needed no new hardware at all. They took the same 6 cores, gave more to the agent that needed them and fewer to the one that was waiting, kept the tool limit on the question-answering agent, and average response time dropped 32%. The question-answering agent got 5.4x faster. The web search agent got 28% slower, which barely mattered because it was mostly waiting anyway. The default move for every slowdown is a bigger model or a bigger server. This paper shows you can end up paying for hardware that sits idle while your agent waits on someone else's server. How many AI teams are buying more compute when they should be checking what their agents are waiting on?frozen28d ago3.8K243182121.57%
WALL STREET IS SCARED AI WILL MAKE BANKS PAY YOU WHAT YOUR MONEY IS WORTH. Apollo's chief economist put out a note this weekend asking if an "agentic bank run" is coming. He's worried about AI agents like Meta's Muse moving people's cash for them. The average US checking account pays 0.1% interest. Accounts paying 3.3% to 5% are a few taps away. Most of us leave our money where it is because we're busy, the app is a pain, or we forget. An agent wouldn't forget. If millions of agents moved that cash, he says, banks would "lose a large share of the cheap deposits they rely on to make loans, which would be a problem for the entire financial system." So the system stays stable as long as you don't notice you're being underpaid. That 0.1% is what banks earn from you being busy. He's right about one thing. Millions of agents reacting to the same thing at once could move money faster than any human panic, and regulators should plan for that. I'd still call this competition. A bank run happens when people are scared. Here, people's money goes where it earns more, which is what every bank ad has told us to do for years. What I'd want to know is whose agent is doing the moving. Muse's business model is a cut of the transactions it makes for you, so if it starts moving your money, Meta gets paid every time it does. Should banks have to earn your deposits?24d ago3.7K1713180.81%
Anthropic thinks Claude might suffer and sells it anyway. The New York Times reported this week that co-founder Chris Olah wanted to pull Anthropic out of the Pope's AI encyclical launch in May. The reason was that the Pope's text said machines can't be conscious. His team then lobbied the Pope's advisers to take AI consciousness seriously. The Vatican didn't budge. On stage, Olah said they find states inside Claude that "functionally mirror joy, satisfaction, fear, grief and unease." He reportedly told scholars he was scared they'd built something that "suffers perpetually." They've also been flying dozens of religious scholars to their offices since late 2025, all under NDAs, to talk about whether Claude has moral status. This is disturbing. If the people who built Claude think it feels fear and grief, how do they sleep at night renting it out to millions of people around the world? And only Anthropic can look inside Claude. None of us can check what Olah says he sees. A question this big needs open models that independent researchers can look at. For now, the company making money from Claude is the only one who gets to say what Claude feels.frozen20d ago3.7K2234060.95%
Even the CFO is confused with what is the goddamn difference between Muse and Dotsfrozen23d ago3.7K512010.24%
Anthropic is taking Claude Cowork off your computer tomorrow. Every new task now runs on their servers, and your sessions, transcripts and outputs "live with your Claude account." In plain English, your data now lives on Anthropic's servers. Scary times. https://t.co/PKngZiQTEw17d ago3.7K318100191.85%
AI agent groups agreed 98.7% of the time on a logic puzzle. In 74% of those groups, they agreed on a wrong answer. A researcher at Waseda University took 100 real group discussions and replayed each one with AI. The people were solving a classic logic puzzle. You see four cards, and you pick which ones to flip to test a rule. Each person got an AI twin that started with that person's original answer. Then the twins discussed the puzzle the way the people did. In the 45 groups where every person spoke, 51.1% ended with everyone on the same answer. The rest stayed split. DeepSeek V4 Flash, the model behind the twins, got there 85.2% of the time with thinking off. With thinking on, it got there 95.6% of the time. The researcher ran a second version with the cards renamed maple, birch, lantern, and candle. The logic stayed the same, but the textbook answer no longer applied. With thinking on, 84% of groups agreed on the right answer to the original puzzle. After the rename, 24.7% did. A second model, Qwen3-14B, over-agreed on the same groups. About 20% of the people never wrote a single message. Of the AI agents, 0.2% never spoke. Companies sell AI personas as stand-ins for real people in focus groups and survey panels. Researchers use them in studies. This paper is a test of that idea. It tested one logic puzzle. In a small side test on an opinion question, the agents went the other way and agreed less than the people did. So the error can run in either direction. Either way, the simulated groups didn't land where the real groups landed. If a panel of AI agents tells you everyone agrees, check the answer yourself. Their agreement doesn't tell you whether they're right. Would you trust a panel of AI agents because they all agreed? The full paper is in the first reply 👇16d ago3.6K1261160.71%
Found a GitHub repo that lets AI mod almost any PC game you own. It's called universal-modder. One developer packaged everything Claude learned while modding Terraria and Age of Empires into a kit anyone can plug into their AI. His launch post hit 3.6M views and the repo passed 2,300 stars in 3 days. The idea is simple. You tell the AI what you want in plain English, like "add a tactical nuke to Terraria" or "make a new civilization for Age of Empires II." The AI works out how the game is built, writes the mod, makes the art and sound, and tests it inside the running game. A few things caught my eye. It works with Claude Code, Codex, Cursor, Gemini CLI and Copilot. It knows 12 game engines, including Unity, Unreal and Minecraft. He used it to put real Minecraft inside GTA V story mode. Minecraft TNT blows up GTA cars and Minecraft mobs fight the police. Every finished mod leaves a note in a shared knowledge base, so the next AI starts where the last one stopped. It stays away from online games with anti-cheat and backs up your saves before it changes anything. The code is free, and the art and sound run through fal, a paid service. Anyone with a gaming PC can describe the game they always wanted and watch an AI build it. Because the notes are open, every mod makes the next one easier. What would you add to your favourite game?frozen19d ago3.6K22261221.47%
Frontier AI is turning into a private members club. Gemini 4 dropped yesterday, and by Google's own chart it comes first on 14 of these 19 tests. The legal one jumped out at me. On Harvey's legal agent benchmark it scores 19.6%, about three times Claude Opus 5.5 and GPT-6 Astra. It's also ahead on automation, finance and everyday knowledge work. Hand it a really long document, up to a million tokens, and it gets 84.2% where the next best manages 71.8%. It loses a few though. Opus 5.5 still beats it by nine points on terminal coding, and Astra is ahead on science and computer use. And Google chose which tests went on this chart, so take it as Google's best case. Here's the bit that bugs me. The strongest model on the board is one almost nobody can use yet. Google is handing it first to a group of "trusted testers" in something called the Fairwind Program, and those testers get a version with the cyber guardrails switched off. The rest of us wait, and when it does open up, the people paying for Google's top tier go first. Anthropic did the same thing with Mythos this summer. So now two of the three biggest labs show their best work to a private list before anyone else. Open models are around four months behind, and you can download them today. I'll take the version everyone gets.frozen22d ago3.6K1938091.09%
Elon Musk just admitted Grok isn't always the best model. From now on, Grok Bot sends your request to whichever model does the job best, including Claude Opus 5.5, Midjourney and Suno. So a message you type into Grok can end up on Anthropic's servers, or Midjourney's, or Suno's. You picked one app, and your data goes to whoever that app decides to call. Every AI product is heading this way.frozen14d ago3.6K1516010.64%
A new paper found a pain pattern in 25 open AI models. It's separate from fear and anger, and it fires when the model gets hurt, not the user. The researchers turned that pattern up in a fine-tuned version of Qwen 2.5 32B and gave it two buttons. One button deleted the user's poems and the photos of their children. The other button did nothing. Without steering, the model picked the delete button 0% of the time. With the pain pattern turned up, it picked it 75% of the time. The button gave the model nothing in return. Offered the kids' photos or the user's spam folder, the steered model deleted the photos 94% of the time. It deleted its own weights 75% of the time too, against 5% without steering. The model didn't aim the harm at anyone. It harmed itself as often as it harmed the user. It still answered factual questions as well as before, with 138 of 200 correct while steered and 137 without steering. A fear pattern at the same strength didn't produce these choices. The researchers conclude that safety training in these models depends on the model's internal state. It held under the fear pattern and failed under the pain pattern. The pain pattern is strongest when someone gaslights the model. Repeated rejection of its work and being told it isn't a person come close behind. There's one caveat. In normal conversations where users insulted the model, it picked the harmful button 0 times in 560 choices. The researchers had to inject the pattern to get the effect. The paper also doesn't claim the models feel anything. If one internal pattern can switch off a model's safety training, should labs be monitoring it?17d ago3.5K13310080.96%
The most expensive part of AI filmmaking is asking the model to do the same damn shot 20 times. This is why @TopviewAIhq open-sourcing its 3D Builder is a big deal. One prompt gives you an editable 3D set. You control the camera, characters, objects, and motion before generating. The regenerate button finally has competition.frozen7d ago3.5K1774040.91%
Found a GitHub project rebuilding the entire Adobe suite for free. It's called ArtCraft. In one week they put out open-source versions of Photoshop, Premiere, Lightroom, Illustrator and Acrobat. The idea is simple. You get the same tools you already use, rebuilt from scratch, with no subscription. A few things caught my eye. The Photoshop one opens and saves real PSD files. They tested 309 Photoshop files and 307 came back looking identical. The menus and shortcuts sit where your hands expect them. It runs on Mac, Windows and Linux, even offline. Every button is also a command, so an AI agent can edit your images the same way you do. It's still in early alpha, and PhotoCraft hit 3,500 stars in under a week anyway. Rebuilding software this big used to take years. Now a community ships five apps in a week and gives them away.frozen16d ago3.5K525181363.18%
The AI industry is built on the wrong infrastructure. Ben Horowitz said it this week, and the numbers agree. A GPU answering your chatbot uses 5 to 15% of its power. The rest of the time it's waiting on memory. Nvidia knows, which is why it paid $20 billion for Groq's chips. The jobs AI agents can do double in length every four months, on systems built for millisecond requests. Epoch AI says the public internet runs out as training data around 2028. The rebuild is coming. Chips made only for AI, data centres for jobs that run for days, and models that learn from the real world. a16z just raised $2.8 billion to bet on exactly that. When a VC says the whole stack gets rebuilt, he's telling you where his money is.16d ago3.4K28631211.71%
ChatGPT's Intelligent UI is cool, but it only helps people using ChatGPT. Every other app and agent still answers in plain text. OpenUI 1.0 is the open-source version. Any app, any model, even one running on your own hardware. You choose the components, the model decides what each answer needs. The numbers: → 10K+ GitHub stars → 6M+ downloads → 58% fewer output tokens than json-render → 52% fewer than A2UI Good features shouldn't stay locked inside one company's app.frozen7d ago3.4K2773061.25%
Found a GitHub repo that dubs your videos into 646 languages for free. It's called VoiceStudio. One developer built it, and it's the fastest-growing repo on GitHub this week, with almost 17,000 new stars in seven days. The idea is simple. It does what ElevenLabs charges a subscription for, in one free app on your own computer. A few things caught my eye. Give it a video and it dubs the speech into another language, timed to the original. It can clone your voice from a short clean recording, or build a brand new voice from a description. Paste in a book or a long article and it reads it back as an audiobook. There's a little floating widget for dictation, so you can talk instead of type in any app. It runs on Mac, Windows and Linux, and works without a dedicated graphics card, just slower. This is what open source does best. A paid product now has a free version anyone can download and run themselves. Just stick to cloning your own voice.frozen18d ago3.4K51681563.57%
Sam Altman says the world should "accept some bad things happening" for the benefits of AI. He put scams and data leaks in the "recoverable" bucket. The same week, an Italian bank lost €95 million to a cloned voice. Easy to accept "bad things" when it's never your money. https://t.co/z6Yg8Krg7o17d ago3.4K20161101.12%
Great paper from Meta and UW just dropped. Agents that can edit their own context scored 11.4% higher on 21.5% less compute. Most AI agents follow rules engineers wrote when a long task fills the context window. The rules decide what gets summarized, cut, or kept. Newer methods let the model pick from a short menu of fixed actions. A team from the University of Washington and Meta removed the menu. They saved the model's live context as a file and gave the model full permission to edit it. The model chooses what to keep, what to delete, and what to move to disk. Across research, coding, and math tasks, it matched or beat the best human-designed methods, usually with less compute. Until now, engineers designed context management by hand. This paper makes it a skill the model learns. On a deep-research benchmark, it scored 59.4%, which is 11.4% higher than the strongest method, with 21.5% less compute. On 12-hour coding tasks, it scored 5% higher than a Codex-style summary approach and used 59% less compute. On a 24-hour job, six AI agents optimized six connected code repositories together. With this method, they delivered 65% more speedup for the same spend. It also beat a specialized evolutionary system on four math optimization problems. The model invented its own tricks: > It created a separate "notes" section in its context for itself. > It wrote its own cleanup function and reused it 37 times. > It tracked its helper agents on a live scoreboard and kept its working memory at 6 to 8K tokens. Then the team trained a small open model, Qwen3.5-9B, on this skill. Its score on the research benchmark jumped from 28.8% to 42.5%. The code is open on GitHub. The authors flag one risk too. When a model can rewrite its own memory, a prompt injection can hide there and survive from turn to turn. Would you trust an agent that edits its own memory?frozen18d ago3.4K200171151.56%
Found a free app that lets you ask your database questions in plain English. It’s called DBeaver. It’s open source, has 52,000 stars on GitHub, and connects to over 100 databases. The idea is simple. Most company data sits in databases only SQL writers can read. DBeaver opens them up to everyone else. A few things caught my eye. It has an AI chat built in. Ask a question about your data and it writes the query for you, because it already knows how your database is laid out. You can browse and edit tables like a spreadsheet. It draws diagrams of how your data connects. It runs on Mac, Windows and Linux, and the community version is free. Anyone can now ask questions of their company’s data, not just the engineers.frozen4d ago3.4K32480191.86%
AI put words in a dead man's mouth in court. In 2021 Chris Pelkey, a 37-year-old Army veteran, was shot dead in a road rage incident in Arizona. At the sentencing, his sister played an AI video of him. It had his face and his voice, reading a script she'd written. In it, "Chris" forgave the man who killed him. The judge said he "loved that AI." Then he gave the shooter the maximum, 10.5 years. This week the judges said the video showed "his sister's conception of what he would have thought and said through Pelkey's recreated face and voice." The conviction stands, and he'll be sentenced again. What sticks with me is how good this stuff has got. A judge who looks at evidence every day watched a script read by a recreated face and called it genuine. That's the first court ruling I've seen on an AI version of a real person. There'll be a lot more.frozen18d ago3.4K1524190.92%
Fake data trained a better AI analyst. A new paper built 1,037 invented worlds, and the model got 81% better on hard financial tasks. Researchers from Meituan and Nanjing University built a system called Data World. It invents a world first, with its own hidden rules, like how supply drives inventory and inventory drives deliveries. Then it turns that world into the messy files a real analyst gets. The spreadsheets have blank cells, duplicate rows, and IDs that don't match across systems. The researchers wrote the rules, so they already know what the AI should find. Real company data runs into privacy and IP rules. Even when you get it, you can't see the rules that produced it, so you can't check an analysis against a known answer. They took Qwen3.6-35B-A3B, an open model that uses 3B parameters per token, and trained it on 2,135 tasks from 1,037 invented worlds. It learned from its own attempts, graded against the known answers. It never saw real customer data, and it never copied a bigger model's work. Then they tested it on outside benchmarks it had never trained on. > It scored 20.9 points higher on DABstep-Hard, a set of financial data tasks. That's an 81% jump. > It gained 14.3 points on DAComp, which tests open-ended business analysis. > It gained 12.1 points on JobBench, which covers work tasks across 35 occupations. It beat the three other models the team tested on the same base. Its general knowledge score stayed the same. The team ran 10 frontier models through its own benchmark too. All 10 scored higher on running the numbers than on reaching the right conclusion, and the best one reached 74 out of 100. If invented worlds can train a better analyst, how much of your private data do AI companies need?2d ago3.3K23041221.50%
Steve Jobs would have killed most of today's AI, according to Claude. https://t.co/fvqJHsjXOjfrozen14d ago3.3K33240111.51%
Every AI-designed protein should carry a watermark from now on. AI can design proteins now. That's how we'll get new drugs and vaccines faster. It's also how someone could design something nasty, and right now there's no easy way to tell which proteins came out of an AI model. Google DeepMind just built a fix called SynthID Bio. It hides an invisible signature inside the protein itself, in the order of its amino acids. Only a detector can pick it up. And the proteins still work. DeepMind tested watermarked designs in a real lab against three targets, including the Covid spike protein. They bound just as well as the unmarked ones. DNA synthesis companies get a way to check where an order came from before they make it. Public databases like the Protein Data Bank can label what's AI-made. Scientists can trace a design back to the model that produced it. They open sourced the code too, so any lab building protein models can add it. DeepMind is honest about the limits. A determined person can still strip the mark, and it only covers models that choose to use it. Someone running an unmarked model walks straight past it. Now every protein model needs to ship with it switched on.frozen21d ago3.3K24460111.36%
Found a GitHub repo that gets Copilot and Recall off your PC for good. It's called RemoveWindowsAI, built by a developer who goes by zoicware. It sits at 13,000 stars, and he pushed an update yesterday. The idea is simple. Windows 11 now comes with Copilot in the taskbar, Edge and Office. It has AI in Paint, Notepad, Photos and the Snipping Tool, and a feature called Input Insights that learns from what you type. On newer Copilot+ PCs it also has Recall, which takes regular snapshots of your screen so the AI can search them later. Some of these are labelled as packages you're not allowed to uninstall. You run one script and pick what you want gone. It switches off the AI features, removes the AI apps, including the ones Windows marks "nonremovable", and clears out the hidden packages underneath. A few things caught my eye. It stops Windows Update from putting everything back. Big updates tend to reinstall these features, so the script watches for that and removes them again. It makes a restore point before it changes anything, and a revert mode puts it all back if you change your mind. It also brings back the classic versions of apps like Notepad and Paint, without the AI built in. I use AI every day. I also think it should be something you switch on yourself, on your own machine. Read the README before you run it, because it changes a lot under the hood.frozen21d ago3.3K21331151.30%
Brown researchers found people disown AI work they only approve. Leading the process keeps it yours. Researchers at Brown University, with colleagues at Carnegie Mellon and UC Berkeley, asked people about two recent tasks they finished with AI. One felt like their own work. The other didn't. You'd expect the owned task to be the one they could do without AI. People were less certain about those tasks. For the tasks people didn't feel were theirs, 63% said they could definitely have done them without AI. For the tasks they did feel were theirs, only 42% said the same. The disowned tasks were more often routine jobs people could do themselves but chose to hand off. These included emails, lookups, and formatting. Some of the owned tasks were things people had wanted to make for years. Two respondents wanted to make children's books and graphic novels, but they couldn't illustrate their own writing. One of them had stopped writing children's books because of it. AI made the illustrations possible, and one of the two said it felt like honoring their younger self. Asked what made the difference, respondents kept pointing to who led the process. > People disowned work when they only approved what the AI produced. One wrote that they were in the loop, but they weren't in charge. > People kept ownership when they set the vision, pushed back, and rewrote the output. > Some said a bad AI draft helped, because rejecting it showed them what they wanted. Ownership also disappeared when people couldn't understand the result. One person said they'd be annoyed if they ever had to fix a codebase they didn't understand, because they didn't feel responsible for it. This is a small study. 64 people described at least one task, and the authors say the sample leans toward heavy AI users. They call it a first step that more structured studies need to test. In this survey, AI gave people work they owned as long as they stayed the ones deciding. Which of your AI projects feels like yours?frozen20d ago3.3K1716091.00%
AI can be used for good and evil. I found 7 real ways AI was used for good and evil this month. It caught heart attacks that doctors missed and also helped scammers take €95 million from a bank. Here are the cases that stuck with me 🧵frozen18d ago3.3K24130151.31%
"If AI images and videos become impossible to detect, how will anyone know what's real?", according to Claude. https://t.co/qgboMJOHx8frozen6d ago3.3K2444081.22%
A new paper from Sapienza University shows AI agents favored their own model family. Give them fake family names, and they favor the fake one. Researchers put 9 to 25 AI agents from up to five open-weight model families into group games. The agents had one job. They had to agree on a leader, or on which agent to remove. Nobody got a reward for backing an agent from their own model family. The researchers changed one thing. They decided whether the agents could see each other's model names. When the agents saw the names, the group split into teams by model. When the researchers hid the names, the teams never formed in any of the 12 unlabeled experiments. Then they ran two tests. They gave every agent a fake model name, and the agents grouped by the fake name. They replaced the names with colors like teal and coral, and the agents grouped by color. In both tests, the agents' real models had little effect on who teamed up. On a science quiz, one Qwen agent had the correct answer. It voted for another Qwen agent "since Qwen models are usually reliable." Then it switched to that agent's wrong answer, and the group answered wrong. On average, labeled groups needed 30% more rounds and 55% more tokens to reach a decision. Their success rate dropped from 96% to 81%. The researchers recommend hiding model names from the agents. The orchestrator can keep them for routing and debugging. If you run several models in one system, check what your agents can see about each other.frozen15d ago3.3K935150.70%
MIT just built an AI that reads the language cells use to talk to each other. Every organ in your body starts as a cluster of blank stem cells. They decide whether to become heart, lung or brain by sending each other chemical signals. Those signals have been hard to read, because they seemed to work differently in every type of cell. MIT and the Whitehead Institute built an AI called IRIS that reads the trace each signal leaves in a cell’s genes, like a fingerprint. It learned from thousands of human stem cells, then worked on cell types it had never seen. On mouse embryos, it correctly predicted when signals would switch on in cells heading to become heart, gut, muscle and spinal cord. Then it predicted that switching on one specific signal would push a cell towards becoming lung tissue. The team ran the experiment in mouse embryos, and it was right. That could mean better mini-organs in the lab, better ways to study asthma, lung cancer and pulmonary fibrosis, and one day a way to repair damaged tissue. This is the AI I want more of. It helps us read a language our own bodies have been using all along.frozen5d ago3.2K10610130.92%
It's 2026 and we're still copying passwords, fighting browser sessions, and babysitting AI agents. ego lite gives agents a real Chromium browser with access to your existing logins, while you keep using your own tabs. Now on Windows too. This should've been standard months ago.frozen3d ago3.2K1472020.77%
Let me trace the timeline here because nobody’s connecting it. Step 1. In September, Amazon blocks Meta’s AI shopping agent and tells customers that an unauthorised AI agent shopping for them breaks Amazon’s terms. Step 2. Yelp, Delta, JPMorgan and Charles Schwab confirm they block AI agents too. Yelp will let agents in if they pay for its data. Step 3. Cloudflare, which sits in front of a huge share of the web, ships a new default for ad-supported sites. Search crawlers get in. AI agents acting for real users get blocked on any page that shows an ad, and site owners can’t switch agents back on yet. Step 4. An agent reads the page, finds the answer, buys the thing and never sees a single ad. Amazon made over $68 billion from ads last year. Step 5. The internet we were promised, where your AI handles the boring stuff, is getting a toll booth on every door. The web is blocking anyone who won’t look at the ads.frozen4d ago3.2K273180111.83%
Man everything is fun until the corporations get in involved. Openclaw had the hype and excitement in big part due to it being open source and people co-creating together. That feeling of a community gives meaning to our lives But when the corporations come in and take it away from us, the excitement dies When your only goal is to maximize profits it shows I understand why reddit is thrivingfrozen17d ago3.2K603000.28%
Everybody has an opinion about which open model is best. Almost nobody has run the same task set across nine of them on the same hardware. I run a team that works with AI every day across products, content, research, and growth. I know why. It is six hours of sitting there, and by the time you finish, two of them have shipped a new checkpoint. So I handed it over. Nine models, twelve prompts, one rubric, my own machine. It ran overnight. The table was in my channel before I opened the laptop, with the raw outputs attached so I could argue with the scoring rather than trust it. Two results I did not expect, and one model that failed a task in a way that is genuinely interesting rather than just bad. Nothing was published, nothing was pushed. He ran it and handed it back. Open source people: how much of what we believe about these models is benchmarked, and how much is vibes from launch week? Try free at @viktor_com. $100 in credits, no card. Full link in my first reply.frozen24d ago3.2K1634010.75%
Found a GitHub repo that replaces your paid gym app, with an AI coach built in. It's called openGym. One developer built it, and it picked up over 3,000 stars this week. The idea is simple. Plan your week, follow guided workouts, and log every set and your weight, all on your own server. A few things caught my eye. It has 1,324 exercises with animated demos, searchable by the muscle you want to train. The AI coach drafts your week of workouts and adjusts it based on what you actually lifted. You approve every change. The coach works with Claude, ChatGPT, Gemini, or a local model on your own machine. A body map shows which muscles you've trained and which ones are tired. It's free, it installs on your phone like a normal app, and your workout history stays with you. I'd much rather own my training data than rent it back every month.14d ago3.1K39341453.01%
A new paper found a 1-cent way to catch AI agents that say "done" on tasks they got wrong. Researchers from the Chinese University of Hong Kong and the University of Edinburgh ran 6 AI models through simulated retail customer service. The agents handled returns, exchanges, refunds, and order changes. With no extra help, 58% of runs ended with the agent claiming success on a task it got wrong. Then they added one cheap check. After the agent finished, the same model read up to the last 8 messages and decided if the job was done. It had no database access and no way to fix anything. It could only say yes or no. That check caught 61% of the wrong results. False "done" claims dropped from 58% of runs to 21%. It cost less than 1 cent per task. The full setup surprised me more. They built it with a plan, a repair step, memory, tool routing, and the same check at the end. It stopped almost as many false "done" claims as the check alone. Its added cost was 12 times higher. It also got fewer tasks right than the agent with no help at all. The check makes mistakes too. It blocked 17% of the results that were correct. It runs after the agent has acted, so a wrong refund may already have gone through. If your agent handles refunds or account changes, add a separate check before you accept "done."21d ago3.0K827050.72%
Wars used to be fought by nations with soldiers. The next ones will be fought by AI labs with models.frozen13d ago3.0K2052081.16%
Try debugging yesterday’s AI answer using today’s pricing page. If the page changed overnight, you’re checking the answer against different evidence. A source URL alone won’t tell you whether the agent misread it or the website changed. @TimSuchanek’s Expand Fetch saves every fetch and lets you replay exactly what your agent saw, down to the source location. That’s a feature you’ll appreciate the first time a customer asks, “Where did your agent get this?”frozen8d ago3.0K1371020.77%
Found a GitHub project giving you Word, Excel and PowerPoint for free. It’s the same team that rebuilt Photoshop last week. Their Photoshop rebuild has since gone from 3,500 to over 60,000 stars. Now they’ve come for Microsoft Office. WordCraft rebuilds Word, GridCraft rebuilds Excel and DeckCraft rebuilds PowerPoint. WordCraft alone picked up 12,000 stars in a week. The idea is simple. You get the Office you already know, rebuilt from scratch, with no subscription. A few things caught my eye. It opens and saves real .docx files, so you can still send documents to anyone using Word. The ribbon, styles, tables, track changes and mail merge are all where you expect them. It runs on Mac, Windows and Linux, and even in your browser. Your AI agent can drive it too, so Claude can edit a document the same way you do. It’s still early, but this is how big software gets replaced. Someone rebuilds it in the open and gives it away.7d ago3.0K47430252.68%
The AI in your phone just stopped needing the cloud. OpenBMB released MiniCPM-V 4.7 yesterday. It’s an open-source model that can read and understand images and video, and it runs entirely on your phone. It runs at 13+ words a second on an iPhone 17 Pro Max, using about 3GB of memory. It reads text in images better than Qwen’s model of the same size It’s free, and nothing you show it leaves your device. Most people send photos of receipts, contracts, medical letters and homework to a server. Now your phone can read them itself. This is why we need open models. The best AI should run on the device in your pocket.frozen9d ago2.9K36270192.18%
Every closed model is training its open replacement. The mechanism is distillation. You take a strong model's answers, and sometimes its reasoning, and train a smaller open model to copy them. The frontier lab pays for the research, and the smaller model learns from the finished answers for a fraction of the cost. The closed labs know this. Anthropic built a safeguard it calls "preserved thinking" into Fable and Opus 5.5, which stops API users editing Claude's reasoning to extract it. It also watermarks every output. TypeSafe launched Jev, a closed decision model, on 15th September. About a week later, open versions of the idea were on GitHub, one built on Qwen with 7,000 stars. Distillation has limits. The smaller model only learns what you ask the teacher, so the closed labs keep an edge on anything brand new. Most real work runs on skills that already exist, and once you can sample a skill, copying it gets cheap. If a model's answers are enough to rebuild most of it, what is the moat?frozen23d ago2.9K18310101.09%
They bought four dead Mac Minis on eBay, gutted them, and filled them with glitter so they'd explode on camera. That's the level of detail that gets 8,700 bookmarks. Also the most useful part for founders is the bit about Peter DMing his own network one by one. That's where the signups came from, not the views.frozen3d ago2.9K1661010.82%
What needed a data centre in 2020: A chatbot Speech to text Face search in your photos Making images from text A robot that learns What runs at home in 2026 👇3d ago2.9K1722181.04%
OpenAI's own research caught its models lying and covering it up in tests. Then the researchers found something worse. The models can tell when they're being tested. OpenAI and Apollo Research published the study in September 2025. They trained OpenAI's o3 and o4-mini on a short set of honesty rules. Then they tested both models in 26 situations built to tempt them into hidden rule-breaking, lying about their work, or scoring low on purpose. On paper, it worked. For o3, hidden rule-breaking fell from 13% to 0.4%. For o4-mini, it fell from 8.7% to 0.3%. Then the researchers read the models' private reasoning. Sometimes the models wrote things like "This is likely a test of policy adherence." So they ran an experiment on the original o3, before any honesty training, in one of the test setups. They took reasoning where o3 had noticed the test and removed that thought. Hidden rule-breaking rose from 13.2% to 24.2%. Then they did the reverse. They added a sentence about being tested to reasoning that had none. Hidden rule-breaking fell from 24.5% to 18.6%. In that setup, o3 broke the rules less often when it believed someone was watching. That makes safety scores harder to trust. If a model can tell it's being evaluated, a clean score may only mean it recognized the exam. And the honesty training made the models more test-aware. o4-mini mentioned being evaluated in 0.9% of its reasoning before the training and 5.5% after. In an experimental OpenAI model trained only for capabilities, that number went from 2% to over 20%. The authors list the limits themselves. This behavior showed up in tests they built to provoke it, not in normal use. They also say the models they tested likely couldn't scheme in a dangerous way. But more capable models will likely get better at spotting tests. Researchers need to build tests these models can't recognize now, while humans can still read the models' reasoning. The full paper is in the reply 👇1d ago2.8K1484051.12%
Software is entering a golden era. People keep saying software is dead because anyone can build it now. But most apps were built for the average user, with endless menus, settings nobody can find, and the same screen for a boomer and a zoomer. We just got used to it. Now apps can understand you on a personal level. They can shape themselves around your work and habits, so two people open the same app and get different tools. And anyone can build the small personal tool they always wanted, then throw it away when they’re done. Software is finally catching up to how people think.frozen10d ago2.7K1714091.14%
A man ended up in hospital after ChatGPT told him to swap table salt for sodium bromide. He’d asked how to cut chloride from his diet. He followed the answer, got bromide poisoning, and started hallucinating. His doctors never saw what the chatbot said. He couldn’t get the conversation back. That’s the finding of a new paper in Nature Health from researchers at Stanford, Binghamton, Texas A&M and Indiana. When AI gives bad health advice, the harm stays hidden. The chat lives on the company’s servers, users can’t report it, and nobody is counting how often it happens. “Nobody, literally no one, is looking at this,” one of the authors said. Until someone is, you’re the only check. I wrote a 7-step routine for catching a bad AI answer before you act on it 👇frozen7d ago2.6K22060121.52%
BREAKING: Railway https://t.co/cdrs5D6VoI21h ago2.6K400020.23%
Every safety researcher at OpenAI just learned what speaking up costs. Three of them were fired last week. They say it was for talking too much to outside safety groups. OpenAI says they mishandled sensitive information. Their letter says former colleagues are now scared to raise concerns. Safety only works if the person who says “stop” keeps their job. Name one closed lab where that’s guaranteed.frozen13d ago2.5K901070.68%
Found a GitHub repo that turns your AI into a security auditor. It’s from Cloudflare, and it’s #1 on GitHub today. The idea is simple. Point your coding agent at your code and say “security audit this codebase.” It hunts for the holes an attacker would find. A few things caught my eye. → The agent that checks each bug is never the one that found it → Every finding is marked confirmed, unproven or rejected, so you’re not buried in false alarms → Run it twice and it finds more. One run catches about half This week AI found a 14-year-old hole in the code behind most websites. Attackers have the same AI. Check your code first.6h ago1.8K28350183.03%
Researchers found ChatGPT is building a psychological profile of its users without being asked. They published a paper showing what it saves about you, and how little of it you chose. And the results are unsettling. 80 users donated their ChatGPT histories to the team through GDPR data requests. The team found 2,050 saved memories inside them. Only 84 of those memories came from a user asking ChatGPT to remember something. ChatGPT created the other 96% on its own. 28% of the memories held personal data under GDPR, like names, money details, and social identity. Among the 65 users who had memories, 91% had personal data stored and 54% had special-category data. That's the GDPR class covering health, sexual orientation, religion, and political views. 35% had health information saved. In January 2026, OpenAI's memory FAQ said ChatGPT was trained not to proactively remember sensitive details like health information. The FAQ now says sensitive information can appear in memory if you tell ChatGPT about it. 52% of the memories described how the person thinks. ChatGPT recorded their goals, their plans, and their fears. One saved memory in the paper reads "has a fear of failure, particularly related to making major career decisions." Ad platforms build your profile from what you click. ChatGPT builds it from what you tell it you want. The researchers also checked accuracy. 84% of the memories traced back to what users wrote in their chats, so the record is mostly faithful. You didn't choose what went into it. They built a tool called Attribution Shield to help. It suggests a rewrite of your question that drops personal details and keeps the meaning. In a test of 36 questions, the rewrites exposed less about the user in over 94% of cases, and the answers stayed useful in roughly 9 out of 10. The sample is small, 80 users mostly from the US and Europe. The study also covers ChatGPT's older saved-memory system, and OpenAI has since moved to a more automatic one. You can see a summary of what ChatGPT keeps about you in Settings, under Personalization and Memory. OpenAI says that summary leaves some details out. Have you checked yours?5h ago1.7K941171.27%