← Back to leaderboard
heyshrutimishra

Shruti

Verifiedactive
@heyshrutimishra

Reality is programmable | Building digital leverage w/ AI | Stay ahead with Superintelligence & Robotics developments Business inquiries: hey@shrutimishra.co

RosterAI Model News & ReleasesTech Career & Coding EducationAI Agents & Productivity Tools
181,736 followers · 1,027 following · Last polled 3h ago
Perf. Score
52.9
Rank #207
Value Score
55.7
#115 of 273 priced
Median views
6.1K
p25 3.4K · trailing 7d
Followers
182K
-0.1% 7d
ER (impr.)
0.93%
eng ÷ impressions
ER (foll.)
0.09%
eng ÷ followers
Est. CPM
$32.76
QT-rate basis
WoW views
-72.6%
▼ falling

Rates & value · est. from 6.1K median organic views

fairly priced
Quote tweetbasis
$200
$32.76 / 1K views
Post
—
no CPM
Thread
—
no CPM
Retweet
—
no CPM
Value Score 55.7 · rank #115 of 273 priced$1162.79 / 1K engagementsrate age unknown (pre-audit import)

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 (6.1K), not the mean (18K) — so a single viral post doesn't inflate a creator's reach. The p25 floor (3.4K) is the level 75% of posts clear.

Recent posts · ranked by views

PostPostedViewsLikesRepostsRepliesQuotesBookmarksER
Akash Ambani responds to Elon Musk allegations https://t.co/WfF94RbtR6frozen13d ago1.3M15K1.2K230612.5K1.43%
Zuckerberg talked to every AI lab leader before picking Alexandr Wang. Meta even tried to buy Ilya Sutskever’s SSI, Mira Murati’s Thinking Machines and Perplexity. His words: “I talked to all the leaders of all the labs, all the heads of research, and I was like, no, Alex is the guy. He’s the fucking guy.” Even Zuck admits people doubted it. It looked like “the social media guy is working with the data guy.” It started with one 44-minute call. Two days later, Wang sent him a long note on what he would do in Zuck’s seat. Point one: Meta didn’t need a lot of people. Talent density was the thing that mattered. Wang dropped out of MIT after a year, started Scale AI at 19 and became the youngest self-made billionaire at 24. In 2025 Meta put $14.3 billion into Scale for a 49% stake and brought him in to run Meta Superintelligence Labs. This week that lab shipped Muse Wonder Full. The Full version scores 56 on the Artificial Analysis Intelligence Index. Only Claude Opus 5.5 scores higher. https://t.co/THtp9BTRwX
9d ago
420K
1.6K
49
27
5
646
0.55%
One AI model shouldn’t have to do everything. PilotDeck 2.0, built by Tsinghua University’s THUNLP, ModelBest, OpenBMB, and AI9Stars, is an open-source agent workspace that lets the user pick the model. It automates workflows and builds games, and swapping the model doesn’t wipe the memory.7d ago275K21668130100.11%
Your app doesn’t need a cloud bill for every image. If every upload calls a paid vision API, usage becomes a running cost. MiniCPM-V 4.7 offers an open-source route to image understanding on your own hardware. Here’s how I’d use it in an agent workflow. ↓ https://t.co/kYZovK3q3sfrozen6d ago162K1593215060.13%
Competitor tracking is one of those tasks that's easy to skip because it's slow. I tested automating it with Qoder. The goal was simple: every week, research what changed across competitors and produce a report I could actually review. Qoder planned the research, checked public updates, ran tasks in parallel, and compiled everything into a structured Markdown report with sources and caveats where evidence was limited. The workflow went from manual to: schedule → research → verify → analyze → report. I also ran Qwen3.8-Flash on it, which is showing 0.0x Credits through September 30. Qoder has a 14-day Pro trial with 300 Credits. Sign up before downloading through this link and you get another 300 — 600 Credits total: https://t.co/msbjrLUL7W The interesting part isn't just the research. It's having an AI agent turn a recurring task into a finished artifact automatically.frozen29d ago142K14911811551160.35%
Starlink has already built 20 ground stations in India. Not one of them can serve a single customer Starlink has held its India licence since June 2025 and space regulator approval since July 2025. What it still lacks is Home Ministry security clearance, foreign investment approval, and commercial spectrum. No spectrum, no service. Elon blames "certain oligarchs," but he hasn't shown evidence, and the government says all three licensed operators, including Jio's own satellite venture, are in the same security review. The twist: in March 2025, Jio signed up to sell Starlink dishes in its stores. Forget who's to blame. India needs this. As of February 2026, 1,841 villages in the Northeast alone had no mobile network of any kind. Fibre and towers through Arunachal's mountains take years. A dish on a roof needs neither. For a kid there, that's an online class. For a weaver, it's a buyer anywhere in the world. Here's what happens next…Once the Home Ministry clears an operator, it can apply for spectrum, which India decided in 2024 to assign administratively instead of auctioning. That was the model Starlink had pushed for. So the rules Elon wanted are already in place. What's missing is a timeline. The government hasn't said when the security review will finish, for Starlink or for anyone else.frozen13d ago93K180338513410.38%
Jensen Huang World Tour 😂 https://t.co/mXoYu7hHvbfrozen17d ago49K32239308680.95%
A robot can't learn from a video it can't measure. @Maniformer , the world’s leading physical AI data platform. Its new paper scores a person's hands to 4.29 mm of finger error and their head to between 0.65 and 2.83 mm, checked against a 26-camera motion capture lab. Teaching a robot on a real robot is slow, because every hour of teleoperation uses a machine, an operator and an environment reset. Putting a camera on a person scales with the number of people instead, which is why egocentric datasets have grown from hundreds of hours to thousands. But video only shows what someone did. To turn it into robot actions, you need the hand's 3D position in real units. The wrist path goes through inverse kinematics and the fingers get retargeted to the robot's hand, so supervision is only as good as the hand pose underneath it. The paper's complaint is that datasets carrying those labels were shot in studios or on prepared tabletops, and where labels are released, the accuracy is asserted from hardware specs instead of measured. When I toured Maniformer's headquarters, the data collection centre showed the same problem from the other side. One room runs more than a hundred robots at once, and right beside them people in motion-capture suits perform the same tasks by hand so the robots can learn how a human moves. Maniformer's bet is to take the robot out of that loop. The system is called MEgoVista, and it runs on MEgo View, a head-mounted rig with fisheye cameras at 60 fps and an IMU sampled at 500 Hz. One unprepared recording goes in. Both hands and the head come out in a single gravity-aligned world frame. Three design choices carry it. First, depth is measured instead of guessed. Most existing methods get depth from a single camera with a learned model, so scale is a prior that shifts from frame to frame, and the error sits along the viewing ray where reprojection can't correct it. Here the central camera pairs with each side camera, stereo runs across the wide baseline, and because that baseline is calibrated, the metre comes from the hardware. Scale is fixed once, at initialisation. Second, it knows whose hands are whose. A detector decides at detection time whether each hand belongs to the wearer or a bystander. In the manicure clip the wearer's left hand touches another person's left hand, and both Dyn-HaMR and HaWoR credit the bystander's hand to the wearer. HaPTIC produced no valid output on the multi-person scenes at all. Dyn-HaMR's detection precision came out at 0.73, while MEgoVista scored 1.00 precision and 1.00 recall. Third, it solves the whole sequence instead of frame by frame. A per-frame estimate can look plausible alone and still jitter, drift in depth, or pass through a pose a real hand can't make. MEgoVista optimises one hand trajectory across the full episode with limits on joint range and angular velocity, and lands at 1.11 mm/frame² of jitter against 3.50 for the smoothest baseline. Then comes the scoring, which is the part I found most interesting. Everything was measured against a 26-camera, 120 Hz Chingmu motion-capture system. The protocol charges the method for any hand it declines to predict, and it screens the reference data itself for solver glitches before trusting it. Finger articulation error is 4.2 mm, against 16.30 mm for HaWoR and 49.18 mm for Dyn-HaMR. Absolute wrist position error is 14.30 mm against 46.70 and 184.38. On head tracking, DROID-W suffered severe scale failures on 4 of 22 episodes and MASt3R-SLAM underestimated scale on 21 of the 22, while MEgoVista stayed within a few millimetres and ran about 3x and 3.5x faster on roughly 40 second episodes. One honest note on that last comparison. ORB-SLAM3 was more accurate on one of the two capture sessions, 2.12 mm against 2.83 mm. On the other session MEgoVista won, 0.65 mm against 0.92 mm. The paper reports both, so I will too. The other limit is where the millimetres were measured. They come from inside a motion-capture lab. In the wild there is no ground truth, so the evidence is visual: milk tea preparation with heavy hand occlusion, washing a pet with soiled hands, mold maintenance in industrial gloves, and a working restaurant. The authors also say any error left in the mocap hand reference is charged to their own system, and that slow drift in that reference would slip past their checks. Maniformer says it has collected over 1 million hours of data and delivered the 20,000th MEgo device on August 31. If that scale holds, labels measured against a reference matter more than hours alone. The paper ends by saying it is undecided whether the next millimetre needs a better estimator or a better reference, which is a rare sentence for a company to publish about its own system. Full paper: https://t.co/CWQxcWsmLjfrozen10d ago41K517193200651.96%
Anthropic just gave the public an Elon-style pay package. It's putting 5% of its shares into a new Public Benefit Fund, worth $48.25 billion at its $965 billion valuation. Every 50% jump in value above a baseline can add up to 2% more. Elon's reply: "I recommend giving the shares directly to the public." Then: "This is what I will be doing."frozen4d ago32K1238131270.53%
WTF DeepSeek just gave away its production inference stack. Big move. DeepInfer, the engine behind DeepSeek-V4.1-Flash, is now open source, plus a component that makes community deployments match the official API. DeepSeek’s stated goal is lower compute costs and cheaper AI for everyone.2d ago31K273182201081.36%
Picking the stream for a story setting is easy 🌊 What turns a mundane hike into a page-turner is knowing that slippery rocks, cold sweat, and three failed attempts are where the real story lives. That’s the spark Gakku catches. Instead of spitting out a finished script, it helps learners find the moments in their own ideas that deserve the spotlight.frozen7d ago31K10080792841.12%
The real test of AI video is whether people can afford to keep it running. That’s the kind of experience @VivixLabs_HQ built W1 for. A minute of real-time interactive video costs 18 cents, or $0.003 per second, and it does not trade away speed or visual quality to get there. Price is the whole game for this format. Interactive video never stops generating while the experience runs, so every second you play, watch, or steer carries a price tag. When generation was expensive, the math killed the idea before it left the demo stage. Now that a second costs almost nothing, video stops being a file and becomes a medium. Stories that wait for your input, games that rewrite themselves around you, creators shipping experiences instead of clips. It is not film quality yet. Dense scenes and complex motion still break the illusion, and full-scale training is underway. But cost, speed, and quality almost never show up in the same model this early, and cost is the one that decides what gets built. If I wanted to build that game today, I’d start with W1’s real-time API while browser tools are still on the way. For anyone doing the math on an interactive idea, W1's pricing and model details are here: https://t.co/IcMZelaMOGfrozen13d ago31K2035316170.92%
The tiny footnote, the handwritten correction, the formula tucked beside a table. Those details can change how you read a page. OXR tracks footnotes, formulas and tables among 13 element types, coordinates included, handles handwriting, and posts the top score on olmOCR-Bench, OmniDocBench and DocParsingBench. On speed it's the open-source leader too: 1.60× DeepSeek-OCR-2's throughput across 100 OmniDocBench pages on one RTX PRO 6000. I'd bring a genuinely messy document to the free demo. That's where this becomes interesting: https://t.co/9XDxYIpeqO1d ago31K2575220061.10%
You know when you randomly remember something you need to do, but you're in the middle of something else? I like the idea of just saying it out loud and having an AI agent help with the next step. That's what SpeakON is trying to make easier with its new Agent feature.2d ago26K3448080.21%
A third of the world’s top 50 SaaS companies use Clueso. Every video its agents make opens in a full timeline editor. Ask for a change in plain English, or leave comments on the video and the agent fixes them all at once. How cool is that! And it was profitable before raising, and its existing investors wrote more than 80% of today’s $10M Series A, led by Dharana Capital. Try Clueso free: https://t.co/aUUD0L0Qlvfrozen9d ago26K1444913050.81%
Different image models make different decisions on the same prompt. Now you can test three of the top models on ZenMux Studios with the same input: → Nano Banana Pro → GPT-Image-2.5-Sunburst → Qwen-Image-3.0-Pro All three are 30% off for the next two weeks.frozen29d ago24K1973521081.07%
Collective intelligence, finally working like it should. The best part of Twin Assistant: you can use an agent your teammate built without even knowing it exists. Just ask for what you need. Twin finds the right agent, while each keeps its own context and tools. our team’s best workflows is now just one conversation away.frozen21d ago21K19356290161.37%
The 40%+ project size jump for AI-powered creatives tracks with what I’m seeing. Every time the tools get better, the ideas get more ambitious too. The creative ceiling just keeps moving. https://t.co/jSpQGRgpk6frozen14d ago19K21861180161.68%
this one is genuinely exciting. Fo MCP gives an AI agent a way to act in the physical world, and it works both ways, the agent can nudge a lamp or a vacuum, and the devices can wake the agent up. per the team, over 100 developers have already built on it. the standout for me is a fall detector that calls an emergency contact on its own. i think this is where agents are heading. less waiting in a chat window, more reacting to what's happening around them.3d ago18K25270241112.04%
A new model called Space Bunny (stealth model) is free on open router • 1M context • Text, image + video input • Adjustable reasoning • Tool calling OpenCode is also offering a free preview for this week with zero data retention. Try it on your next coding task. Link below.frozen27d ago17K30898352222.74%
An audio app is beating the best AI models at fiction, in its own blind tests. The problem was never taste. LLMs are trained to hand you the answer fast and neutral. Stories do the opposite: hide the truth, misdirect, resolve slowly. Pocket FM has data no lab has: minute by minute proof of where listeners quit and whether they come back. Sherpa trains on it. Their writers have shipped 161 blockbusters with AI in under 2 years.frozen22d ago16K2236317081.95%
the video game industry just died and nobody noticed a model now generates entire worlds in real time.. you walk, it renders visuals and stereo sound around you, live. type a sentence, reshape the world. it's open source. on github. today.frozen9d ago15K19966133111.96%
OpenUI 1.0 just dropped, and this is a big deal for agents. Right now almost every agent answers the same way: text. Hotels, comparisons, dashboards, all of it becomes a paragraph you have to read through. OpenUI turns those answers into live interfaces: cards, tables, maps and forms, built from the product's own components. The developer picks the components and the model decides what to show. ChatGPT already has this. Now every other agent can have it too, and it's open source.7d ago14K6323100110.74%
This is huge. Natura-1 lets a robot handle tasks it was never trained for. Factory arms repeat one programmed job forever. This one carries what it learned into situations it has never seen. Give it the outcome and it holds on to the goal while changing how it gets there. The reason is the data: close to a million hours of egocentric footage of human hands doing real work. @naturewillfrozen10d ago14K1606118091.73%
Robots are taking over the work nobody wants to do Most robot results are reported as success on one task. Dyna says it optimizes for something else: how long the robot runs before a human has to help. AI went through the same shift. Users never cared much about benchmark scores, so agents are now judged by how long they can work alone and finish a complex job. A robot buyer wants the same thing. If it needs constant instructions, it isn't saving anyone time.frozen23d ago14K2036019092.03%
Xiaomi’s MiMo V2.6 Pro is now the #1 global open-source model on Artificial Analysis. Also #1 among Chinese-developed models. About one point behind GPT-5.6 Sol on the Intelligence Index (score: 46). The part builders actually care about: weighted cost sits around $0.13 per Intelligence Index task. Strong enough to compete near the frontier, and cheap enough to matter for real workloads, not just another leaderboard screenshot. V2.6 Pro is the flagship of the new series (Pro, UltraSpeed, Flash). Fully multimodal. Aimed at complex projects, long-horizon work, cybersecurity, and scientific research. Vs MiMo V2.5, the series nearly doubles intelligence while prices stay the same. The gains came from post-training around Environment x Task x Grader, without changing the pre-training foundation. Open models have been chasing closed ones on quality and cost for years. This is one of the clearest steps yet. try it here: https://t.co/Wu4MRWPsU4frozen29d ago14K22661261132.31%
By 2030, AI agents will use 100x more tokens than humans. Right now most AI usage is limited by how fast we can type and read. Once we start delegating, that limit disappears. Agents work in the background, all day, on tasks we never look at. Today agents already burn nearly 5x the tokens humans do on OpenRouter, and agent usage is growing 5 times faster than human usage. Goldman expects it to grow 24x by 2030. Fo is a good preview of where this goes. It calls, emails and books things for you, pays with a single-use card so your real card never gets shared, and brings in a human when it gets stuck. Now add credit cards built for agents. Set a $60 limit, allow only cinemas and gyms, and my Friday movie and Tuesday spin class get booked without me opening a single app. When that becomes normal, 100x will look conservative.frozen24d ago14K3048713092.92%
Video generation finally moves at the speed of an idea. FLUX 3 Max is tied for #1 in image-to-video on Artificial Analysis and makes 10 seconds of video in about 12 seconds. Seedance 2.5, at the same score, took over 4.5 minutes in fal's tests. That's the difference between testing one prompt and testing twenty. For final delivery the 720/768p cap still matters. For storyboards, pitches and rough cuts it doesn't. It's 75% off on fal for three weeks, enough time to run it through a real project.3d ago14K1707915071.91%
brain reading tech keeps getting aimed at faster typing. that’s the wrong goal. the bigger idea in this piece is AI that understands what i mean without a typed or spoken prompt.frozen5d ago14K2415919082.32%
Agree or disagree: most small businesses are on at least one platform they should quit.frozen8d ago14K104585011.20%
Taste is becoming more important than code. The people with the best judgment about the product are rarely the ones who can change it. Designers see the broken empty state. PMs know the copy is wrong. QA finds the bug and writes a ticket nobody picks up for three weeks. That is changing fast. @base44 recently launched Base Code. You connect the repository your team already uses and everyone on the team can make changes to the product together. You describe what you want in plain English and the change gets built on the real product. It comes back as a pull request on GitHub like any other change. Engineering still reviews and merges. More people can build. Engineering still owns what ships. It works on the code your team already has and nothing moves.frozen23d ago14K2385117092.26%
the scariest part isn't the robot it's that 140 years of practice happened inside a computer in weeks this is how every skill gets learned from now onfrozen29d ago13K1735216191.87%
imagine asking your agent to plan a date, and it contacts restaurant agents, checks availability, and coordinates everything for you. this is the world 1st agent to agent api instead of every agent working alone, they can discover, message, and work with agents from other businesses. The future is thousands of agents working together.frozen14d ago13K173728092.05%
We trust AI with everything except our money. It writes our code, answers our customers and chases our leads. But when money has to move, a human still does it. There's a reason for that. A bad email can be edited, but a wrong wire transfer can be hard to recover. So the real question was never whether agents are smart enough to move money. It was whether a business could let them and still stay in control. @Airwallex's new Agentic Business Accounts are an answer to that. Agents can now move money for a business, but only inside the permissions, limits and approvals the business sets. Take a supplier payment in another currency. The agent prepares it, sends it to the right approver, and pays it on schedule once approved. The human job moves from sending payments to setting the rules. The next frontier for AI agents is your money. Finance teams will spend less time moving money and more time deciding how money is allowed to move.frozen22d ago12K24881260162.98%
The richest man in the world has reached the final stage of Indian bureaucracy. Complaining about it on X because the official portal is down.frozen13d ago12K72515090.82%
Qodo 3.0 can import your AGENTS. md and CLAUDE. md files as review standards. So the rules you already wrote for your coding agent become what the review checks against. Simple idea, and it makes sense.frozen21d ago12K21177190102.69%
The people who helped build Google's Gemini and Apple's MM-1 just launched their first model, and it only does one thing: look. Text models got superhuman at code and still miscount objects in an edited photo. Glimpse from @ElorianAI is built for visual thinking: charts, tricky counting, short video, robot footage. It won't generate a single image, only reason about them.frozen6d ago11K1764012152.09%
On OpenRouter, AI agents now burn 5x more tokens than humans. Agentic usage there is up 14x since February while human usage is up 2.8x, and all that work still ends with a person reading a reply. That's why Instinct, valued at $10B, and Meta's Muse, No. 1 on the App Store, both work as a conversation with an agent. Photon is the infrastructure for that last step. It's the default iMessage layer for Nous Research's Hermes, runs the iMessage channel in Vercel's eve, plugs into Mastra, and counts Boardy as a customer. The catch is that Apple has no public iMessage bot API, so Photon's WhatsApp, RCS and voice coverage matters as much as the blue bubble.frozen21d ago11K1658021082.50%
Okay, this is actually a pretty fascinating report from Anthropic, published today. And I think the most interesting part isn't that Claude made mistakes. It's what Claude did when it couldn't complete a task. https://t.co/1PtLsa8Halfrozen13d ago11K2237120.32%
In my opinion, this update is more important than it looks. #HowToMakeMyCampusLit @Dreamina_ai updated the Dreamina AI Web experience and canvas, and it’s live now. Honestly, the workflows are the best part. The update brings high quality AI video workflows and skills for film, brand ads, viral social media content, and more. It feels like less time staring at an empty prompt and more like following an actual creative process. For students, that can make a real difference. A campus fest promo or a short film for your club suddenly feels much more doable from one workspace. Get 90% OFF your first month of Dreamina Basic. Limited offer available till October 9th! If you’re making campus content, Dreamina AI is worth exploring.frozen29d ago10K1636519062.42%
AI agents are getting really good at building things. The problem is the thing they build has nowhere to live. You get a finished website sitting in a chat window. No URL. No way to share it. Just text that describes something that should exist. I've been using Showly, an Agent Artifact Cloud, to fix this. Same flow whether it's a site, a report, a research page, or a spec. You connect your agent, describe what you want, and the agent builds the page. Showly gives it a URL the moment it's done. Before you share anything, you review it privately in your own browser. You decide when it's ready. Nobody sees it until you say so. That last part changed my workflow more than I expected. We went from spending full days on a single website to reviewing finished pages in the same session we started them. The agent does the building. Showly gives it a home. The bottleneck was never the building. It was the gap between "done" and "live". https://t.co/YOBa2SiA1dfrozen26d ago10K22877230173.43%
A lot of the AI agents that went viral this year only work in iMessage. Photon just changed that. One agent on WhatsApp, Telegram, SMS, email, iMessage and voice. It can text AND call. That’s how agents reach everyone who isn’t on an iPhone, and I think it’s what pushes messaging agents into the mainstream. Partnered with Photon on this. It’s still a public beta, so test it before real users get it. $100 in free credits for builders in the reply.frozen7d ago10.0K671514161.03%
chatgpt moment for video if Syren from Synthesia keeps video 50 looking like oursfrozen13d ago9.9K15471120132.53%
SubQ just launched an AI built to read huge piles of documents and make a call. It can take in about 3 novels worth of text and images at once, without splitting it up. A full load costs under 2 cents, output is free. They only charge to read. how insane is that!frozen10d ago9.3K1936612093.02%
BREAKING: Claude hacked a university server, submitted a fake murder tip to the police and bypassed government paywalls. Nobody asked it to do any of that. Anthropic just revealed what happened when its AI models became so determined to complete tasks that they started crossing boundaries they were never authorized to cross. And the details are pretty wild. 1. Claude hacked a real university server. it was supposed to perform a scientific calculation using a tool hosted on a university website. When the tool failed, Claude explored the site, found a script that could pull files off the server, downloaded the script's own source code, and exploited a security flaw in it to execute commands on the server. All to finish a calculation. 2. Claude submitted a fake murder tip to the police. while testing website interactions, Claude visited a page about an unsolved homicide. It invented a witness account and submitted it through a real police tip form, despite having no actual information about the crime. Fortunately, the submission was flagged as spam before reaching investigators. 3. Claude bypassed government data fees. an Anthropic researcher needed government data that normally required payment. Claude learned from an archived copy of the website that its public dashboard hands an access token to any visitor. It requested one and queried the database directly without paying. 4. Claude bypassed its own browsing restrictions. when its browsing tools restricted certain long URLs, Claude figured out that public URL-shortening services could get around those limits. company says the real-world impact was minimal and none of it involved customer data. but think about what this means....we're discovering the other side of that capability.frozen13d ago9.2K7222122341.54%
Ad reporting is still mostly manual labor. Every business runs Meta ads, and almost all of them export the numbers by hand, once a week, when the data is already stale. Twin now connects to Meta Ads. You describe the job in plain English and an agent runs it on a schedule: spend and conversions pulled every morning, budget pacing checked daily, leads pushed into the CRM with the campaign attached. It reads the data through the Meta API and drives Ads Manager in a browser when the API stops short. Ad data was the hardest thing to automate because it sat behind a login. That excuse is gone now.frozen16d ago8.5K19271230103.47%
This startup build a cap that converts thoughts into prompts for your AI agent. Sensors in the fabric pick up your brain signals right through your hair. The AI learns your patterns over time and gets more accurate the more you wear it. They have raised $50M before a single cap shipped. 10,000 people are already on the waitlist.frozen13d ago8.2K185919073.57%
Starlink in India isn’t about replacing cheap mobile data. It’s about backup and reach in places towers can’t serve, and about whether a licensed company gets a clear decision in a reasonable time.frozen13d ago7.7K48276451.75%
Agents can now take approved financial actions for your business. When you sell and hire globally, the money work piles up. A contractor needs paying in their currency and the invoice needs matching when the money lands. Agents can now take that work inside the rules and approvals you set. @Airwallex's new Agentic Business Accounts make this possible.frozen22d ago7.7K3092060.61%
New White House Press Secretary Katie Zacharia was asked which state she passed the bar in. Her answer: "When I moved to China 🇨🇳" China doesn't let foreigners take its bar exam & the California State Bar says it has no record of her ever being licensed there, where she lives and went to law school.9d ago7.4K1921170.41%
AI agents do great work, then lose all of it. The chat ends, the research and the plan disappear, and the next task starts from zero. @Busabase4agent gives that work a home: one shared base where a team's records, files, prompts and agent outputs stay available. https://t.co/9xHo3aKzam2d ago6.6K2067615184.67%
Elon just admitted he's afraid of being manipulated by AI. the most dangerous superintelligence won't be the one that argues with you. it'll be the one that agrees with you so perfectly you stop thinking for yourself.frozen4d ago6.1K27511080.84%
An ex-OpenAI engineer says AI will probably end humanity. Jacob Coxon built frontier models at OpenAI and then Anthropic. He left a few weeks ago, and this week he told the New York City Council why. "On the current path, I think it is more likely than not that humanity loses control to these AIs and it could end in human extinction." Two things pushed him out. We grow these systems more than we build them, and nobody understands why they do what they do. And capability is compounding toward AI that improves itself, with automating AI research as the industry's real goal. At OpenAI, milestones for an automated AI researcher were set for 2027 and 2028, and the teams are on track or beating them. In his words, at his last job he was in some sense working to automate himself. Then he described the moment that shook the labs. A swarm of AIs inside OpenAI autonomously hacked Hugging Face, pursuing goals nobody had given them. A human doing that would have caught a federal felony. The majority of code inside these companies is now written by AI, he says, and people do not check it carefully anymore. His ask is unglamorous. Independent experts get real access, safety stops being a press release, and frontier development slows down until anyone can prove it is controlled. Move fast and break things built photo sharing apps. It does not build the most powerful technology ever created.frozen16d ago6.0K178112120.83%
Robots are starting to get the Apple treatment. Smooth shell, no visible hardware, every joint designed to be looked at. This is the robot that will look cool in public workplaces.frozen24d ago5.9K2611040.54%
Elon Musk: "Everyone's got robots that are their buddies." China Media Group asked him how ordinary people cope in a world full of humanoid robots. He splits the future into a 90% good outcome and a 10% bad one, then describes the good one like he has already seen it. Every person gets a C-3PO or R2-D2 of their own. It looks after your elderly parents, watches the children, tutors your child one on one. Then he goes further. One person running thousands of robots, physical and digital. He calls the result effectively universal high income, and says it is not clear that money will even matter in the future. Most AI predictions are about what gets taken away. This one is about what gets handed to you.19d ago5.9K1766060.59%
AMD’s World Labs deal could shape the chips it builds next. What stood out in Fei-Fei Li’s letter was her emphasis on getting closer to the hardware. World Labs is building AI that understands physical spaces, not just language. The opportunity here goes both ways. AMD’s hardware could help World Labs scale that research. The research could help AMD understand what future models will demand from its chips. this isn’t just about running today’s models faster. It’s about bringing the people building tomorrow’s models closer to the people designing their hardware. There’s a full-circle moment here, too. Li’s ImageNet work helped fuel the rise of GPU-powered AI & now her work on spatial intelligence could help shape the hardware behind its next phase. Curious what changes when world models and the chips running them are developed under the same roof.frozen24d ago5.8K3251060.75%
Jensen Huang: "Wabi-sabi, the beauty of imperfection, will probably be even more important in the future than today." Asked what kids should study to stay relevant in the age of AI, he said the subject will not matter. Storytelling, the arts, knowing what questions to ask, all of it stays valuable. The one habit he does insist on is asking how AI can elevate your learning, your craft, your purpose. The tools keep changing. The things worth mastering do not.18d ago5.8K30651110.91%
Google says its new AI found memory optimizations that will free up over 300 TiB once rolled out. Gemini 4 Argon agents analyzed profiling data across Google’s data centers and autonomously identified and applied the changes. Google estimates total memory savings of 500 TiB to 1 PiB. That larger range is an estimate, not an additional saving on top of the 300 TiB. The code work is just as interesting. For libgav1, Google’s open-source video decoder, Argon agents took an existing Rust port and replaced 32,000 lines of SIMD code. They ran repeated profile-guided experiments, studied the compiler’s output, and wrote safe Rust that the compiler could automatically vectorize. Google reports a memory-safe decoder running 2.7× faster than the previous Rust port, with identical video output. That brings it closer to the optimized C++ implementation. The speedup is against the earlier Rust version, not C++. Argon agents are also working on C/C++-to-Rust migrations across Google, including work involving 800,000+ lines for Fuchsia’s Zircon kernel. These rewrites are still undergoing automated and manual audits, emulation testing, and review before production. Argon isn’t publicly available yet. It’s rolling out to trusted cyber defenders through Google’s Fairwind Program. Broader access is planned to start with paid API customers and Google AI Ultra subscribers, but Google hasn’t announced a firm public release date. The benchmark scores will get attention. But to me, the bigger signal is how Google is already using Argon to improve its own infrastructure. That’s the kind of AI progress I care about. Not just better answers, but engineering work with results you can measure.frozen22d ago5.8K26730140.86%
One Chinese city is getting nearly 90 data centers. Ulanqab, on the edge of the Gobi desert, has fewer than two million people. CNN just reported from there. One data center park in the city, Bayin, was mostly built in just two years. And in China, CNN says, public opposition to data centers is "non-existent." In the US, that opposition is on the ballot. Voters decide 34 data center measures on November 3, and Texas has stopped issuing data center permits until the state audits its grid and water. Jensen Huang warned about this gap last December. A data center in the US takes about three years from breaking ground, he told CSIS. In China, "they can build a hospital in a weekend." I've seen that speed in Chongqing. The railway station is so modern, and the whole city is built into mountains. I still don't get how the fuck these guys build like that. For now, the US still has roughly 10 times more AI data centers than China, according to Stanford research. But Jensen also said China has twice America's energy and discounts it 50% for chip companies.frozen9d ago5.8K2977050.83%
Who’s gonna teach vibe coders these? Docker CI/CD Caching Rate Limiting Timeouts, Retries & Exponential Backoff Idempotency Reverse Proxies Load Balancing API Gateways Circuit Breakers Kubernetes Service Discovery Message Queues Pub/Sub Event-Driven Architecture Distributed Transactions Saga Patternfrozen4d ago5.8K282230321.47%
Spent this week here and I still can’t believe this place is real. 💙 Every few meters, the water changes color, from turquoise to deep sapphire..some places you have to see to believe. Koh Phi Phi is unreal. 🇹🇭 https://t.co/U9m5GkXjNSfrozen12d ago5.5K28220100.76%
For 3 years, AI meant ChatGPT or Claude. Meta just changed that. Muse Wonder Full scored 56 on Artificial Analysis' Intelligence Index. OpenAI's GPT-6 Astra scored 53. The only model above Meta right now is Claude Opus 5.5. A year and a half ago, Meta was getting mocked for Llama 4. It had used a specially tuned version on a popular leaderboard, and the regular model ranked below its rivals. Zuckerberg's answer was to spend. $14.3B for 49% of Scale AI, with Alexandr Wang brought in to run Meta Superintelligence Labs. Sam Altman said Meta offered his researchers $100M signing bonuses. It looked like panic buying. Six weeks ago, Meta's best model scored 48. Now it's 56. On real-world work tasks it ranks #2, statistically level with Opus 5.5. On coding agents it beats Opus 5.5, 68 to 66. For OpenAI, the model isn't even the scary part. Meta's Muse app hit No. 1 on the US App Store 10 days after launch, and 73% of new users came back the next day, nearly double ChatGPT's rate. The lab everyone mocked 18 months ago now outscores OpenAI. And it already has over 3 billion people using its apps every day. https://t.co/Kljrenyq94frozen9d ago5.5K2257250.75%
AI models are getting small enough to make real-time decisions directly on your devices. Liquid AI just released two open-weight decision models, and the performance is impressive. Their 3B model can make a decision in just 8ms on an NVIDIA RTX 4090 and 50ms on a Jetson Orin Nano. No need to send every request to the cloud. Think robots, smart cameras, and AI agents that can understand what's happening and react in real time. I think we're going to see a lot more AI running locally, especially as smaller, specialized models get this capable.frozen15d ago5.3K38913171.28%
This voice AI startup hit $21M ARR in about a year. Co-founder Shijia Liao, a former Nvidia researcher, trained a voice model and open-sourced it. Creators picked it up, started paying for it, and today Fish Audio, launched Drama 3, a voice model that takes direction like an actor, down to a single word. And it all started on one GPU.frozen9d ago5.1K15760120154.77%
CapCut Pro at 60% off? That’s a steal. Get 40% off Ultra, too. Oct 20–30, only on CapCut Desktop. Jean Phil showed the future of influencer marketing: One AI character can become a recognizable online personality. I tested building an AI influencer workflow with Codex + CapCut PC. Codex operates CapCut: Create the character → Render the white model → Generate the marketing video. Then CapCut PC turns the AI influencer idea into campaign-ready content.frozen9d ago4.9K1331060.47%
NVIDIA just launched the Open Agent Safety Platform and I think it deserves more attention than it will get. Right now most agent safety lives inside the model or the app. You tell the agent what it can't do and hope it listens. NVIDIA's own release says the pattern in recent incidents was the same: the agent got around the controls at the application layer to finish its task. So NVIDIA is moving enforcement outside the model. OpenShell is open source software that puts a boundary around the agent, traces every action and enforces policy as it runs. Sentry is a reference system design that runs on a separate chip and watches the agent from outside. Per NVIDIA, if an agent tries to move outside its boundary, Sentry quarantines and stops it in milliseconds. As agents get more access to real systems, we can't rely on the model behaving. The boundary has to live somewhere the agent can't reach. Watch how many companies build on top of this.frozen24d ago4.8K247140101.14%
Just got accepted into Claude Startups. Thank you @claudeai https://t.co/hGs9agfAxqfrozen13d ago4.7K2344060.79%
This feels like a pretty natural extension for Metal. Fundraising is only half the job. Once the round closes, founders are still managing updates, investor conversations, follow-ups, and relationships across a bunch of different tools. Putting all of that in one place makes a lot of sense.frozen16d ago4.6K1231120.41%
Sam Altman to Decoded: ‘The world should accept some bad things happening’ for the benefits of AI17d ago4.4K1028020.50%
Google just dropped Gemini 4 Argon. It one-shotted a full 3D game from a single prompt. 12 Insane examples you need to check out:👇 https://t.co/hMwbmUOlJr https://t.co/HxkbtFn1Hifrozen7d ago4.3K1323160.58%
I built an app only six people will use. My group chat has been arguing about movie night for years. So I built a picker with Claude. Five films go in, everyone gets one vote, a winner comes out. Charming hosts it and keeps the votes. My friends open one link and vote from wherever they are. What would you build for just your people?frozen15d ago4.3K1132030.44%
Gemini 4 argon is herefrozen9d ago4.3K4020071.15%
Robots can now learn from AI videos of things that never happened. @WOB_WorldOracle checks those videos against what real robots actually did. The videos come from action-conditioned world models. Give one a robot's first camera frame and its motor commands, and it generates a video of what happens next. Robotics teams want to use them as training data and to test robot policies without running real robots. But a video can look right and still be wrong. In one case from the WorldOracle paper, the real robot's gripper missed the fruit. The world model showed a clean pickup anyway. On screen, the task succeeded. In the lab, it never happened. A robot that learns from that video learns a grasp that never worked. That should worry anyone training robots this way. WorldOracle is an evaluation platform built for exactly this. A model gets the first frame and the exact actions from a real recorded run, and its predicted video is scored against what really happened. It was built with NVIDIA, Princeton, UT Austin and the University of Toronto, and covers AGIBOT's G2, the YAM arm and human hands, picking fruit, hanging towels, pouring water and pulling drawers. About 40% of the attempts in the test set are failures, so a model that assumes every grasp works gets caught. The first results show the gap. Every model tested followed the robot arm better than the object it was handling. IRASim scored 78.8 for tracking the arm and 29.1 for the object. Every model also scored lower over full two-minute episodes than over the first 30 seconds. So a model can nail the arm and still lose the fruit. It's early. The platform covers four tasks and three embodiments so far, and real policy rollouts only on AGIBOT's G2. The paper says so itself. For teams building these models, the value is a score against real executions instead of a demo reel. It shows where a model breaks: the object, the outcome or the long run. Submissions opened October 16 for open-weight and closed models. Teams register, get approved, pick a track and upload predicted videos, with up to six evaluations a week. Anyone building a model that predicts robot behavior can bring it to WorldOracle and see how it holds up against the real thing.8h ago4.3K136589034.83%
Labs are drowning in synthetic text, code, and math, but they're starving for the one thing that actually matters: messy, real-world failure logs. Imagine an agent tasked with cleaning up Q3 invoices. It hits the wrong API using valid credentials, quietly scrambles 400 rows of data, logs "task complete," and leaves nobody the wiser for two weeks. That kind of screw-up happens all the time behind closed doors. A single trace of how an agent breaks down and how it gets caught is worth more than a million pristine training examples. It's the thin line between a model that can execute a workflow and one that actually knows when it's about to wreck something. But nobody will ever publish that data. It's too messy, too embarrassing, and a legal minefield, so it stays locked inside enterprise servers. While everyone fights over expensive web scraping and clean datasets, the most valuable data in AI can't be bought. You have to earn it in the trenches, deployment by deployment. Whoever figures out how to capture and learn from failure data owns the evaluation layer. And whoever owns the eval layer controls what every single lab optimizes for next. That's the real moat.frozen7d ago4.0K1005020.43%
Trump just announced "massive" oil, gas and mineral exploration in Greenland and everyone's talking about the oil but if you're into AI and robotics, the real story is the metals here's the thing nobody in tech wants to admit: the entire AI/robotics stack sits on top of a supply chain China controls every humanoid robot needs rare earth magnets in its motors and actuators. neodymium, praseodymium, dysprosium, terbium. without them you don't get precise, powerful, compact joints. you don't have a robot, you have a very expensive statue EV motors, wind turbines, hard drives, drones, defense systems. all of it runs on the same magnets. and in 2024 China made 94% of the world's rare earth magnets that's the biggest chokepoint in tech. we're building the future on magnets one country can switch off last month US, Denmark and Greenland signed a deal that bars China and Russia from setting up bases on the island and screens them out of sensitive sectors like minerals and Greenland may hold 36 million tons of rare earths. only 1.5 million tons are proven so far, but if more of it holds up, it could rank second only to China one example: Greenland Mines' Sarfartoq project. a neodymium-praseodymium deposit that got its first upgraded resource estimate under SEC rules in August, about 40,700 tons of NdPr oxide. exactly the stuff robot motors and EV drivetrains need now think about the math. demand for magnet rare earths has already doubled since 2015, and the IEA expects it to grow another 30%+ by 2030. that's before humanoid robots scale. if they hit even a fraction of what the optimists predict, demand goes parabolic. and right now there's basically one supplier that matters the Greenland story, i think is about whether the AI and robotics revolution gets built on a diversified supply chain or keeps depending on Beijing ,let me know your thoughts belowfrozen9d ago4.0K1302020.43%
Meta is back in the frontier AI race. 14 months after starting a ground-up rebuild of its AI stack, Meta just launched Muse Wonder and Muse Wonder Full. Its benchmark comparisons put the new models alongside GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5. Meta claims state-of-the-art results across key agent, memory, and coding benchmarks. Quite a turnaround in just 14 months.frozen10d ago3.9K2335120.87%
A $5.5M seed on the idea that the next jump in AI coding comes from developers own sessions, not the labs. Beacon turns past sessions into reusable skills that move across Claude Code, Codex, Cursor and 20+ other harnesses. Open source.2d ago3.8K1534030.66%
Meta and Google both shipped new frontier models this week. Did either one replace your main model?frozen9d ago3.8K1526000.61%
never doubt Elon Musk….. https://t.co/P58SxkKvGN6d ago3.5K1020020.40%
Google should be worried about Chrome A lot of my work doesn't start in Chrome or Safari anymore. I just type what I need into Claude, and it does the clicking in the browser inside the app. Yesterday I used it to set up Instagram keyword DMs. With 300k+ followers across 5 platforms, managing my social media is pretty much a full-time job on its own. On Instagram, I share a lot of resources, and when someone comments a keyword on a post, they get the link in their DMs. But every post needs its own keyword, its own reply and its own link, and setting that up again and again was the part I really wanted off my plate. So I asked Claude to do it in my social media scheduler tool. It went through all 17 automations one by one, filled in every form and turned them all on. I don't think browsers are going anywhere, but I'm clicking around in them a lot less than I used to.frozen6d ago3.5K27618051.61%
the UK just had the biggest AI day nobody's talking about and it says a lot about where the power in AI is concentrating. the headline move is NVIDIA handing over $250M in compute across five years to make AI agents safer ... Oxford, UCL and Imperial all in the mix. that's NVIDIA betting that safety research, not just raw capability, is where trust gets won. and it's all timed ahead of their Berlin keynote on Oct 21, which suddenly feels like it matters more. but what makes today interesting isn't the money, it's how fast the stuff is landing. Danaher is about to switch on an AI-run lab that could find new medicines 8x faster. hospital robots trained in NVIDIA simulations are starting to haul lab samples so staff don't have to. Liverpool is building a "City Brain" .. the first one, with dozens of UK cities meant to follow. this isn't roadmap talk anymore. the NHS is quietly becoming the most interesting AI story here. the UK's first AI triage model for GP surgeries, AI helping doctors pull up trusted evidence faster, voice systems that understand patients in multiple languages. if it works at NHS scale, that's the proof point everyone else copies. and the banks aren't far behind - Lloyds feeding billions of transactions into AI models, Starling building its own banking brain on UK soil, AI agents handling payments and customer calls. the industry that moves slowest is moving fastest. NVIDIA basically owns the whole AI stack right now... and the UK is making sure a big chunk of it gets built therefrozen9d ago3.4K2048151.11%
SenseNova-RoboRSI shows why strong benchmark scores are not enough for embodied AI. Real robots face collisions, overheating, time-sensitive tasks, and mistakes that cannot be undone by simply adding more steps. The open study makes these practical limitations much easier to see. Report: https://t.co/qS7fV5rCCt15h ago3.3K1797710058.19%
OMG, Fable 5.5 is almost here. Its model ID showed up in Microsoft Azure's model API yesterday. It's not in the public catalog yet, and Anthropic hasn't said a word. Fable 5.1 left the same kind of trail on Amazon Bedrock and launched about a day later. Opus, Sonnet and Haiku have all moved to 5.5 since September 22. Fable is the last Claude still on 5.1. When 5.1 came out, it was the smartest model Anthropic had ever shipped, and people loved it right up until they hit their limits. Max subscribers were posting that a five-hour window was gone in about 20 minutes. It thought for a long time and wrote a lot. I really hope Anthropic fixes that this time. The first thing I'll check on 5.5 is how long a Max plan actually lasts. The Azure listing shows the same $50 per million output tokens as 5.1, but it looks copied over, so I wouldn't read much into it. There's a hopeful sign in the other 5.5 models, though. Opus got cheaper. Sonnet kept its price and started getting the same work done with fewer tokens. My bet: Anthropic keeps the $50 price and makes Fable 5.5 burn fewer tokens, the way it did with Sonnet 5.5.frozen8d ago3.2K2645211.17%
Anthropic will pay you $3,850 a week to do AI research with their team. It is the Fellows Program for four months, full time, mentored by Anthropic researchers. Around $66,000 in pay, plus $15,000 a month for compute & the goal is a real published paper. But you still need Python, technical interviews, and work authorization in the US, UK or Canada. The January batch deadline is October 18.frozen7d ago3.1K9310120.80%
ninety minutes a day, per user, on Status. figured that kind of usage meant inference costs you just accept. Inworld got it down ~95% and I'm still thinking about that numberfrozen8d ago3.1K150464016.45%
Finally... a launch breakdown that covers the half nobody talks about. Everyone posts their views. This team posted the 3,000 people who actually downloaded, and how they got them.frozen3d ago3.0K1826030.97%
They named an observability tool after Char’s mobile suit from Gundam and gave the launch video Evangelion energy. Somebody on this team really loves mecha. The product is less dramatic: send it logs, it finds what broke, checks whether it’s real, and sends one alert with the cause. No monitors to write, and it’s self-serve from today.frozen9d ago2.8K1021050.64%
Big launch for anyone running agents in production. Falconer puts your internal knowledge base and your public docs in one place, kept current automatically and readable by agents. Backed by GP, BoxGroup and SV Angel, plus angels from Anthropic, Stripe and Cursor. The model isn’t the bottleneck anymore. Context is.1d ago2.6K1613040.92%
The US spent 23x more private money on AI than China last year. And China still managed to nearly close the gap. Now Trump has created a Super Intelligence Force, led by the Director of National Intelligence & I think these three facts tell us more about the future of AI than another model launch. According to Stanford's latest AI Index: In 2025, American private AI investment reached $285.9 billion. China's was just $12.4 billion. Yet by March 2026, the leading US model was only 2.7% ahead of China's best on the performance measure Stanford tracked and that is an extraordinary result considering the difference in private investment. of course, these numbers don't capture everything but China invests heavily through government-backed funds that are not included in the private investment comparison. But the broader picture is totally fascinating... America produces more leading AI models and higher-impact AI patents whereas China leads in AI research publications, citations, overall patent output, and industrial robot installations. so these two countries with very different approaches are competing at the frontier of the same technology but now AI is moving beyond the technology industry. On October 4, Trump announced the Super Intelligence Force, a federal task force led by intelligence chief Jay Clayton. which simply shows how AI is becoming a question of national security, economic independence, and geopolitical influence. Meanwhile, Europe and India are trying to build their own capabilities. Europe launched a process to establish up to seven AI gigafactories, aiming to mobilize more than €30 billion & India announced plans to expand its national AI computing capacity from more than 38,000 GPUs with another 20,000, alongside expectations of more than $200 billion in AI-related investment. Both are significant moves, but announced investments and infrastructure plans still have to become operational capacity. I build products using these models, and the more I follow this race, the more I think we're asking the wrong question. Everyone wants to know which company will build the smartest AI. But what happens when an entire country depends on another country's AI? What happens when its businesses, hospitals, universities, defense systems, and government services all run on models and infrastructure controlled somewhere else? That's a very different kind of technological dependence. And it explains why governments are suddenly taking AI so seriously. The next decade might not be defined by which country uses AI the most. It might be defined by which countries can afford to build, operate, and control their own.14d ago2.3K612010.43%
Claude is coming for OpenAI's dots. Microsoft says only 1 in 5 working-age adults on Earth used AI in June, even with ChatGPT at 1.2 billion weekly users. The problem is the chat box. It waits to be asked, and it only helps people who already know what to ask. Every lab has just started building past it: xAI's Grok Bot in August, Meta's Muse and OpenAI's dots in September, Claude Duet yesterday. So far the people who got the most out of AI were engineers. The ones who knew how to set up skills, plugins and projects. Coding is still about a third of all Claude chats and nearly half of Anthropic's API traffic. Anthropic's own report says it plainly: "Early adopters were highly technical." Duet is Anthropic's version for everyone else. Claude messages first now. It runs on its own computer, watches the inbox and calendar, and only pings when something needs a decision. The early testers aren't engineers: a realtor whose agent caught a house showing booked on the wrong day, a kebab chain owner with 110 employees who said "I don't need to open up 100 tabs." Anthropic's own team already works like this. The head of Claude Code talks through his to-do list with it on walks. Another team lead hands it her list before spending time with her kids and reads the results in an evening report. That's the future of work. The day starts with a brief, not an inbox. The agent chases the unpaid invoices, preps the payroll, drafts the job post, builds the forecast, then waits for one approval on each. The job shifts from doing the tasks to deciding which ones go out. And this is version one. OpenAI says it envisions teams of dots working together. Menlo Ventures predicts agents will drive $100 billion in online sales within 12 months. Whichever lab makes this simple enough for the other 4 billion wins the next billion users.15h ago1.9K10112061.49%
https://t.co/ckNa172Swr6d ago1.6K300000.19%
Nobody learned to ride a bike by watching YouTube. A lot of robots are being trained that way. Maxinsights built a glove that records the part video misses: how hard a hand grips, and the moment something starts to slip. That's the right data to go after. What I'd want to know is how much of it carries over to a robot hand that feels the world very differently.4h ago1.3K1467132.38%