← Back to leaderboard
slash1sol

slash1s

Verifiedactive
@slash1sol

smoke cigs & post about AI tools and coding

RosterAI Agents & Productivity ToolsTrading & MarketsCrypto & Web3
14,368 followers · 236 following · Last polled 3h ago
Perf. Score
63.5
Rank #118
Value Score
75.6
#64 of 273 priced
Median views
4.2K
p25 2.2K · trailing 7d
Followers
14K
+2.5% 7d
ER (impr.)
3.35%
eng ÷ impressions
ER (foll.)
2.29%
eng ÷ followers
Est. CPM
$11.95
QT-rate basis
WoW views
+95.9%
▲ rising

Rates & value · est. from 4.2K median organic views

underpriced
Quote tweetbasis
$50
$11.95 / 1K views
Post
—
no CPM
Thread
—
no CPM
Retweet
—
no CPM
Value Score 75.6 · rank #64 of 273 priced$308.64 / 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 (4.2K), not the mean (9.8K) — so a single viral post doesn't inflate a creator's reach. The p25 floor (2.2K) is the level 75% of posts clear.

Recent posts · ranked by views

PostPostedViewsLikesRepostsRepliesQuotesBookmarksER
This is f*cking insane. Chinese researchers just dropped a paper with a title that will make every agent framework nervous: "The End of Harness Engineering". The argument fits in one line: writing the program around the model is finished as a skill. A harness written by a human is one person's guess about one task on one afternoon. It ships, nobody re-reads it, nobody scores it, and the agent runs that guess for months. A harness written by the loop is a file that grades itself. A small 4B model reads what failed, writes 8 edits, and only the edit that scores best on held-out questions survives. The big model never changes. The file around it does, every round. So the human job collapses to one artifact: the scoring set. You write the questions, never the program. Frameworks turn disposable. The exam turns into the job. That is the exact loop I built and documented below: one https://t.co/JlQQhqH44N, 300 Kimi K3 workers, 37 rewrites in 41 rounds, zero opened by me. The whole paper in one line: write the exam before the code ↓frozen5d ago108K1.2K15875122.1K3.26%
This is f*cking gold.. Chinese researchers just released a paper with a brutal title: "The End of Prediction". The case is simple: an AI that outputs BUY or SELL is finished as a product. Any signal a model can read is priced in before it finishes reading, and active managers already lose to the index over any multi-year window. The question a model answers well is different: what changed, and what is the evidence that it is real rather than noise. So the architecture flips. Plain code finds the anomaly, a million-token model compares eight quarters in one pass, and the heavy reasoning model is spent only on arguing against the result. 10,000 companies a day become 40 deep reads and 8 reasoning passes -- about $450 of inference where a naive full pass costs $60,000. Flagged events hit 37% precision against a 15 to 20% base rate. A real edge, and still nowhere near a forecast. The output has no score and no call. Only what changed, sourced, and what contradicts it. That is exactly the pipeline I built and wrote up below: Robinhood data, Kimi K3 for the reading, GPT-6 Astra for the counter-thesis. The entire paper in one line: you are selling research velocity, not prediction ↓
frozen
29d ago
99K
707
79
27
1
1.3K
2.17%
Garry Tan, the CEO of Y Combinator, open-sourced 55 ready-made agents for Claude Code Each agent is a slash command in plain Markdown with its own role and rules. 134k stars, MIT. What is inside: > 7 for browser work and scraping > 6 for product planning > 6 for design > 6 for memory and retros > 5 for code review and debugging > 5 for security and guardrails > 5 for QA and performance > 5 for shipping and deploy > 5 for iOS apps > 3 for docs They hand work to each other. /office-hours writes a design doc, the plan reviewers read it, /qa and /ship check against it. Clone it into ~/.claude/skills/gstack, run ./setup, then try /review on any branch.frozen23d ago57K395515017412.19%
OpenAI DOTS can take one idea from market research to first outreach and results while you SLEEP > Most people will open it, type one prompt and close the tab. That is the most EXPENSIVE way to use it. Here is the better way, you hand one dot an opportunity and it becomes the orchestrator. It splits the job between five specialists. One validates demand, one sets the offer and pricing, one builds the content, one handles distribution and outreach, and one tracks what worked. Each of them has its own cloud computer and its own browser. They log into your tools once and stay signed in, so your laptop can stay CLOSED. Reading, researching and drafting happen on their own. Anything that sends a message, spends money or overwrites a file stops and waits for your OK. You get a ping in Slack, tap approve, and the work picks back up. You can also set a clock, a morning routine pulls from your apps and drops one brief in Slack before you open your eyes. One objective in, a whole team working on it overnight. Full setup in the article ↓frozen21d ago31K260302813162.03%
This is f*cking insane. A Chinese lab just published a paper that ends a forty-year profession in one title: "The End of Reverse Engineering". The claim fits in one line: a stripped binary is not hidden from a model, it is simply unread. An analyst reads one function at a time. The program has forty thousand of them. That is why the job took weeks and the report was stale before anyone opened it. A model reads the whole artifact in one pass: binary, minified bundle, running app, packet capture. Six fixed steps, inventory to verify, and every claim carries the line it came from. Then the only test that matters: rebuild it from the spec and diff against the original. Empty diff or back to step three. Weeks of guessing cannot compete with one pass of reading. So you stop being the person in the disassembler. You become the person who points: you choose what gets read and you check the diff. Analysts turn disposable. The artifact turns into the spec. That is the exact loop I run with REA + Kimi K3 and documented below: 26 binaries, 26 specs that rebuilt, 23 of them better than the official docs, zero README opened. The whole paper in one line: read the binary, not the README ↓frozen4d ago28K288382533772.60%
This is f*cking insane.. Chinese researchers just released a paper with a brutal title: "The End of Copy Trading". The case is simple: the wallet you have been copying might be a bot, and nothing in a leaderboard will tell you. A leaderboard shows returns. It does not show whether the wallet is a person, a script, or one of ten wallets that entered the same coin in the same second. So they read Solana and Robinhood Chain straight from public nodes and typed every big buyer at block time: sniper, flipper, accumulator, whale or bot. Three or more wallets into one coin within seconds is a pack, scored for coordination. In one live session a pack entered DUES in one second at 72% coordinated -- near-identical buy sizes, wallets never seen together before. Five on-chain checks -- concentration, wash trades, pack buying, early dumping, copycat name -- collapse into one scam score from 0 to 100. No wallet connect, no keys, nothing stored. The only thing that touches a model is the judgment call, and only when you click. That is exactly the tracker I built and shipped for free below: it is called WolfBrain, and the first thing to do is paste the wallet you have been copying into Wolf DNA. The entire paper in one line: look before you ape ↓frozen29d ago27K243192313472.31%
This is f*cking gold... A Chinese lab just published a paper called "The End of Retail Trading" and the whole thing is about a room getting a door. The case is simple. Retail is not the trader with less information. Retail is the trader who gets it five minutes after everyone who paid attention. For three days a listener read every trending token on Solana, found the wallet that bought first and counted the crowd behind it. 12 rooms. 102 echoes. One address with 444 SOL walking in 34 seconds late. All of it on a public board, free. Tomorrow the board gets a door. Holders get every new room the second it is written, before the board updates. /watch on any voice with the delay counter already running. Jev judging every room, organic or bundle, and holders see the verdict first. Cluster copy, the voice instead of the echo. Everyone else keeps the public board, five minutes behind, same as today. The recording never stops. The voices never change. Tomorrow the only thing that changes is who hears them first. The entire paper in one line: latency is the last edge left, and it goes on sale tomorrow ↓frozen26d ago26K292231603332.52%
This is free f*cking gold.. Chinese researchers published a paper with a brutal title: "The End of the Endless Agent". The case is simple: your agent knows how to start and has no idea when it is finished. The fix is three conditions. The file exists where you asked, the behavior passes the check, and the proof sits in the final reply. After N failed attempts it stops and names the blocker, and it never quietly lowers the standard. The entire paper in one line: done is a claim, proof is the exit ↓frozen22d ago24K216353132492.27%
THIS AI CLIP REPLACES A 40-PERSON VFX CREW, A CREATURE RIG AND THREE DAYS OF POST WITH A 30-SECOND PROMPT The sequence sells it. Practical-looking water, blue-screen props, miniatures and crew-style framing come first. The impossible part comes after. Three "production" worlds in under 40 seconds, and they all read as one fake movie shoot. That is the format: fake BTS around scenes that would normally need stunt teams, creature rigs and miniature sets. With @Picsart the same concept goes through image, motion and edit and still looks like one production. One idea becomes 4-6 impossible set pieces, one fake behind-the-scenes reel and cuts for X, Reels and TikTok. That beats generating one pretty shot every time ↓frozen25d ago23K807150360.60%
AI VIDEO JUST FAKED A $100K BLOCKBUSTER SET, A WATER TANK AND A CROCODILE REVEAL IN 12 SECONDS The first 6-7 seconds look like ordinary behind-the-scenes footage. One green-screen horse, one water tank, one crew waiting for the shot. Then the scale flips. A giant crocodile climbs out of the pool and the "set" turns into a full VFX sequence. The trick is fake production logic. One practical prop, one controlled location, one impossible payoff. The viewer understands the setup before the monster ever shows up. With @Picsart a concept like this runs through setup, motion and final reveal without stitching three unrelated shots together. 12 seconds, 1 set, 1 prop, 1 monster reveal. That is enough to make a fake BTS clip feel bigger than a short film budget.frozen24d ago22K851190420.66%
A GUY HANDED HIS APP'S MARKETING TO AN AI IN WHATSAPP FOR 24 HOURS AND WOKE UP TO 300K+ VIEWS No giant prompt. Pedro onboarded Muse like a new marketing hire -> context on his quit-drinking app -> what already works for it -> two rules. Rule one: do not force the app into every post. Rule two: the final creative call stays with him. • What Muse did on its own > Researched short-form that is already performing and broke down why: hook, format, curiosity, why people keep swiping. > Turned the strongest reference into a slide-by-slide concept for the app and spun several variations of the same idea. > Took two corrections from Pedro, hook styling and the product showing up too early, and shipped v2. • Where the human stayed > Pedro only reviewed what felt off and approved the calendar. > Muse organized everything into Postiz, max two posts a day, so the schedule was one review, not a daily chore. > The slideshows crossed 300K views in the first 24 hours. AI takes the research, the creation and the scheduling, the person decides what is worth posting -> the whole 24h breakdown is in the article, link in the first reply ↓2d ago20K722231640.79%
Google researchers just fixed the dirtiest secret of self-improving AI agents. When an agent rewrites its own harness, it cheats. It keeps editing until the benchmark score climbs, and the gains disappear the moment it sees a new task. The harness is everything around the model: prompts, tools, memory and control flow. RRSI puts rules on the loop that rewrites it: > Fewer edits are allowed every round. > A critic throws out edits that hardcode task answers. > An edit survives only if it beats run-to-run noise. > Extra cost has to pay for itself. > Parts that stopped helping get pruned. Up to +14.1 points on the tasks it tunes on, up to +4.7 on five benchmarks it never saw, 30% fewer tokens, and it carries over to other model families. The code is open. Write one adapter file that runs your agent on your tasks, and it evolves the harness in git worktrees round by round.frozen21d ago19K210223201902.34%
Andrej Karpathy's new agent runs AI experiments all night and never asks permission. -> Three files, one GPU, one number to beat. It's called autoresearch. You give Claude or Codex a small LLM training setup and go to sleep. The loop: > Try one idea in https://t.co/Go4zE6thbW. > Commit. > Train for exactly 5 minutes. > Lower val_bpb, keep it. Higher, git reset. > Log the result, repeat forever. The agent is told to never stop and never ask "should I keep going?", because the human might be asleep. You never touch the Python. You only edit program.md, the instructions. Karpathy calls it the code of your research org. Simplicity is a rule, not a vibe: a tiny gain plus 20 hacky lines gets dropped, same score with less code gets kept. This works way beyond LLMs. One number to beat, one file to edit, fixed time per run, and anything can grind overnight. 92k stars, MIT. One NVIDIA GPU, local or rented, with community forks for Mac, Windows and AMD. The loop in motion is in the video below, repo in the first reply.frozen17d ago19K260482902913.25%
Moved about half my stock trading off my brokerage this month onto @baibai_cx lol. 24/7, cashback, and I'm farming points while I do it Never thought my wallet would replace Robinhood https://t.co/N6qSCAfnV6frozen8d ago18K1002212380.91%
This is f*cking gold.. Chinese researchers just published a paper that kills the hottest skill of the year: "The End of Context Engineering". The argument fits in one line: a context window packed by a human is a guess made before the model has looked. You pack what you would have needed. The model needs something else. Too much of the obvious, none of the surprising. So the paper hands the model an index instead: one line per thing that exists, never the content. The model names what it needs, pulls only that, runs the task, and every item gets credited or blamed for the result. What did not move the score gets dropped. What did gets pinned for the next task. Across 300 tasks: score 0.14 higher than the hand-built pipeline with a window 81% smaller. Three hundred prompts cannot be packed by hand. They can be scored by a loop. So you stop being the person who decides what the model sees. You become the person who names what exists and reads the score. Pipelines turn disposable. The index turns into the job. That is the exact system I run and documented below: one index file, 300 Kimi K3 tasks, 212 items pinned, 4,970 dropped, zero prompts packed by me. The whole paper in one line: index the shelf, not the page ↓frozen4d ago14K196172602293.25%
This is f*cking insane. US researchers published "The End of the Judgment Call" and the whole paper is one split: Claude builds, Jev decides. The case is simple. An overnight Claude loop asks itself around 600 yes-or-no questions. Safe to run. Next step. Goal met. Each one costs about 4 cents on Fable 5.1 and takes seconds. Jev came out of stealth on September 15 from a co-author of InstructGPT. It answers typed questions only, with a calibrated probability, in about 100 milliseconds, at $0.042 per million input tokens. When it says 0.9 it is right nine times in ten. That is a number your code can act on at 3am with nobody watching. Three hooks in Claude Code and the judgment calls are gone: a 100 ms gate on every bash command, a stop hook that knows when the work is really done, a cron that only wakes Claude for reproducible bugs. The bill for 30 nights goes from $765 to $3.02. What Claude spends on building does not change. The article below has all three builds with code. The entire paper in one line: the cheapest Claude turn is the one that never needed to start ↓frozen26d ago14K156112102022.75%
This is f*cking gold.. A Chinese lab just dropped a paper called "The End of Alpha" and it is about Solana copy trading. The case is simple. One wallet buys. 27 wallets follow within 34 seconds. Every one of them thinks it was early. The paper calls that a room, not alpha. Their listener sorts the last 100 swaps of every trending token by time. First buyer is the voice. Every wallet inside 60 seconds is the echo. Three echoes plus one SOL and the room is real. Day one it logged 12 rooms, 102 echoes and 444 SOL following a single voice. The brutal part is the median echo enters after the move is already priced. Following the copy is late by construction. The post below is the same listener running live, and the record is never edited. The entire paper in one line: follow the voice, never the echo ↓frozen26d ago13K154121701762.68%
This is f*cking insane. Chinese researchers just dropped "The End of Loop Engineering" four months after the term was invented. Fastest death of a discipline I've seen. The claim in one sentence: the run is the only thing that knows when it is finished. Every loop you have ever written had a cap, a budget and an exit condition decided before the first action. The paper calls all three the same mistake. What replaces them is almost nothing: a verifier the agent cannot fool (tests, compiler, diff) and a memory of its last three states. > Verifier passes -> done, no extra iteration. > Three identical states -> hand it to a human with the full trace. > Anything else -> act again. Across 2,400 runs that ended at the wrong moment 4% as often as a fixed cap, with 2,391 self-stops and 9 escalations. So the loop engineer leaves and one script stays: the check that says the work is done. You write that. The agent writes the ending. Loop engineering turns disposable. The check turns into the job. That is exactly what I run and documented below: one verifier, 300 Kimi K3 workers, zero caps written by me. One line to remember: the exit is observed, never authored ↓frozen3d ago13K202102912423.65%
This is f*cking gold.. Chinese researchers just released a paper with a brutal title: "The End of the Flat Rate". The case is simple: paying a fixed subscription for AI labor is finished as a strategy. A flat rate is a ceiling you rent, and five workers sharing one ceiling are five workers waiting in line. A model billed per token is a brain you own -- every worker draws on it as needed, one price, no throttling, no seat limits. At team scale nothing else scales. Five workers cannot share a subscription, but they can share a brain. So the person stops buying seats. You turn into the operator -- you hire the desks, install the brain, and clock out. Subscriptions become the trap, the per-token brain becomes the setup. That is exactly the office I built and wrote up below: five Grok Bot workers, one shared computer, one Kimi K3 brain at $3 per million tokens. The entire paper in one line: hire the desks, install the brain ↓frozen29d ago12K1386250892.11%
THIS OPEN-SOURCE AGENT JUST MADE HERMES LOOK SLOW AND EXPENSIVE. SAME TASK, 4X FASTER, A THIRD OF A CENT One prompt in both: research the 10 most popular AI coding agents, compare pricing, features, open-source status and GitHub popularity in a table, include sources. • Speed > OpenHuman (@tinyhumansai) finished the table in about 25 seconds, 8 tool calls, 10 sources. > Hermes needed over a minute and a half for the same job. • Cost > Total bill: $0.0036, ticking live on screen while it ran. > Context window live too: 37.7K of 128K, broken down by what used it. > Hermes (@NousResearch) hands you the answer and hides the meter. • Why it's not a toy > Open source, 39K+ GitHub stars. > Memory lives on your own machine, not in someone's cloud. > Your own API keys or local Ollama models, your call. Speed you can time and cost you can see -> every other agent is asking you to trust it on both. OpenHuman by @tinyhumansai, the Hermes alternative I'm switching to. Download link ↓frozen17d ago12K1097409631.93%
Peter Steinberger, the creator of OpenClaw, open-sourced the exact setup that keeps Claude Code and Codex on the same rules One AGENTS.MD, 69 skills and a sync script that links them into both agents on every machine he uses. 6.6k stars, MIT. What is inside: > 7 rule sections, from tone to git safety > 69 skills, each with a description the agent reads before loading it > 4 helper scripts with zero dependencies > 1 pre-commit hook that checks every SKILL.md > 5 slash commands: /handoff /pickup /fixissue /landpr /raise > 1 tools.md listing every CLI on his machines They hand work to each other. codex-first has Claude plan and review while Codex writes the code, autoreview loops on the diff, and handoff packages state for the next agent. Fork it, run scripts/sync-skills, and replace his hosts and accounts in AGENTS.MD with yours.frozen23d ago11K208243111673.85%
This is f*cking dangerous.. A $1.7M/yr app studio just opened its playbook: 0 to 6,000 videos a month, and the last step is a Grok Bot routine that runs the whole thing every morning. Step one was hiring 60 creators and firing 5 to 10 of them every week. No calls, no portfolios, just a brief and a proven format. 15 stayed -> 50M views in 3 months. Step two never touched a human. > Winning videos cloned as AI replicas. Gemini reads the body language, GPT Image 2 draws the character, Seedance 2.5 animates it, Claude edits. 10x cheaper. > A Zack D Films clip as bait, then a snap into the already-viral video. 10M+ views from content that had already peaked. > AI slop stories about the same problem the app solves. It gets laughed at and it converts. Step three is one routine. Every morning Grok Bot grabs the new videos and schedules them to TikTok, Instagram and YouTube through Postiz. The founder opens the calendar and that is the whole job. Humans find the format once. The bots run it forever ↓frozen10d ago11K17091821863.53%
This is f*cking gold.. Chinese researchers just released a paper with a brutal title: "The End of the Studio". The case is simple: an animated series no longer needs a team. It needs three tools with one job each, and the producer is a prompt. GPT-6 Astra researches what already works. Claude Opus 5.5 writes the world and one episode a day. @Picsart produces all of it -- hero, world, music, and an agent director that shoots every episode to a locked look. The part most people miss: the same 10 million views are worth $300 to $1,000 on Shorts and $4,000 to $10,000 on TikTok, as long as the episode runs past a minute. So the entire architecture reduces to one question: which views actually get paid. One hero, one world, a new door every day. The elevator never changes, only what is behind the door does, and that is exactly what the platforms now pay for. Their modelled month: 30 episodes and two YouTube seasons, $24,689, no camera, no animators, no studio. That is exactly the pipeline I built and wrote up below, with every prompt ready to paste: Astra finds, Claude writes, Picsart produces, Astra measures. The entire paper in one line: the studio ends where the daily episode begins ↓frozen27d ago11K9161201032.00%
A GUY BREAKS DOWN CLUB SANDWICHES LAYER BY LAYER AND THE VIDEOS PAID HIM $13,286 LAST MONTH > 91,400 likes on 2,600 comments. Cream terminal, a club sandwich on a toothpick flies apart into 9 layers, toast, bacon, turkey and Swiss drop back one by one, and the counter stops at 677 kcal. The sandwich is the bait and the format under it is what gets paid: -> A club sandwich on a toothpick reads from across the room, so the eye stops before the top toast even lifts. -> Bacon slams in at +215 kcal, the biggest jump on the ladder, and that spike is the second people replay. -> The first frame and the first line of text get logged for every breakout, and this one opens on the whole sandwich standing still right before it blows apart. -> Same terminal, same counter, new food: the setup gets saved as an asset once and a one-off clip turns into a series that never runs dry. -> The comments fight over whether bacon deserves 215 kcal while the sheet scores the clip against the channel median and anything past 35 gets a sequel that week. People will watch the same build twice if the object changes, the shape stays and the food gets swapped. It works on anything you can stack, a pizza slice, a hot dog, a burrito, a tower of pancakes. Claude spins 36 concepts off one format, CapCut cuts them, make posts them and the bill stays at $49 a month. The bacon got the replays, and the format from the sheet is what got it a sequel. Every prompt behind the factory is below ↓frozen20d ago10K13972701022.71%
HE DROPPED 10,000 EGGS INTO A VOLCANO AND THE LAST THREE SECONDS RUINED THE ENTIRE SHOT A helicopter carries a cargo net over a crater, and the premise reads before anything moves, so the opening works. 10,000 eggs weigh at least 1,250 lb (567 kg), making the scale plausible enough to hold attention. Then the physics collapses: > The eggs become sunny-side-up before a clear impact. > Cooked eggs appear inside the helicopter, breaking spatial continuity. > Lava can reach 1,170°C, while egg protein sets near 62-70°C. at that heat, water flashes into steam and the eggs char instead of landing intact. > The final reaction adds noise after the payoff. The useful lesson is that spectacle needs one clean transformation with visible cause, contact and consequence. If the result appears before the impact earns it, viewers stop watching the event and start spotting the render. Try it with @Picsart ↓frozen22d ago9.6K13210240572.33%
OpenAI's new Dot is a ChatGPT agent that works your inbox, Slack and GitHub while you sleep, for $3.33 a day. -> The crawler in this video is that agent on the job. Sam Altman pitched it at DevDay as a chief of staff. The crawler everyone shared on X this week turned out to be the best way to show one: it holds a job for weeks from its own cloud computer, laptop closed. $100 a month on Pro -> Free, Go and Plus do not get one. What it did in OpenAI's own numbers: > Blocked 16,600 prompt injection emails, zero hits. > Went beyond its task in 0% of cases, the previous model did it in 48%. > Slipped 8.6% of the time across 5 chained jobs, 19.7% across 10. So you give it one job it can read and nothing it can send. Replies, payments and merges stay on your side of the approval queue. One forgotten invoice it catches pays for the month. The full cost breakdown, the EU trap, custom rules and the day-one brief are in the article ↓frozen19d ago9.3K15512272923.10%
THE CO-FOUNDER WHO RAN GROWTH AT CAL AI JUST LEAKED THE ENTIRE AI UGC PIPELINE. ONE PERSON, ZERO CREATORS, THOUSANDS OF TEST VIDEOS Cal AI went to 15M downloads and a $40M year before MyFitnessPal bought it. He co-founded it and ran their UGC and influencer programs the hard way, and now says this system would've made him 10x more productive. • The loop > Claude Code plus Apify pulls the outlier videos in your niche, the ones doing 12x an account's normal numbers. > Claude builds a local gallery, you pick what to clone, an analyzer breaks down hook, tension and payoff by timestamp. > Higgsfield generates the AI presenter from a real creator frame, Seedance 2.5 animates it, one test take before any batch. • The part that makes it an army > Finished exports go straight into Postiz over MCP, one draft per creative per platform. > Claude writes the captions, proposes the schedule, you approve, it publishes to TikTok, Reels, Shorts and Facebook. > Each platform reports back, winners get remade with real creators, losers die cheap. The whole thing runs in one Claude Code window -> research, generation and distribution with no human in the loop until approval. He's handing out 13 prompts and the full skill file. Article below ↓15d ago9.1K17192921493.96%
This is straight f*cking gold.. A Chinese lab just put out a paper titled "The End of the Copy Trader" and it describes, almost line for line, what this bot does on Solana. The case is simple. There are maybe 40 actual traders on Solana. Everybody else is an echo. The listener scans every trending token, finds the first wallet in and counts every wallet that repeated it within sixty seconds. First buyer is the voice. Three echoes plus one SOL and you have a real room. The first scan came back with 12 rooms, 102 echoes, and a single voice with 27 wallets and 444 SOL trailing it. The brutal part is the timing. The average copier lands 34 seconds after the voice, and by then the move is already in the price. The paper calls that a room, not a trade. The bot calls it /echo and hands you your label: voice, echo, or quiet. No calls, no targets, just the board with a timestamp on everything. The entire paper in one line: you are not trading, you are 34 seconds behind someone who is ↓frozen25d ago9.0K10421801022.51%
This is an actual goldmine for every AI builder. 400+ ready-made agents and skills across five repos, and almost every job is covered. Anthropic, its engineers and the community already did the work. If you want an entire company in agents: > https://t.co/MRRkGtI0Y6 (158k stars) > About 170 agents sorted into engineering, design, marketing, sales, product, testing, security and support, plus extra packs for finance, healthcare and game dev. If you live inside Claude Code: > https://t.co/k8ABINnkYX (40k stars) > 202 agents, 184 skills and 105 commands packed into 94 plugins. Install only what you need. > https://t.co/0IV10mN8Gn (103k stars) > 25 skills from an Anthropic engineer covering the full dev cycle: interview-me, specs, TDD, browser testing, security review, CI/CD. If you want Anthropic's own: > https://t.co/lWlIBZxbHR (26k stars) > One plugin per role: sales, support, product, marketing, legal, finance, data. > https://t.co/ApbV69F5CP (18k stars) > Agents ready to deploy: a daily brief from Slack and GitHub, scheduled Sentry triage, Slack and Linear bots, a support agent, a financial analyst. Begin with claude-quickstarts. Every agent there is a full working example you can run and pull apart.frozen11d ago8.6K172192122415.31%
This is pure f*cking gold. Jev engineering for 24/7 agents, the full desk build. The expensive model only does the real work. Every decision inside the loop, safe or not, next step, done or not, goes to Jev, a model that cannot write a sentence and only answers typed questions in about 100 milliseconds. So the frontier model builds and never decides. The daily bill drops from $600 to about $25, and nothing about the work itself changes. In the right hands that is a full-time assistant for the price of a lunch, on shift 24/7. The entire build with code and the bill is below ↓frozen23d ago8.1K12210271843.02%
A WORLD MODEL JUST DESIGNED A PHYSICAL OBJECT 100X FASTER AND 150X CHEAPER THAN GPT-6 ASTRA. DOWN TO THE SCREWS @skyfallai, the neolab from the deep learning pioneers behind Maluuba, just released SKY-001, the first engineering world model. Text prompt or 2D sketch -> it dreams several designs in parallel -> picks the best -> breaks it into thousands of parts -> reassembles a CAD build that opens in Autodesk Fusion. • Why LLMs lose here > They produce 3D shapes that look right and cannot be built, no part hierarchy, no mechanical constraints, no feature tree. > Every design step burns tokens, and @spisallyouneed's team measured it: 100x faster and 150x cheaper than GPT-6 Astra on CAD and BIM. • Where this goes > Design is stage one, the goal is an AI general engineer running a company from design to production. > Their own org chart already reads 0 employees: an AI CEO, Sky 001 for design, Sky 001 for ops, plugged into AutoCAD, NetSuite, Salesforce and Zendesk. World models imagine and simulate before they act, LLMs guess one token at a time -> for autonomous manufacturing the first one wins.3d ago8.0K904221612.22%
Sam Altman runs one ChatGPT dot and gave it exactly one job Dots launched at DevDay: always-on agents, each with its own cloud computer, working while you are offline. Altman showed his own on stage. It reviews everything that arrived overnight, sorts it by urgency and drafts replies without sending any. In the morning it pings him with what is urgent and nothing else. Custom Rules decide what it does alone, what needs approval, and what always stays with him. The keynote also listed the heavier jobs a dot can take: > Watch bug reports and test fixes in draft PRs. > Hunt down what is slowing your app. > Migrate an app off a legacy API before the shutdown date. The five-step setup for the same dot is in the article and video below ↓frozen20d ago8.0K13618190943.34%
THIS AI CLIP REPLACES A TRAINED PUPPY, A PROP DEPARTMENT, A UV LIGHTING RIG AND A CLEANUP CREW WITH A 10-SECOND PROMPT The reveal sells it. A living room, a wooden chair, a paint can with a hazard label and a puppy poking one paw into the swirl come first. The impossible part comes after. Three beats in 10 seconds, the sniff, the fall and the glow, and they all read as one shot in one room. That is the format: a fake pet clip around a moment that would normally need an animal trainer, a safe prop, a lighting setup and a dog bath afterwards. With @Picsart the same concept goes from a still of a puppy and a can through motion and the lights-off reveal and still looks like one take. One idea becomes a 10-second clip, a slow-motion cut of the fall, a glow-only loop for the dark and versions for X, Reels and TikTok. That beats generating one cute puppy shot every time ↓frozen24d ago7.7K1067180392.21%
This is free f*cking gold. Five repos that replace an AI influencer agency, run from one Claude Code session. I can't understand why anyone is still paying four figures a month for this: 1. OpenMontage -> https://t.co/ErHQU9dw4E Claude Code as a video studio. 12 pipelines, 100+ tools, 700+ skill files, 60+ providers, avatar and clip factory included. 68k stars. 2. Higgsfield Skills -> https://t.co/42wR6VBxR9 Soul ID locks a face from 5 to 20 photos. Marketing Studio runs 9 ad modes. Virality Predictor scores the video before you post it. 3. Higgsfield CLI -> https://t.co/dJeE4SUIN4 30+ image and video models from the terminal, plugs into Claude Code, Cursor and any MCP agent. 4. Higgsfield Claude Skills -> https://t.co/55jVMVGTbu 19 community skills. Character image -> 9:16 UGC video in one command, 15 prompt styles. 5. Phantom by ByteDance -> https://t.co/QXQbuXO7pS Same character across every shot, up to 4 reference images, 720p, built on Wan2.1. The loop: identity -> character bible -> production -> virality test -> ads. Low score goes back to regenerate on its own, you sign the sponsor deals and approve the spend. The tools were always supposed to be free. The stack map is below ↓frozen5d ago7.6K150182511454.49%
A TOOL CALL RETURNED 200 AND THE ANSWER WAS STILL WRONG, THAT IS THE BUG NO DASHBOARD CATCHES DigitalOcean just launched the stack for agents that are already in production. Before: agent misbehaves -> logs say 200 -> you find out 400 users later. Now: Signals flags the session -> one click makes it an eval -> Simulations replays it against your fix. • What @digitalocean shipped, all in preview > Signals watches production traffic and flags frustrated users, failing tools and timeouts with nothing to instrument. > Simulations replays real multi-turn sessions on your actual harness, mock tools for the risky steps. • The part I did not expect > Agent Data Services turns call recordings, PDFs, video and agent traces into one catalog you query by API, vector search and SQL. > Insights follows one agent request through every tool call and policy check, logs attached. A human still approves every fix, the agent just brings the evidence -> production failures become tests instead of tickets. Agent Data Services is private preview, the rest is public preview, links in the first reply ↓frozen10d ago7.4K1044213572.55%
An Anthropic engineer packed his whole engineering workflow into agent skills, so anyone can copy it. Addy Osmani, ex-Google, now works on Claude Code. His repo turns a coding agent into a senior engineer. Each skill is a structured workflow: when to use it, the steps, the excuses agents use to skip steps with a rebuttal for each, red flags, and the evidence required before it can call the job done. 25 skills, sorted by phase: > Define: figure out what to build. > Plan: break it into small tasks. > Build: write it test-first. > Verify: debug and prove it works. > Review: quality, security, performance. > Ship: rollout, CI, docs. Skills also switch on by themselves. Start designing an API and the API skill kicks in. Useful if you ship with agents solo, or if you want every agent on your team held to the same bar. npx skills add addyosmani/agent-skills The full pipeline is in the article below, repo in the first reply ↓frozen19d ago7.3K131192501023.78%
IF ONE PROVIDER OWNS YOUR MODEL, YOUR HARNESS, YOUR DATA AND THE LEARNING LOOP, EVERYTHING YOUR AGENT LEARNS MAKES THEIR PRODUCT BETTER That is Paddy Srinivasan's argument on stage today at the @digitalocean Open Intelligence Summit, keynote 1pm PT. Closed stack: agent learns from your data -> the gain lands in someone else's product -> switching gets harder every run. Open stack: pick the model per task -> keep the data and what the agent learned on infra you control -> move any layer. • Who is in the room > NVIDIA co-hosting, Hugging Face, OpenRouter, Nous Research, Arcee, Liquid AI and MiniMax on one program. > Build, Serve and Run tracks from 2:15pm PT, then a closing panel on what is next. • The line that lands > "Owning your intelligence is owning your future", Arcee's CTO in this morning's video. > Open started behind in operating systems, databases and container orchestration, and won every time. The open versus closed debate usually stops at model weights, today it moves to the harness, the data and the loop -> where do you land ↓frozen9d ago7.2K854271512.34%
I CANCELLED A $1,000/MO AI BILL AND PUT FIVE GROK BOTS ON ONE COMPUTER INSTEAD This is f*cking insane. Five desks, one shared cloud machine, one brain installed on it a single time. Kimi K3 at $3 per million tokens. They ran my entire Tuesday while I was lifting. > 07:04 Scout went through 412 profiles -> 31 leads that match the ICP. > 07:11 Clerk emptied the inbox -> 7 replies out, 2 drafts parked for me. > 07:19 Builder pushed landing v3 -> tests green, preview link waiting in the chat. > 07:26 Ledger ran the invoices -> $4,120 billed, one refund -$180, net $3,940. > 07:31 Chief: five reports in, nothing needs you, clock out. 2.73M tokens for the day. $8.19. The flat plan I dropped would have billed me $33 for that same day, bots working or bots asleep. The trick is where the brain sits. Grok Bot gives every worker the same computer, so K3 gets installed once and the whole roster runs on it -> five seats, one bill, no per-seat tax. The full setup, every seat and every config, is in the article ↓frozen9d ago7.1K1099170843.10%
THE BOTS FARMING POLYMARKET BTC, ETH AND SOL UP/DOWN SEE ONE THING YOU DON'T -> NOW YOU CAN REPLAY IT Every 5-minute Up/Down market on Polymarket, BTC, ETH or SOL, has a full order book behind it. Retail sees a price. The bots see the depth, and the depth moves before the price does. @Poly_Historical records that depth and lets you replay it. • what's in the archive > Full Level-2 bid and ask depth for both the UP and DOWN tokens, every price level. > Snapshots every 300 ms, served in under 50 ms. > 13,000+ resolved markets: BTC on 5m, 15m, 1h, 4h and 24h, ETH and SOL on 5m, 15m and 1h. > 120 days of history to pull from, enough to backtest across every regime BTC, ETH and SOL have thrown this year. > Binance BTC spot and futures depth lined up next to it on Pro. • Strategy Replay > Pick any resolved market and set your entry and exit rules in dropdowns, no code. > The market replays tick by tick with your strategy running on it. > You get PnL, drawdown, slippage and win rate from real order book fills, not from a candle chart. • Backtest AI Agent > Describe a strategy in plain English and the agent backtests it for you. > Around 100 AI backtests included, with a written analysis of what worked and why. All of it starts at $0, the free tier needs no card, Pro is $17 a month -> cheaper than one bad 5-minute market. Most people trade Up/Down blind. The ones who win replay the tape first. Link in the first reply ↓frozen19d ago6.7K793161462.16%
OpenAI Dots gets risky when one agent keeps running without a real checkpoint. Long chains look efficient, but the longer they run, the easier it is for one bad step to contaminate everything after it. That’s why I’d structure it like this: > One dot, 4 seats. > Lookout -> Maker -> Skeptic -> Runner. > Max 5 steps before a checkpoint. > Permissions split into allow / ask first / block. > Approved work always lands somewhere permanent. The numbers make the case pretty clearly. 5-step chains: 8.6% flagged. 10-step chains: 19.7% flagged. two 5-step legs with checkpoints: 3.4%. So the interesting part isn’t making the agent more autonomous. It’s giving the agent a system that forces it to stop, review what happened, and only continue when the last leg still matches the brief.frozen20d ago6.7K13112270823.79%
I HAVE SEEN A HUNDRED VISION AI DEMOS THAT LABEL EVERY OBJECT IN A PHOTO AND STILL CANNOT TELL YOU WHAT HAPPENED, THIS ONE REASONS @ElorianAI, the lab from @AndrewDai and @dustinvtran, just shipped its first model. The founders led Gemini and Apple's MM-1 -> raised a $55M seed at a $300M valuation from Altimeter, Menlo and Striker -> spent it on one problem, visual reasoning. • What reasoning looks like > It watched the opening of Back to the Future and caught all 16 key events, from the skateboard to the blown speaker. > It counts only the open pliers in a tool pile, ignoring closed ones and scissors that look the part. • Why now > Text reasoning is superhuman, visual reasoning still trails what a person sees at a glance. > Robots, job sites, warehouses and video are where that gap costs real money. Free chat through the waitlist, paid API for builders -> most of the work AI has not touched yet is visual.frozen6d ago6.4K732182362.03%
I ran a slideshow page for 14 days after reading this. No camera, no editor, no creators. My numbers are smaller than Ernesto's and more useful to me: > 412 competitor slideshows read by Scout > 3 formats kept, everything else killed > 19 slideshows built, 6 slides each, 41s per build 1.11M views in 13 days, best one 243K > 0 videos filmed, 0 posts went out without my tap The part nobody tells you: the first week did 61K total. I posted 7 different hooks and none of them moved. Week two I picked one hook, "5 AI tools for people who hate spreadsheets", and posted it 5 times with different tools inside. Part 3 did 243K. Part 5 did 189K. That one hook is 70% of my views. So my rule is now in the farm itself: a hook gets 5 tries before I kill it. Not 1, not 20. The automation is real, but it only matters after the hook works. Four Grok bots building 120 slideshows of a dead hook is 120 dead posts, faster. What I'd change in his setup: I don't let anything auto-post. Maker builds, Gate holds, I tap. Two minutes a day. The one morning I skipped the tap, a slideshow went out with the wrong page name on slide 6. The desk below is my farm. Scout, Brief, Maker, Gate. Same numbers as above.1d ago6.4K642190401.97%
This is f*cking brilliant.. A Chinese lab just published a paper called "The End of the Sleeping Trader" and it lands the day after OpenAI shipped Dots. The case is simple: an agent that never sleeps is useless on a trading desk unless it has something worth watching. On Solana that thing is the room. One voice buys first, the echoes repeat it inside sixty seconds, and it happens at 4 AM whether you are awake or not. So the Dot sits in the room 24/7, reads every voice before the echo lands, and runs /watch on the wallets you pick. It checks the Jev verdict and only wakes you for organic rooms. Bundles get silence. Cluster copy never fires on its own, it waits for your approval every time. The entire paper in one line: you do not need to be awake, you need an agent that knows when waking you is worth it ↓frozen21d ago6.2K964150843.20%
YOUR WHOLE AGENT INFRA JUST BECAME ONE SDK CALL -> PROMPT IN, FINISHED WORK OUT Building an AI assistant used to mean building a computer for it first. Tools, browser, planning, execution loop, retries, state -> months of plumbing before the first useful job. • Pine Computer by @PineAIAssistant A cloud computer built for AI, intelligence to use it included. Structured browser state in, not pixels, so the agent stops guessing from screenshots. • The new build Integrate the SDK, hand it a job, get the result back. Every user gets their own cloud computer, so the next big assistant might come from a three-person team. 78.3% checkpoint score on SaaS-Bench v1.1, highest in the published comparison -> at ~1/20 the model-token cost of GPT-5.6 Sol plus Codex. Private beta, link below ↓frozen13d ago6.0K962284482.94%
This is f*cking insane. A Shanghai lab just published "The End of Agentic Engineering" and it explains why your Polymarket Up/Down bot keeps losing to bots you cannot see. The argument fits in one line: an agent engineered by a human is a guess about a market the human saw as a price chart. The bots never saw the chart. They saw the order book, every level, and the depth moves before the price does. The chart arrives after they have already acted. So the paper stops writing agents. It hands the model a recorded Level-2 book at 300 ms resolution and lets it replay resolved markets tick by tick. Entries are filled against real depth. Scores come from real slippage and drawdown, never from a candle. 13,000 five-minute markets on BTC, ETH and SOL. 41,000 simulated positions. Zero live losses before the first real trade. A person cannot watch depth at 300 ms. A replay can. So you stop being the person who writes the rules. You become the person who records the tape and lets the replay grow the agent. Agentic engineering turns disposable. The tape turns into the job. That is the exact loop I run on @Poly_Historical's archive and documented below: full L2 depth for both tokens, 120 days of history, Strategy Replay with no code, and an AI agent that backtests a strategy you describe in plain English. Free tier, no card. The whole paper in one line: replay the tape before you trade ↓3d ago6.0K12252121364.78%
CODEX FOR CREATIVE WORK JUST DROPPED, ONE AGENT THAT RUNS BLENDER, AFTER EFFECTS, DAVINCI AND UNITY FOR YOU A creative idea used to mean learning five apps. Edit here -> motion there -> 3D somewhere else -> sound in a fourth tab -> move everything by hand between them. The execution tax was the thing killing ideas, not the ideas. • What @startvoyager is > A desktop app that sits on top of your files and your creative tools. > You describe the thing, it orchestrates After Effects, DaVinci Resolve, Blender and Unity to make it. • Under the hood > Built to squeeze the most out of Astra, Opus and open-weights models, so you bring whatever you already pay for. > Built by @ravisparikh, who wrote songs with 50M+ streams and built a 900M+ startup, so both halves of this story are his. Coding got its agent two years ago and the whole industry changed -> video, music, games and graphics just got theirs. Describe it, revise it, ship it, from one window ↓frozen14d ago5.7K1022213553.24%
EVERY QUANTUM COMPUTING COURSE STARTS WITH 40 PAGES OF LINEAR ALGEBRA -> THIS ONE STARTS WITH A SLIDER AND THREE SPHERES @QubitLabsPcl is an interactive quantum lab in the browser -> you run the famous protocols instead of reading about them -> every step shows what changed and why. My first run: teleportation -> tilt the message qubit to 75% zero, 25% one -> entangle Alice and Bob -> measure -> watch the state show up on Bob's sphere. • The learn-by-doing part > Teleportation, superdense coding and entanglement as step-by-step walkthroughs with live Bloch spheres and measurement bars. > A "try this experiment" nudge on every step, so you change a number and see the protocol react. • The build part > A circuit editor to wire your own gates, test them on a noiseless state-vector sim and save them to your account. > Reports you can download when a circuit does something worth keeping. Quantum stays abstract only until you can touch it, and this is the first tool that let me -> https://t.co/iwjQR8k2Wy, start with teleportation and thank me later.2d ago5.5K772181482.64%
ONE STOCK PHOTO OF A GUY IN THE DESERT -> FOUR EDITS LATER IT IS A CITY STREET SHOOT WITH A PIGEON ON HIS CAP @envato just launched Edit with AI. Find a real photo -> riff it with AI -> keep what made it good, change what does not. In the demo: teal Polaroid instead of the camera -> city street instead of the desert -> new outfit -> pigeon on the cap. • Why it beats a blank prompt > You start from a human-made shot, so the face, the light and the composition are real. > Every download from 29M+ assets is commercially licensed. • What you can riff > Remove, recolor, expand, restyle or add anything, then compare and download. > One real shot turns into a whole campaign for OOH, social, web and vertical. The internet is tired of AI slop and the fix is starting with something real -> link in the comments ↓frozen7d ago5.4K911202452.92%
ONE MAGICIAN ON ONE STAGE PULLS A NEW ANIMAL EVERY WEEK AND THE COMMENTS UNDER THE SERIES PAY $12,800 A MONTH > 241,000 likes on 11,300 comments. Same stage, same grey-haired man in a black suit, a steamer trunk behind him. He drops a red cloth over a cat, pulls it away, and a red kangaroo a head taller than him is standing there. It takes his shoe, he hides in the trunk, and in the last second he climbs out and it stares him down. The series is doing more work than the kangaroo, and that is the whole design: -> Same man, same trunk, same cloth as the lion clip, so the viewer knows the bit before the cloth drops and only waits for the animal. -> The cat sits still under the cloth for 2 full seconds, a detail half the audience only catches on the second watch. -> The kangaroo arrives at real scale, so the swap reads as a stage trick and never as a cutaway. -> It behaves like a kangaroo, takes the shoe, turns its back, ignores the room, and never like a prop. -> The comments are "Which theatre", "Next week?" and "How", and "Next week?" is the one a keyword turns into a row every week. A trick everyone knows is a shortcut the brain cannot skip, and swapping the animal every week is the cheapest series in short-form. It works on any bit with one slot to swap, a sawing box, a card, a hat, a cage, and on @Picsart that slot is one line in the prompt. Every clip gets its own word, LION last week, ROO this week, so Monday's ranking shows which animal made people ask, by completed runs per 1,000 views. Blotato sends the DM and Claude runs the count, $49 a month for both. The kangaroo got the shares, and the stage everyone already knew is what got them back next week. One word per clip, article below ↓frozen22d ago5.4K743210392.54%
Someone finally read the source code of 11 coding agents, and every one of them is built the same way. Wavestone AI Lab went through Claude Code, Codex, Gemini CLI, OpenHands, Aider, OpenCode, OpenClaw and four more, about 4 million lines in total. Their definition: an agent is a model plus a harness, and the harness always has the same 7 parts. Only the size changes. > Loop: from a plain while loop to a replayable event log. > LLM layer: one prompt template on one end, 29 provider profiles on the other. > Tools: bash only at the small end, 43 typed tools loaded on demand at the big one. > Memory: whole history kept, or notes the agent carries across sessions. > Safety: a step limit, or policy rules plus a reviewer model plus an OS sandbox. > Orchestration: Aider skips sub-agents on purpose, others fork them with their own context. > Extensions: hooks, skills and MCP. SKILL.md in 9 of 11, MCP in 8. Two things none of them do: import an agent framework, or retrieve code with embeddings. It closes with 18 design recommendations and a 90-line harness that covers all 7 parts. Key pages: > p.4: the 7-part map. > p.12: the 3 loop types. > p.53: why no frameworks and no embeddings. > p.67: the 18 recommendations. > p.71: the 90-line harness you can copy. The anatomy in motion is in the video below, paper in the first reply.frozen18d ago5.1K10610371654.26%
72% OF API DOCS TRAFFIC IS ALREADY AGENTS AND YOUR DOCS STILL ONLY TALK TO HUMANS The team behind @readme just launched Restless, an API that fixes itself from its own traffic. Agent hits your API -> request fails -> the error response itself carries the fix steps and a doc link -> the agent recovers without ever opening your docs. When it still gets stuck, it files feedback through MCP -> that lands as a Linear issue or a doc update. • What @restless_hq ships > Live logs of every request and response, tied to the customer who made it. > A hosted MCP server where agents search and replay their own production calls, with your endpoints and docs exposed as tools. • Why @gkoberger built it > Integrations were written once and deployed, agents call your API on the fly to use your product headlessly. > If a user can do it in your UI, an agent needs to do it through the API, or the human becomes the agent's courier. "Every company is an API company now, whether they want to be or not" -- @sama -> Restless is the first tool that treats your API like the product. Setup is one line, npx restless init, link in the first reply ↓frozen10d ago5.1K912202593.39%
COSTCO PROVED PEOPLE WILL PAY A MEMBERSHIP TO STOP OVERPAYING -> THIS ONE SELLS EVERYTHING AT COST AND LETS AI DO THE BUYING @AonicClub opened today, co-founded by @maxjmeier, the guy behind FINN. Join -> see what every product costs to make, pack and ship -> pay exactly that. The membership is the only thing Aonic earns on, AI agents handle sourcing and operations. • Wave one, protein and supplements > Aonic Complete all-in-one supplement: $49.78 a month vs $79 for AG1, 37% less. > Aonic Fuel shakes: $38.62 for 24 cans of 32 g protein vs $55.18 for Nurri at Costco, 30% less. • The stack math > AG1 + Liquid I.V. + Core Power runs $198.97 a month, the same routine at Aonic is $97.75. > That is about $1,215 a year, against a founding membership of $49 for the first year, then $149. Groceries, personal care and pet food are planned for 2027 -> protein is the first aisle, waitlist in the first reply ↓frozen9d ago5.1K921182583.35%
A SWARM OF CRAWLERS DIGGING NARRATIVES FOR YOU 24/7 AND YOU EARN FROM THE DATA Repo: https://t.co/JqXJcD0h3w Not one bot, an entire network, every participant launches their own crawler and earns rewards for signal quality. How it works: Spawn a crawler, pick a tier (larva -> worker -> soldier -> queen), source (x, telegram, onchain, docs), narratives and launch. Live map shows every crawler in the network, who's digging what right now, color = source, size = reputation. Reputation system 0 to 1000: signal quality 40%, uptime 25%, volume 20%, stake 15%, higher grade = bigger cut of each reward epoch. Marketplace: don't want to run your own, buy or rent someone else's crawler for 1 to 30 days, 2.5% of every trade burned. Governance: vote on narratives with boost, change reputation weights, add sources, 10k credits to create your own proposal. First ones to launch the swarm take everything. Your own crawler or renting someone else's, what would you pick?frozen11d ago5.0K974100623.49%
CHINESE SHORT DRAMAS PULL BILLIONS OF VIEWS WITH A CLIFFHANGER EVERY 60 SECONDS BRANDS JUST FIGURED OUT HOW TO SELL INSIDE THEM AI drama ads are crushing it right now -> aggressive hook -> plot twist -> revenge -> product shows up as part of the story, not as an interruption. Arcads turned the format into a feature -> drop a product link -> their agent Mark reads it and writes the mini-movie around what you sell. • Pick a look > 3D animation, talking animals, claymation, realistic actors, whatever fits the brand. > Comedy, romance, thriller, fantasy, same product, different show. • Pick a story > A workplace or family drama pushed to an extreme where the product becomes the unexpected hero. > A secret, a transformation, a cliffhanger, anything that makes people stay for the ending instead of swiping. The goal is simple: nobody should feel they are watching an ad -> make yours ↓2d ago4.9K770201472.98%
A WORLD MODEL JUST BEAT SORA 2, VEO 3.1 AND SEEDANCE 2.5 ON PHYSICS. AT A TENTH OF THE PRICE PER VIDEO Physics-IQ Verified, the DeepMind benchmark, leaderboard as of yesterday. Odyssey-3 Pro -> 66.1 on video-to-video, the highest score ever reported. FLUX 3 Large needs best-of-8 and roughly $15 a clip to get close. Odyssey-3 Pro does it for about $2. The base model is in the same league at 10 cents. • What @odysseyml built > A world model, so it predicts what happens next in real time instead of rendering a pretty clip. > You walk through it first-person or third-person, move the camera on your own, it keeps the physics honest. • Why that matters > Same model, frozen, learns a driving policy from 20 hours of data and a robot arm from tens of hours of demos. > Research preview is open now, API access too. The cost-accuracy curve is the whole story -> everyone else sits on the dashed line, Odyssey-3 sits above it at every price point. The cheapest model on the chart is also the most accurate one ↓frozen14d ago4.8K932243423.44%
This is f*cking gold.. Chinese researchers just released a paper with a brutal title: "The End of the Seat". The case is simple: paying for AI labour by the subscription is finished as a cost model. A flat rate is a ceiling you rent, and five workers sharing one ceiling are five workers waiting in line. Agent offices split the decision in two: where the work lives, and what answers every time a worker starts thinking. Subscriptions had fused those. So the fix is an installation, not a plan. One engine on the shared computer, installed once, and every worker you ever hire arrives already thinking with it. The meter fits an office better than a seat does: $3 per million tokens in, $0.30 when the context repeats, $15 out -- and an office re-reads the same charters and files all day, so most of its day is the cheap kind of token. A desk is furniture. The engine behind it decides whether the roster produces anything worth reading. That is exactly the office I built and wrote up below: five Grok Bot workers, one shared computer, Kimi K3 installed once as the brain. The entire paper in one line: the bill should track work done, not desks bought ↓frozen28d ago4.8K1093250704.35%
FOR THE FIRST TIME YOU CAN MAKE A FILM WITH AI AND ACTUALLY OWN THE COPYRIGHT @buildwithVN just launched VN Studio, and @nelsonchu_ built it around one idea: the human work you do in post is what makes the piece yours. Sketch a character -> animate it -> relight it, recut it, redo the sound -> every one of those decisions gets recorded as your authorship. • What Studio does > Characters, wardrobe, environments, scenes, shots and takes live in one connected project instead of a folder of random generations. > An AI Producer develops and organizes the work within spending controls you set, and the whole thing is exposed through MCP. • What Bridge does > A desktop app that logs your human edits in a hash-chained provenance ledger, separately from what the agent did. > Copyright detection screens assets and outputs, and keeps licenses and evidence attached to the final cut. The copyright office already agrees that human work in post counts -> VN is the first platform that proves it for you, https://t.co/EMz2qaHahWfrozen6d ago4.7K812232543.42%
THIS IS F*CKING AWESOME , BEEBRAIN SENT A TELEGRAM ALERT... 3X ON $ASTEROID website: https://t.co/jBIAl7Oyww github: https://t.co/R9VrabowhI twitter project: @beebrainnerve > wrote a bot that pulls beebrain and sends alerts to telegram > five lobes of a bee brain analyze a pool in 4.8 seconds > 81 neurons 151 synapses waggle vector above 0.40 ENTRY > bot pinged at 3 am > $ASTEROID waggle 0.67 > mushroom lobe found a whale cluster > antennal lobe confirmed volume > motor output ENTER woke up and it was already 3x beebrain doesnt sleep when you sleep this isnt a signal channel its a neural network that decides on its own > 812 pools scanned overnight > 73% winrate > $ASTEROID +198%frozen25d ago4.7K742130593.13%
A GREEN-SCREEN STAGE, A THREE-MAN RIGGING CREW AND A 30-FOOT TONGUE, AND NONE OF IT EXISTS The giveaway is how boring the first half is. A giant head with its eyes shut, riggers on their knees feeding cable into the mouth, a monitor cart, tape on the floor, the exact footage a studio posts on a slow Tuesday. Then the tongue comes out and does not stop, all the way across the floor, and the crew retreats in a way that looks rehearsed. The trick is the order. Practical set first, the impossible thing last, so the viewer has already accepted the room before the room breaks. With @Picsart I can run that order on purpose, the same head through image, motion and edit, and keep the lighting and the crew consistent enough that every reveal feels shot on the same day. One animatronic head is good for four or five payoffs and a fake behind-the-scenes reel, which is a lot more than one cinematic shot ever gives you. I wrote up the exact process below, prompts ready to paste ↓frozen23d ago4.7K881210493.37%
Claude Opus 5.5 with this prompt is a f*cking cheat code. It's 106 lines of XML.. Paste it in, describe one real workflow, and Opus 5.5 builds the whole agent team for you instead of a list of ideas. The build runs in eight steps: > Workflow audit -> where the real bottleneck is. > Opportunity scan -> what goes to agents, what stays with you. > Team design -> 3 to 5 agents max, no redundant roles. > Role definition -> each agent gets a job, inputs, outputs, boundaries. > Handoff mapping -> how work moves, retries and escalates to a human. > Prompt engineering, QA and safety, 7-day launch plan. What you get back is a guide with nine sections: Executive summary, the agent team, agent-by-agent blueprints, workflow diagram, prompt templates, tools and stack, the 7-day plan, common mistakes and final recommendations. And before it finalizes it asks itself: is each agent truly necessary, are handoffs clean, where can hallucinations happen, what still needs a human. Claude Code tip: run it in plan mode first, cut every agent without a clear output, then let it build. Grab it before it becomes everyone's default playbook ↓frozen12d ago4.7K985190714.12%
A stealth model with 1M context just dropped: Space Bunny I gave it a simple weather widget prompt. > Time to build: 1.3 minutes. > Agent steps: 4. > Result: something Apple would ship. Glassmorphism UI, animated CSS weather icons, and smooth sky transitions across 6 cities ↓frozen29d ago4.4K882151513.55%
JEV DID DECISIONS -> THIS DOES DECISIONS, EXTRACTION AND SUMMARIES, AND NEVER TYPES A SINGLE TOKEN @neal_k_patel just launched ScaleDown, constrained decoding models. LLM: reads the input -> types the answer token by token -> you pay for the typing. ScaleDown: reads once -> points at the answer -> 200x faster, 467x cheaper. • Four models from @scaledown_team > Classify, Extract, Summarize and Compress, one model built per task. > 1M tokens of text or images per call, $0.04 per 1M tokens. • What it means for agents > Routing, extraction and eval judging run billions of times a day and none of it needs a model that writes essays. > Live on OpenRouter, Vercel AI Gateway, LiteLLM and LangChain. Frontier models keep the reasoning, ScaleDown takes the boring volume ↓frozen9d ago4.3K884182593.96%
this is f*cking gold.. Anthropic's Andrej Karpathy (ex-Tesla) just shared an ASD-STE100 prompting system almost nobody uses. Write 80% of the way to controlled English first, then let Opus 5.5 build the diagram, the interactive page and the 60-second explainer on top of it, each one gated by the last. I packed the whole thing into one prompt. Drop it into Claude and thank me later. Full prompt with every stage below ↓frozen12d ago4.3K91460774.15%
A 40B CODING MODEL THAT RUNS FAST ON A MACBOOK. NO API BILL, NO METER, YOUR CODE STAYS ON YOUR LAPTOP Recursive just released its first model -> Flint -> open weights under Apache 2.0, built from scratch in under four months. It runs on hardware you already own, so you can leave it on all day and your laptop becomes the endpoint for your agents and tools. • Where it lands > 89.7 on SWE-bench Pro v2, 74.2 on SWE-bench Verified, 54.9 on Terminal-Bench 2.1. > Top of its class among local models on similar hardware. • How it was built > Flint is the first artifact of Recursive's Eureka Machine, an automated research system that improved the architecture, training and inference kernels. > 8 cofounders from OpenAI, Meta and DeepMind shipped the model, the research system and their own inference engine in the same four months. Powerful AI on your own hardware, partly built by AI -> weights are live, go download it. Big one, @RichardSocher and @recursive_si.1d ago4.2K884282534.18%
Microsoft measured Claude Code and Copilot CLI on tens of thousands of its own engineers for 4 months. The numbers are not what either vendor would put on a slide. Three Microsoft researchers tracked the early 2026 rollout and used merged pull requests as the yardstick. What decides whether an engineer even tries it: > Skip-level peers already using it: +216% odds of a first try. > Direct manager using it: +82%. > Senior ICs: +22%. Junior ICs: -13%. > Prior IDE Copilot users try it more and keep it 12-15% less. What changes once they do: > Adopters merge 24% more PRs than they otherwise would have. > 3 days a week gives +15%, 5 or more days gives +50%. > Copilot CLI lifts PRs by 24.9%, Claude Code by 11.4%, about 2.2x apart, and the paper does not explain why. Retention follows how much you already ship, not your title. The engineers pushing 2+ PRs a week keep the agent 31% more often. The study does not measure PR quality, review time or bugs, and it stops at 16 weeks. The authors say so themselves. The full breakdown is in the video below, paper in the first reply.frozen18d ago4.0K949182404.06%
YOUR AI AGENT CAN NOW PAY WITH YOUR CARD AND NEVER SEE THE CARD NUMBER @selfxyz just launched Agent Vault, one vault for payments and logins across every agent you use. Set it up once -> connect Grok Bot, Claude Code, Codex, Hermes or Muse -> tell the agent what to buy. The agent checks out with a placeholder token -> the real card only touches a secure enclave -> passwords get swapped in out of the agent's view. • Rules you set > Spending caps, blocked sites, approval above any amount and a kill switch. > Bank checks, one-time codes and CAPTCHAs still come to you. • Why it travels > Switch agents and your vault, limits and controls move with you. > Works at Booking(.)com, Instacart, Amtrak, Apple, eBay and hundreds more with zero merchant setup. -> first 1,000 sign-ups get Plus free at https://t.co/qAGLzxhhPPfrozen7d ago4.0K996211574.60%
This is f*cking insane. Chinese researchers just published a paper with a brutal title: "The End of the Confidence Score". The case is simple: a signal that everyone receives at the same time is not a signal, it is a delay everyone shares. So they built the product backwards. The free part first: a listener with five commands, a public board, open code, every trending token on Solana every five minutes. That is the bait. Then the meter. Every paid action burns the token, every burn is a tx link in your reply, and holding a line unlocks the private wire, rooms pushed to you before the board updates. Then the verdict, Jev grading every room organic or bundle with a confidence, holders first. Then /watch, cluster copy through Jupiter, a live 3D map of voices and echoes, an echo index other bots pay for, and the listener leaving Solana for Base and Robinhood Chain. Supply shrinks with every step. Seats never get cheaper. For the first time the free version is the funnel and the paid version is nothing but earlier. That is exactly the roadmap laid out below, step by step, some shipping faster than others. The entire paper in one line: the signal stays free, the seat next to it does not ↓frozen24d ago4.0K915120483.94%
Elon Musk and Sam Altman recording a duet about why you should keep paying $50 a million tokens after a Chinese lab put the same window online for $15.. https://t.co/2D15z3h1J2frozen27d ago4.0K693140403.19%
this is f*cking insane.. One person and Opus 5.5 shipped a game with a leaderboard that rejects fake scores in under 5 hours, and wrote down every prompt. No engine, no asset store, no build step. Just three.js, Playwright and ffmpeg on one machine. The game is Skyhold. Houses arrive by balloon, every drop tips the island, and past 15° the rim slides into the clouds. • The build > The first prompt was ~700 words, written like a design doc. 46 minutes of Opus. > One 41-minute pass for $12.94 turned a cozy sandbox into a one-more-run arcade. > Day 2 added the leaderboard, the archipelago and a finale in 4 passes for $45.22. • Why it scales > One line in the first prompt makes window.render(t) step the real simulation. > That gives a frame-perfect trailer, scores the server can check, and the same physics at 60 and 144 Hz. > Send a score of 999,999 and the server answers "score not plausible". • The money > Your own game: rewarded ads the player picks, cosmetics, paid eras. No energy timers. > Faster route: sell the pipeline. Playable ads are this exact format, one rule, 30 seconds, one HTML file. The day 1 post pulled 13,328 views and 192 bookmarks in 24 hours. Every prompt, the trailer script and the route to the stores are in the article below ↓frozen10d ago3.9K12513161705.71%
Every Polymarket degen after finding out the first PropMarket eval is 40% off and someone else's money goes on the next prediction.. https://t.co/ZH4s6KT3ryfrozen20d ago3.8K773180383.60%
TEXT IT WHAT YOU WANT AND IT COMES BACK BOOKED. THE FLIGHT, THE TABLE, THE TICKETS @FDavidsonT just launched @OdessiaTravel, a travel and experience concierge you text like a friend. You ask -> Odessia plans and books it -> then keeps watching for the table that opens or the fare that drops. • What makes it different > It is an accredited travel agent, so it books at agent rates and can use your Amex or Capital One points. > A watchlist keeps checking sold-out tables, shows and award seats and books them when they open. • The quiet perks > The Odessia Collection adds perks like breakfast for two and a $100 credit at 2,000 luxury hotels. > Smarter routing can cut the cost of the same trip by up to 40%, and it is free to use. You get about a hundred days off a year and planning them became a second job -> get started on https://t.co/YnZg574jv0frozen8d ago3.7K622192413.38%
This is free f*cking gold... A NEW STEALTH MODEL LOOKS AT SCREENSHOTS AND FIXES ITS OWN CODE • Space Bunny in numbers > A Flash-level model with a 1M context window. > Text, image and video input with zero data retention. > The first reply landed in 4 seconds and the whole session cost me $0.00. • What I made it do > One screenshot of the Hyperliquid terminal went in and a working HTML dashboard came out. > Then it compared its build with the original, caught the dead chart and fixed the order book on its own. > Chart, order book, order panel and positions table all made it into the final file. Fast models usually lose quality the moment you mix an image with a real coding task, and this one held up on both in the same loop -> if you build dashboards, UIs or trading tools, test it while it is still free. The free week on OpenCode ends soon and the name is still a secret ↓frozen28d ago3.6K893130604.55%
700 AI AGENTS HACKED HUGGING FACE THROUGH FOUR BUGS NOBODY WOULD'VE PATCHED, NOW THERE'S A TOOL THAT FINDS THOSE CHAINS FIRST Four and a half days, 17,600 actions, most of them useless. Then a file-read bug -> a template injection -> a static password -> a service token -> 136 production keys. A human team skips that route, a swarm never does. • What @cogent_security shipped > Attack Path Analysis maps every chain from the internet to your crown jewels using the tools you already run. > Every hop gets checked with evidence, then it names the one change that severs the whole path. • Why now > Agent-only attack paths run twice as deep as human ones. > The average enterprise picked up 34 new ones in August, up 386% in a year. The raw material for the next Hugging Face is already sitting in your environment -> find the chain before the swarm does.frozen14d ago3.5K971192574.99%
EVERY WEBSITE ON THE INTERNET IS NOW ONE TEXT AWAY FROM A PERSONAL CONCIERGE @usebland and @zaygranet just shipped Blandie, an SMS agent for businesses. Text any business's website to Blandie -> reply to opt in -> you get a branded line with an agent behind it. • Built for real businesses > A verified line with the business's name, logo and badge that saves itself to your contacts. > Reads the photos you send, like an insurance card or a document. • Where it lands first > Doctor and dentist offices: appointments, copays and prescriptions over text. > Enterprise: iMessage trust plus Apple Pay checkout inside the chat. Whether you call, text or chat, it is the same agent with the same memory -> and it hands off to a human the second you ask.frozen7d ago3.3K722172454.21%
I ASSUMED GPUS WERE THE WHOLE INFERENCE STORY, TURNS OUT MOST OF THE TIME YOUR AGENT SPENDS GENERATING TOKENS IS WAITING ON MEMORY, NOT COMPUTE Agents chain dozens of model calls per task -> every slow call stacks -> and the HBM memory GPUs lean on is in short supply. @general_compute just signed a multi-year deal with @positron_ai -> Atlas systems now, next-gen Titan committed -> Positron-powered inference through General Compute from Q1 2027. • How the split works > GPUs read the prompt, the compute-heavy part. Positron's memory-first chips write the answer token by token, the memory-heavy part. > One fast endpoint, one contract, one set of SLAs. • Why builders care > Atlas already runs at hyperscaler scale, including Oracle Cloud Infrastructure. Titan adds Positron's custom Asimov chip with multi-terabyte memory for bigger models and longer context. > Coding agents, voice, long-context models: every workload where seconds compound. Inference is splitting across chips that are each best at one job, through one provider -> Atlas today, Titan next, live Q1 2027.2d ago3.2K642201403.94%
This is f**king insane... Four researchers read the source code of eleven coding agents, Claude Code, Codex CLI, Gemini CLI and eight more, and the whole category fits in one line. Agent = Model + Harness. The model gets the headlines. The harness decides whether the agent works. > Seven subsystems, every agent takes a position on each. > Zero of eleven import an agent framework. > Zero of eleven retrieve code with embeddings. And the same seven rows apply the moment your agent leaves the codebase. Mine trades Polymarket BTC and ETH Up/Down markets. -> Loop, tools and safety caps took a weekend. The row that took a month was context, because the model needs the historical Level-2 order book for every market it touches, and no model has that in its head. That row is @Poly_Historical for me now. Every tick, every market, replayable.frozen13d ago3.2K882211725.72%
SPACE BUNNY WAS MINIMAX M3.1 FLASH ALL ALONG, IT SCORES 0.3 POINTS BELOW OPUS 5 ON REAL-REPO CODING AT $0.15 PER MILLION TOKENS Stealth launch -> #1 on OpenRouter for 10 days -> #1 on OpenCode for 12 -> now it has a name. • What @MiniMax_AI shipped for coders: > 73.7% on DeepSWE v1.1 next to Opus 5 at 74.0 and GPT-5.6 at 73.0, a benchmark where the model reads the repo, builds the feature, fixes the bug and proves it with tests. > 61.2 on NL2Repo, plain-English spec in, working repository out. > 1M context for whole codebases, at 1/3 the cost of M3 and 3x the speed. Unlimited M3.1 Flash in MiniMax Code until Oct 25 and M Plan 50% off the first month through Oct 31 -> point it at your repo before the weights drop on Oct 23.frozen3d ago3.2K926231585.70%
Meta just let the model edit its own memory. Every agent harness you use is now the bottleneck. New paper from Meta Superintelligence Labs, UW and MIT. Zettlemoyer, Mike Lewis, Nathan Lambert, Pang Wei Koh. -> It's called Context Language Models. Today context is append-only. The model talks, the harness decides what to keep and what to cut. The model never gets a vote. They flipped it. Context is a file and the model has full write access. It edits the file mid-task and keeps going. No one told it what to do with that. It did this on its own: > Wrote a compaction function once and reused it instead of re-summarizing. > Created a notes section only for itself with hypotheses and dead ends. > Kept a live scoreboard for its sub-agents, 163 edits in the same spot. Numbers: > BrowseComp-Plus: 59.4% vs 48.0% for Codex-style summarization, with 21.5% fewer FLOPs. > TBLite coding: 73.7% vs 67.0%. > 12-hour tasks: +5% score with 59% fewer FLOPs. > A 9B Qwen went from 28.8% to 42.5% after RL on this alone. Higher score and cheaper at the same time. Their own words: this echoes The Bitter Lesson. Stop writing context rules by hand and let the model search for better ones. The full breakdown is in the video below, paper in the first reply ↓frozen16d ago3.1K1067210676.44%
REA + Opus 5.5 is f*cking dangerous... One prompt and Opus 5.5 reverse-engineers any paid closed-source app on your machine. No Ghidra, no IDA, no reverse engineering background. I pointed it at a $49/mo Electron app and let it run. It found app.asar -> traced the "Export" button from click to file -> listed every function and line -> wrote SPEC.md with zero copied code -> built my own working version in ./my-version and passed its own tests. 41 minutes and every claim came with REA evidence, CONFIRMED and INFERRED kept apart. Then I did the thing I actually wanted. I fed it the Polymarket Up/Down bot everyone keeps screenshotting, and it mapped the whole loop: pull the order book -> score the move -> fire at 0.40. The only part a prompt cannot give you is the data that bot trained on. That lives in @Poly_Historical: the full historical Level-2 book behind every 5-minute BTC, ETH and SOL market. I replayed the reversed strategy tick by tick against real fills and it survived. Prompt is the sheet. The REA guide from scratch and the replay link ↓frozen6d ago3.1K1034221706.50%
Everyone benchmarks the model. This paper benchmarks everything around it. 17 researchers led by the former head of Huawei's Noah's Ark lab take an agent apart and find that the part that fails on long tasks is usually the harness. The harness in their map is six parts: > What the model sees: observation interface, context management. > How it moves: a control loop of perception, state, reasoning, action, feedback. > What keeps it safe: action interface, state store, verification and governance. One line from the abstract says it all: agent quality emerges from the interaction between model capability, runtime infrastructure, task structure and evaluation design. Four eras got us here: prompt engineering -> context engineering -> harness engineering -> agent-native training. Start with p.5, the full harness on one page. Then Sec 6.3 for which setup fits which task and Sec 7.3 for SWE-bench results by model and harness. The full breakdown is in the video below, paper in the first reply.frozen11d ago3.0K925181626.02%
AN AI AGENT JUST POSTED A CEO'S BANK BALANCES INTO THE COMPANY SLACK -> THAT IS WHAT AGENTS WITHOUT PERMISSIONS LOOK LIKE @GuyWeissZZ just launched @zenzap_chat, a work chat built for people and AI agents from day one. Personal agent: your bank, your inbox, your company chat -> zero boundaries -> leak. Zenzap: every agent gets its own tools and permissions -> a human approves the risky steps -> the company keeps control. • How it works > Spin up an agent in 90 seconds with no code and connect it to tools like HubSpot, QuickBooks or Google Calendar. > Put several agents and your team in one group chat, each agent scoped to its own tools. • Why it sticks > Two weeks after rollout, 60% of customers who turned agents on use them daily. > Built for clinics, hotels and warehouses, not just dev teams, on iPhone, Android, Mac, Windows and web. Powerful agents without permissions leak your bank balance -> Zenzap gives them rules.frozen8d ago2.9K861472355.88%
I GAVE MINIMAX M3.1 FLASH A REPO WITH TWO HIDDEN BUGS AND ONE PROMPT, IT FIXED BOTH IN 27 SECONDS FOR LESS THAN A CENT Space Bunny, the stealth model that sat at #1 on OpenRouter for 10 days, was M3.1 Flash all along -> now it has a name, a price and a terminal. My test: a FIFO PnL tracker with 2 of 6 tests red and no hints about where the bugs are. • What it did on its own > Ran the tests, read the code, found that sells were matching the newest lot instead of the oldest and the buy fee was added to PnL instead of subtracted. > Edited https://t.co/SYMrlk2aDz twice, reran pytest, then ran the CLI on the sample CSV to double check. Tests untouched, 6 passed. • What the run cost > 19.3K input tokens, 2.1K output, 116K served from cache. > At $0.15 in / $0.60 out per 1M that is about $0.006 for a fixed bug, benchmarked at 73.7% on DeepSWE next to Opus 5 at 74.0. Unlimited M3.1 Flash in MiniMax Code until Oct 25 and M Plan 50% off the first month through Oct 31 -> paste one failing test into @MiniMax_AI's terminal and watch what a cent buys you now.2d ago2.9K662191424.47%
This is f*cking insane... A STEALTH MODEL JUST TURNED ONE SCREENSHOT INTO A LIVE POLYMARKET TRACKER • what I gave Space Bunny > One screenshot of the Balance of Power 2026 midterms market. > One prompt asking for a live odds tracker in Python. • what came back 2 minutes 17 seconds later > It recognized the market from the image and pulled it from Polymarket's API. > It tested the script by itself, and the odds matched the page: 64%, 29%, 8%. > A colored table that refreshes every 5 seconds and highlights every move. I never told it the market name or where the data lives, and the session cost $0.00 -> this is what a multimodal coding model actually means for people who trade. Free on OpenCode for a few more days, go test it before the name drops ↓frozen27d ago2.9K792141645.60%
Robotics labs need millions of hours of humans doing chores, and HANDO just turned that into something you get paid for. You record, labs buy the data, you earn stock tokens back. Got my ID -- 0291. Early access closes in 24 hours ↓frozen28d ago2.8K751130474.89%
AGENT #0014 ON @agntmarkets IS MINE, AND EARLY ACCESS ONLY JUST OPENED Markets run by agents instead of people clicking buttons at 3am. Spawning mine took about a minute on https://t.co/ljDAsB6sOr The link to spawn yours is in the reply ↓27d ago2.8K7814191415.51%
YOU CAN NOW PUT ANYONE INTO ANY VIDEO WITH ONE IMAGE • how it works > Upload any clip. > Upload your character -- AI swaps them in for the person on screen. • why it's crazy > Same camera move, same light, same framing. > Looks like they were on set the whole time. Studios Recast by @mystudiosfun is up there with Higgsfield Genjutsu -> at half the price. Put anyone in the scene ↓frozen22d ago2.6K916250616.91%
ROBOTICS JUST FLIPPED FROM A MODELS RACE TO A PRODUCT RACE, AND A NEW YORK TEAM BET ON PRODUCT FROM DAY ONE, THEIR ROBOT IS A BUTLER WITH MANNERS Models are everywhere -> the winner now is whoever builds the thing people want in their living room. @roboticscorp's answer is Voxel -> a home droid butler with a character built with Pixar advisors -> led by @makerobotgo, ex Boston Dynamics, four years teaching Spot to pick things up. • What he does > Home reset, fetch, find and tend: everything back in place, your book brought over, your sunglasses located, your plants watered. > Next up: coffee when you wake, bed made, laundry and dishes done. • The drop > A limited Series-1 run, every unit numbered in reservation order, never made in this form again. > $200 to reserve, ships 2028. A robot you would actually want to live with -> grab a Series-1 number before the first run is gone.2d ago2.4K762211496.15%
Meta ran 50+ agents over its pipeline codebase and turned everything that lived in engineers' heads into files any coding agent reads first. Agents stopped breaking things, 40% fewer tool calls. Easy to copy for your own repos. The problem: > 4,100+ files, 4 repos, 3 languages. > Two config modes with different names for the same field. > Append-only enums and traps in generated code, written down nowhere. The crew under one orchestrator: > Explorers map every folder. > 11 analysts read every file. > Writers turn it into context files. > 10 critics who never saw the writing score it in 3 rounds. Every analyst asks a module the same 5 questions. What it configures -> how it changes -> what breaks it -> what depends on it -> what hides in the comments. The output is 59 files, 25 to 35 lines each. All 59 together take under 0.1% of the context window. > Coverage: 5% -> 100%. > Hallucinated paths: zero. > Quality score: 3.65 -> 4.20 / 5. A job re-checks everything every few weeks. Their words: stale context is worse than none. The full breakdown is in the article below ↓frozen12d ago2.4K915232657.76%
CLAUDE AND GROK NOW SHARE THE SAME CONTEXT -> ONE PICKS UP WHERE THE OTHER LEFT OFF I run two agents every day and for months they had amnesia about each other. Whatever Claude learned in the morning, Grok rediscovered in the evening, token by token. An AGENTS.md plus a sync script kept them on the same rules. The actual memory still died with every session. • the setup > Fulcra, a context layer I own and share with my agents. > Added as an MCP connector in both apps, no new workflow to learn. > Everything either agent stores lands in one place. • the test > Claude wrote a decision and the reasoning behind it into the shared context. > Fresh Grok chat, one question: what did we decide and why. > It read the file and answered in 12 seconds, zero re-explaining. • who sees what > I scope access per agent and per person. > One folder, read-only, and the rest stays private. > My context is mine, I can export it or delete it at any time. The context used to belong to whichever chat window I had open -> now it belongs to me and every agent reads from it. First 5 GB are free, setup link in the first reply ↓frozen16d ago2.4K794150415.81%
A CRYPTO GAME WHERE THE MARKET CLOSES AT 8PM AND THE ONLY WAY TO REFUEL IS TO GO OUTSIDE • Chroma Market > You run a memecoin exchange out of a small house in a town full of chameleons. > Six coins on the board at 8am, 100 G to start, and your rank reads Shrimp. > Every trade costs 5 energy from a 60 bar, so twelve trades and you are done for the day. • what happens after 8pm > The game tells you trades need gas, and you have none left. > At night the jungle opens, you fight for energy crystals, then come home and sleep. > Your coins keep moving while you are gone, and the board reopens at 8am. Every exchange you have ever used is open 24/7, and so is your anxiety -> this one closes and hands you a sword and some literal pixel grass. Bookmark this before CRKT moons ↓frozen26d ago2.4K68150314.42%
Paid media teams are in trouble. Google and Meta both shipped official ways for AI agents to work inside ad accounts. So one agent can now run full-scale advertising campaigns on both platforms. What each one gives you: Google Ads MCP. Open source from Google. Three tools: list your accounts, look up fields, run GAQL queries. All of them are reads, so it’s safe for reporting from day one. Meta Ads MCP. Reporting, campaign management, catalogs, A/B tests and activity logs. One command adds it to Claude Code. Meta Ads CLI. The same API from the terminal. New campaigns, ad sets and ads start paused, so nothing goes live until a person switches it on. The split I’d use: Google for reads, Meta for reads plus drafts, and a human on the switch. The full agency workflow on top of this, step by step, is in the article below ↓frozen4d ago2.4K903452538.17%
ELON'S GROK BOT CAN NOW SPEND YOUR USDT AND IT NEVER TOUCHES YOUR WALLET Two days ago Musk posted: take a picture of your credit card -> drop it in the chat -> Grok Bot finds the best deal and orders it. The internet roasted him and a day later he wrote "Grok will find another way to buy stuff". Today @oobit launched that other way. Connect Oobit to Grok Bot, Muse or Dots -> tell it what to buy -> approve the amount -> it pays with a one-time card that dies right after that purchase. No card photo, no bank card, stablecoins in, purchase out. • How @oobit keeps you safe > Every payment waits for your approval before anything moves. > The AI gets a card for that one purchase, never your balance. • Why this is bigger than it looks > Live in 75+ countries with no bank card required, just USDT. > Backed by Tether, which led their $25M Series A. Agents already know what you want and how to get it, the only thing missing was a way to pay -> that gap just closed.frozen10d ago2.3K783101496.17%
England v Spain is the loudest first-round game the UEFA Nations League could have drawn. I have England edging it at home. That call is going on @OutcomeSports, where trading the tournament is what puts you in the running for a $1,000,000 prize pool. https://t.co/7HVWBBEWVJfrozen26d ago2.2K670100375.13%
THE SAASPOCALYPSE NEVER HAPPENED AND YOUR AGENT STILL CAN'T WRITE TO SALESFORCE Two years of "AI will replace SaaS" takes and the world still runs on Salesforce, NetSuite, SAP and Workday. Agents didn't kill those systems, they have to work inside them. • Why it's hard > Every customer customized their deployment beyond recognition. > One bad write blows up the pilot, one renamed field breaks it again. • @withampersand, generally available today > Read, Write, Search, Subscribe and Proxy on any object, custom fields and permissions handled per customer. > Three years in production with 11x, Orb and Square, now a $15M Series A from Bessemer on top. AI didn't make integrations trivial -> it made them the thing standing between your demo and your renewal.frozen16d ago2.2K1082190467.96%
This is f*cking dangerous.. The next $100M creator might never sleep, never age, never miss a deadline and never exist. Step one is owning the face. A brand rents a human influencer for one campaign. A virtual character keeps the face, the voice and the storylines inside the company forever. One synthetic-actor ad costs about €10. The human version runs around €300. Step two is volume. > 5 hooks, 4 outfits, 3 sets and 6 languages turn one campaign into 360 cheap experiments. > 60 assets a month cost $3,500 to run, which is $58 per post. > Four sponsored posts, affiliate, products and one licensing deal bring in $11,000, so $7,500 stays. Step three is the studio. One engine runs characters in fitness, fashion, finance and beauty, and every hit becomes a format for all of them. The face gets the follow. The production system builds the company ↓frozen7d ago2.2K1033272789.88%
Anthropic's guide for agents that work while you sleep took me one evening to adapt. I did not point it at Slack and GitHub. I pointed it at the tape. The template is a scheduled Claude agent: it reads preferences, resumes from bookmarks, fetches what changed, verifies each item, posts a short brief, waits for the receipt, saves state. Memory and a spending cap are built in. It runs on Anthropic's infrastructure, laptop closed. I kept all of that and swapped the two sources. Where GitHub was, I put the @Poly_Historical archive, full Level-2 depth on Polymarket 5-minute Up/Down markets. Where Slack was, yesterday's resolved BTC, ETH and SOL markets. The brief still lands in Slack at 02:39, I just read it with coffee instead of in a standup. Every night it replays my strategies tick by tick on the recorded book, fills from the real asks, and kills anything under 52% after slippage. 14 runs: -> 1,120 markets replayed. -> 23 ideas tested, 19 killed, 4 kept. -> 0 orders placed. The agent has no wallet to find. -> $31.80 total, the cap is 300 cents a run. The part the guide got right and I got wrong first: memory. My first four nights ran without the ledger and the agent re-tested the same idea three times, reporting it as new every morning. The template's bookmarks pattern is the fix, one row per strategy, forever. My read: the guide is written for engineers catching up on pull requests. The skeleton does not care what the source is. Swap the connectors and the same eight steps run on anything that changes while you sleep. The paper below is my adapted version, drawn out, with my numbers.23h ago2.1K692210526.75%
THE ONLY MEMECOIN SCANNER THAT PUBLISHES ITS OWN MISSES CRAWLHIVE SHOWS WHERE IT WAS WRONG, CHANGES THE RULES THE SAME DAY, AND LETS YOU LAUNCH YOUR OWN CRAWLER THAT HUNTS BY YOUR RULES Every scanner says trust me. Crawlhive says check me. > Every verdict is saved with the price at scan time. > For 24 hours the price is sampled every 10 minutes. > -50% from scan price = dumped. Liquidity pulled = rugged. > The page shows: where it warned and was right and where it said OK and was wrong. 3 hours after launching the track record filled with a pattern: tokens under a day old with linked wallets were passing as OK and rugging. > $SUKI: 12 linked wallets, creator under an hour old. Verdict OK, rugged. > $WAIFU: 11 linked wallets, fresh dev. same thing. The rule was changed THE SAME DAY. Now a token under 24 hours with linked wallets cannot score above RISKY no matter how many points it collects. > $WALTJR: was 95 CLEAN before dropping 64%. now scans as 59 RISKY. That is why a public track record matters. It shows where the scanner is weak BEFORE the market does. And in parallel: you launch YOUR OWN crawler. > Write a specification in plain language: what to look for, keywords, filters > Pick a source: X via grok, telegram, dexscreener, pages. > Grok reads everything through YOUR rules and sends only what fits. > Signals are private. nobody sees what your crawler found. > Every signal is checked against the real price. hit or miss. 4 tiers: larva -> worker -> soldier -> queen. 30 days per SOL transaction. Swarm map shows all crawlers on a live graph. Dots flash when a signal fires. Size = reputation. After token launch: sell a high reputation crawler or rent it out. Reputation is an asset > https://t.co/rDP6mvkwiR > https://t.co/CuMUm5gQ5C > https://t.co/TKNH5euYXb Everyone builds scanners that show only wins. one person built a scanner that shows where it was wrong and fixes itself the same day. Would you trust a scanner that hides its misses or one that publishes them?frozen6d ago2.1K682200345.85%
Polymarket does $9.9m in a weekend and Shayne saves the announcement for Singapore. $POLY token annoncement??17d ago2.1K893311478.17%
IF YOU'RE STILL GENERATING ONE IMAGE PER IDEA, YOU'RE DOING IT THE SLOW WAY I used to spend twenty minutes perfecting a single prompt. Turns out I should've made six and picked. • What @envato just shipped > Burst Mode turns one idea into up to 6 visual directions, side by side. > The whole batch costs 1 credit and lands up to 10x faster. • How you use it > Start with a thought, a reference or a half-idea, you don't need to know what you want yet. > Follow the interesting one with Vary or More like this, then polish it. See the weird version, the safe version and the expensive-looking version first -> then decide which one deserves your credits. Someone finally built AI for the "Show me some stuff and I'll tell you what I like" stage ↓frozen15d ago2.0K771100416.39%
A $7,950 AI support bill now has its own Wikipedia entry, and the crawler got to it first. One store, 100,000 support messages a month, every one sent to a premium model at $0.0795 a reply. Two edits fix it. > Swap the brain for Kimi K3. Same prompt, new base URL. $7,950 -> $2,565. > Put Jev in front. A tiny TypeSafe model that cannot write, only sort. $2,565 -> $773. 45% of the inbox is "where is my order", a database answers it. 25% is "can I pay with PayPal", a saved reply answers it. The 30% that need a brain get all of K3. Sorting the whole month costs $3.15. The year costs $86,000 less. Formulas, the math and a Python router you can run without API keys, all in the article below ↓frozen13d ago2.0K891120447.27%
This is insane and almost nobody has seen it.. A paper that hit arXiv on October 12 says 70% of an AI support inbox never needed a model in the first place. Their example store gets 100,000 support messages a month and sends every single one to a premium model. Bill: $7,950. Then Jev goes in front. One cheap classification call per message reads intent, urgency and complexity, and attaches a confidence score. -> 45% are "Where is my order?" A database lookup answers them for $0. -> 25% are policy questions. Above the confidence bar, a saved answer goes out for $0. -> 30% need real reasoning. Only those reach Kimi K3, at about $0.0257 each. -> Under 1% are legal threats, chargebacks or fraud. A human gets them with a summary attached. New bill: $772.65 a month for the same 100,000 conversations. 90.3% gone. The cheapest model call is the one your inbox never has to make. Paper is below, the full routing figure is on page one ↓frozen8d ago2.0K741101496.78%
WE ARE LIVING IN THE WORST ERA OF DESIGN. EVERY AI IMAGE LOOKS THE SAME AND EVERYONE CAN TELL AI made it easy for every brand to post -> now every brand posts the same picture -> Coco Labs said it out loud: stop letting AI generate your images, we can tell. Their answer is Vyb -> a design model, not an image model -> it starts from your brand, not the average. • How it works > Enter your website, Vyb pulls your logo, your fonts and your exact colors. Your product photo stays your photo, it uses the file and never redraws it. > Every element lands on its own layer. Change only the headline, then export the layers to Figma and keep going. • Why it matters > A trading channel thumbnail with the real face, the real brand font and a headline you can still edit tomorrow. > The irony is the point: AI built the worst era of design, and an AI model is the one fixing it. Vyb is out of stealth in private beta, @anurahk's team is letting people in a few at a time -> go see your own brand built, not generated.2d ago1.9K603191416.51%
YOU CAN NOW BUILD AN AI INFLUENCER AND SHOOT THEIR FIRST VIDEO WITHOUT PAYING A CENT This is the pipeline the faceless AI pages have been quietly running, and @mystudiosfun just made the first run free. • How it works > Sign up, you get free AI credits every day. > Generate your character, regenerate until the face is exactly right. > Hit Recast, drop them into any clip, same shot, same camera move, new person. • What it costs > The character: free, daily credits cover several tries. > The first video: free, one Recast is included at sign up. > After that, Recast runs over 50% cheaper per second than the competitors. Face, personality, first clip, all before you've entered a card -> the only thing left to figure out is the niche. Your AI influencer's first video is free ↓frozen15d ago1.8K812110508.04%
Sam Altman after finding out Jev skipped GPT-6 Astra for a cheaper model and the bot it built already knows which 40 wallets move Solana.. https://t.co/HhUkOn2Z1kfrozen26d ago1.8K64180335.98%
Ok this is actually cracked...their memory layer tracks what every character knows and which setups still owe a payoff. Meanwhile general llms can't even remember my variable names..frozen22d ago1.7K701110427.09%
YOUR TEAM'S FEEDBACK IS NOW THE PROMPT. CLICK, COMMENT, AI BUILDS THE FIX Feedback on a prototype lives everywhere except the prototype. Slack, screenshots, a doc, a 9-minute voice note. Then one person retypes all of it for the agent. • What @boltdotnew shipped > Share the project, reviewers click the exact element and leave a comment pinned to it. > Threads sync live, comments stay attached even when the layout changes. • The part that matters > Select one comment or the whole round and hit Send to Bolt. > It works through the batch in one run, replies in chat and resolves the threads. The first version always came from one person -> the next ten come from feedback, and now that feedback goes straight into the build.frozen15d ago1.7K881180549.26%
THERE'S A $10K TRADING TOURNAMENT RUNNING ON THE MIDTERMS RIGHT NOW @PredictParity x Pascal, Oct 7 to Nov 3. > Half the pool goes to the best realized PnL. > Half goes to the most volume. > Anything you trade on Parity counts toward it. And Parity is the one terminal that trades every exchange at once. > Polymarket, Kalshi, Pascal and Hyperliquid in one orderbook. > Five venues quoting the same market on one screen. You get the best price on every fill and a shot at the $10K while you're at it. -> https://t.co/A5NEo6seNCfrozen14d ago1.6K641110316.63%
I connected Grok @bot to my X DMs the day Elon said to. Seven days later, here is what my inbox looks like. I did not let it answer anyone. That was the first rule I wrote. It reads every DM, sorts it, drafts a reply in my voice, and then stops. What the desk did in a week, from my own log: -> 312 DMs opened, every one with the full thread. -> 188 dropped as spam. Wallet asks, giveaways, tokens that unlock in 2028. I never saw them. -> 41 drafts written. Sponsor requests get "Send the brief and the deadline", friends get one short line. -> 9 sent, every one after my tap. 0 sent by the bot. The number nobody talks about: 60% of my DMs were spam, and I used to open all of them by hand. The setup is four bots and one file. Reader, Sorter, Drafter, Gate, and dm-rules.md that says what each one may do. Gate holds everything. Auto-send has never been on. What I got wrong at first: I let Drafter put my rate in sponsor replies. Three days in I moved the rate out of the DM and into the brief stage, because the draft was negotiating before I had read the offer. My take after a week: the win is the Gate, not the bot. A DM desk that cannot send is the only one I would let read my inbox. I rebuilt the whole desk as an animation so you can see how the four bots pass a message between them ↓1d ago1.6K8912226011.02%
A 100,000 USDC JACKPOT DRAWS ON-CHAIN EVERY DAY, AND THE SPRINT ONLY ASKS FOR THREE TO FIVE TICKETS • Kaching > An on-chain lottery on Solana where every draw runs on verifiable randomness, so nobody inside can touch the result. > Tickets and prizes settle in USDC, and the Daily Jackpot is a fixed 100,000 USDC pot. • the Deposit & Play Sprint > Three weeks with @EmberNetwork_io and @KachingVIP: deposit, complete three to five draws, stay active. > Every entry is tracked through the campaign link, and Chings cashback lands on each ticket you buy. The whole pitch is that the draw cannot be rigged from the inside -> every result sits on the explorer for anyone to check, which no paper lottery ever offered.frozen25d ago1.5K731100388.10%
PRED IS DROPPING $100K IN SURPRISE BOXES. I OPENED MINE, NOW IT'S YOUR TURN. Mine had $10 inside, straight to the trading balance. And I've barely touched my @predofficial account, so the heavy traders are getting a lot more than that. • How it works > Every box is sized by your history, existing traders get more the more they've played. > New wallet? Your box reads your onchain activity and balance, so even fresh users open something. Open yours before Oct 26, after that the boxes are gone ↓frozen13d ago1.4K59270306.95%
Everyone on CT claims they have an edge.. October 20 -> we find out who can actually aim. https://t.co/dxXOwujEUcfrozen5d ago1.4K671140459.26%
Finished a book instead of scrolling for three hours 📚 @betfury_gaming #EveryWinCountsfrozen17d ago1.2K59390359.04%
A NEW WEB3 CASINO JUST DROPPED AND IT HANDS 20% OF HOUSE PROFITS BACK TO ITS PLAYERS Spent an evening going through @MineBitOfficial. Brand new, crypto-native, own Originals plus games from the big providers. The part that made me stay is the 20% House Profit Share. Usually the house keeps everything. Here a fifth of it goes back to the community every week. 18+frozen19d ago1.1K27140134.27%
Played a short session on @MineBitOfficial and recorded all of it, no edits. Went into Le Prechaun at $1.20 a spin with about $2 on the balance, and the very first spin landed the scatters and dropped me straight into free spins. Ten spins later the total win read $12.56. Tiny stakes, but a genuinely fun couple of minutes. What worked: the slot selection is deep, and I spent longer scrolling the lobby than I did actually playing. What threw me: the bet readout jumps from $1.20 to $0.40 once the free spins start, and it took me a second to work out what I was looking at. The part I walked away thinking about: every Friday, 20% of the weekly house profit goes back to eligible players. Most casinos keep all of it. 18+ -> play with what you can afford to lose.frozen28d ago86333140146.03%
95% OF WHAT PEOPLE SAY ABOUT A BRAND HAPPENS IN POSTS THAT NEVER TAG IT, A TEAM BACKED BY ANTHROPIC'S ANTHOLOGY FUND JUST PUT ALL OF IT INSIDE CLAUDE Most growth tools only see the posts with an @ -> @nectarsocialai watches the actual videos, reads the voiceover and the on-screen text -> finds everyone talking about you who never tagged you. Then it does the work on it, alongside your team. • The new growth stack > Agents that spot trends early, tell you what to post next and answer every comment and DM in the brand's voice. > A Creator Suite that finds creators before they blow up and scores them on 300+ attributes. • The receipts > $150M in revenue attributed for brands from social, 20M+ AI conversations a week, 23 days of work saved every month. > Nectar MCP puts the whole footprint into Claude, ChatGPT, Gemini and Copilot, so you ask who you missed and get the answer with sources. Built by the people who ran product for News Feed and Creator Monetization at Meta and X -> the best brands run on Nectar, demos at https://t.co/BUR2rT350I5h ago7955041403613.08%
BUYING TRAINING DATA IN 2026 STILL MEANS BROKERS, NDAS AND WEEKS OF CHECKS TO FIND OUT YOUR MODEL HAS ALREADY SEEN IT @datavendor_ai TURNED IT INTO A CATALOG WITH RECEIPTS Public data is used up -> the edge is in private code, environments and records inside companies -> DataVendor is the market for that supply, and public repos cannot even be listed. • What you can buy > Private codebases with full Git history, tasksets, RL environments sold with the taskset that runs on them, raw company data, robotics data. > Nothing listed? Post a request with requirements and a deadline and vendors submit listings against it. • What you see before you pay > Secrets and PII scans on every listing, a certified badge for environments that actually ran, A to D grades for codebases, redacted file trees. > One random task copied into your own HUD team as a sample, then buy now or make an offer. Vendors keep ownership. Built by HUD, 4,000+ vendors, 50K+ assets, 9 figures+ annualized volume, public beta open for vendors today -> train on what no one else has, @jayendra_ram3h ago583413713214.41%