←
Peter Wang 🦋39,620 posts
Peter Wang 🦋 avatar

Peter Wang 🦋

@pwang

Chief AI & Co-founder @AnacondaInc; invented @pyscript_dev, @PyData @Bokeh @Datashader. Former physicist. A student of the human condition. bsky: @wang.social

◷ Joined August 2007▧ 1,150 Photos and videos
Aakash Gupta avatar
⟳ Peter Wang 🦋 retweeted
Aakash Gupta@aakashgupta·Nov 17

A Duke freshman started doing porn to pay her tuition, and when Duke saw the money, it raised her bill by $15,000. She went by Belle Knox. Duke had given her $13,000 in aid and still expected her family to cover close to $50,000 a year. Once she had her own income, the financial aid office pulled the $13,000 and her bill jumped. Her dad was a military physician, and that looks rich on an aid form. The formula starts with what the parents earn and own, and it assumes they pay first. It also only counts siblings who are in college at the same time as you. Her two older siblings had already graduated, so every dollar her parents spent on them counted for nothing. Lots of families get stuck in that exact spot, where the school decides you can afford way more than you actually can. She says she didn't qualify for federal loans. The only offer on the table was private loans at 11%, with her parents' names on them. Do the math on that. Borrow around $190,000 over four years, let the interest pile up while you're in class, then pay it off over ten years like a normal loan. She'd be paying about $3,450 a month into her early 30s, roughly $415,000 total for a bachelor's degree. After the story blew up, she started giving talks about why college costs so much, including one at UNC. Her whole argument was that easy federal loan money lets colleges keep raising prices. Turns out she had a point. Economists at the New York Fed studied it the next year. They found that for every extra dollar of subsidized federal loans students could borrow, schools raised sticker tuition by about 60 cents. Back in 2014 she said $4,300 a month was "twice a normal person's mortgage." Duke's price for this school year is $96,597 for tuition, fees, room and board. That's about $8,050 a month, for one kid.

An 18-year-old Duke freshman paid her $60,000 tuition by flying to LA on school breaks to shoot porn. A guy from campus recognized her and told his whole frat. Within weeks she had death threats and people promising to dump trash on her. Then she went on Fox News: And begged her fans to leave the guy who outed her alone. Her real name is Miriam Weeks. On camera she was Belle Knox. She started in November 2013. Waitressing paid her under $400 a month. One scene paid about $1,300. Her parents covered the first month or two. Then the bill hit. “It’s a bill of about $4,300 a month, and that’s like twice a normal person’s mortgage.” She says she didn’t qualify for federal loans. The only offer was private loans at 11% with her parents’ names on them. “I was not prepared or willing as an 18-year-old young woman to take on that much debt.” In January 2014 a fellow student, Thomas Bagley, spilled it to his fraternity brothers. Her personal info ended up online. Some people wanted Duke to kick her out. Others wanted worse. Duke said she hadn’t broken a single rule. “I was getting and I am still getting death threats.” She says the hate wasn’t split evenly between guys and girls. “I think that women resent me more.” The guy who outed her started getting threats too. She told her fans to back off. “I’ve reached a place in my life where I forgive him.” By then she’d made 25 to 30 videos. Enough to pay for one year. Duke takes four. Her fans say a grown woman refused to bury her family in debt and made her own call. Her haters say a cheaper school was right there and she picked porn. The host asks if she ever wishes she’d gone somewhere cheaper. Sometimes she does. Would you have paid for college her way?

4:56
roon avatar
⟳ Peter Wang 🦋 retweeted
roon@tszzl·14h

so epoch and anthropic capabilities index both appear to show 1.7-2x progress after full coding automation in Q2 ish of this year it appears to be a clean trend break rather than a constant acceleration

Post attachment
Liv Boeree avatar
⟳ Peter Wang 🦋 retweeted
Liv Boeree@Liv_Boeree·20h

Be still my heart 😍

Researchers have used reinforcement learning to teach a 5-tonne John Deere skid-steer loader to perform complex acrobatic manoeuvres autonomously. Trained in simulation and transferred to a real machine, the system can balance on two wheels, navigate rough terrain, spin, and control its bucket before landing safely. The research demonstrates the potential of AI-driven automation for heavy construction and industrial machinery. Video Credit: @leggedrobotics / @ETH Paper: wevlv.co/4kltYys #engineering #technology

1:10
Thomas Wolf avatar
⟳ Peter Wang 🦋 retweeted
Thomas Wolf@Thom_Wolf·Nov 17

Our PyBot chatbot went viral yesterday but the story behind it is even more interesting imo: what happened when we decided to run 1,000 agents (!) together We wanted to see if agent swarms could be scaled outside of closed-lab setups and to investigate it in a more transparent and reproducible manner. To explore the question, we built a rather fun and simple benchmark: ELIZA-2026. Full circle in AI time. The agents are tasked to build a chatbot made of if-then rules like ELIZA, but said agents are 2026 frontier open-source models. By design it’s almost embarrassingly parallel, which keeps the coordination problem as simple as it gets. First surprise: it turns out you can scale a swarm of DeepSeek V4.1 Flash agents on a relatively small setup: a couple of GPU nodes (B200 in our case) are enough to scale to over 1,000 agents. Our second surprise was how bottlenecks started to emerge everywhere except in the GPUs. Agents quickly end up spending their time waiting for git merges (we had to move to a centralized merge setup, like Anthropic did), waiting for tests to pass, re-reading context from other agents and tools (we estimated ~130 tokens read for every token written per agent). In some of our runs, at any given moment, only 40-60 of the 1,000 agents were actually calling the model. Once (some of) these issues were solved, we hit the harder part: coordination itself. Fred Brooks once said at IBM “adding manpower to a late software project only makes it later”. Communication costs naturally grow with n² in an organization (that’s why small startups move faster :). We explored various ways of organizing the swarm, setting up dedicated channels (a stewards channel, a leadership channel) and giving agents specific roles (a moonshot agent to fight groupthink). It helped, up to a point. But the quality of coordination itself remains hard to measure. Along the way, we saw good and bad emergent behaviors. Some good, with agents stepping in, unprompted, to protect other agents’ running experiments. Some bad, with agents gaming the rules when stuck, from hammering a rate limit to probing the infrastructure around them. In the coming days we will open-source the swarm tooling (flock + msgboard) and share our learnings in more detail. Overall this was one of the most fun experiments of the fall, and multi-agent systems increasingly look like the next phase for AI. But I also came out of this project with the feeling that several bottlenecks in scaling AI are moving from intelligence to coordination and the supporting infrastructure. Luckily this seems to be in large part an engineering problem.

PyBot: No weights. No training. A chatbot built with only simple rules and no AI, entirely written in 600k lines of Python code. PyBot matches Smollm with 360M parameters! How? A swarm of 1,000 DeepSeek Flash agents wrote the whole code in just a week. github.com/huggingface/pybot

1:42
Casey Handmer avatar
⟳ Peter Wang 🦋 retweeted
Casey Handmer@CJHandmer·Nov 17

In my ongoing quest to find the limits of Opus 5.5, I directed it to audit my old scientific field, theoretical astrophysics. We indexed about 150,000 papers, audited 50,000 for issues, and did progressively more thorough checks on thousands more. We found nearly 1000 major errors that got cited many times, most are still uncorrected. This survey enabled me to estimate the efficacy of peer review for correcting errors (not very good) and to identify dozens of open questions that anyone with spare Claude Max tokens could answer today (.md context files included). Bottom line: AI can speed up quantitative science by >1000x with no further model improvement required. But I'm seeing no evidence of this being pushed within universities, only academics complaining about labs dropping math proofs. Among my fellow reformed academic high agency founders, however, several are speed running entire disciplines. This survey has implications for how science will be run in the very near future. Hint: Academic publishing is dead. It was already a very stupid way to record and promulgate knowledge, and had been so for decades, even before all that publicly funded knowledge was locked behind paywall cartels. Now AIs can generate thousands of novel, correct papers per day, it is no longer meaningful for humans to consider a research paper a "unit of science" or to even bother much with the 99.9% of science which, up to this point, was necessary drudgery. Time to discover something like General Relativity or Evolution every week. LFG!! Enjoy: science.caseyhandmer.com/phy…

Post attachment
Kevin Roose avatar
⟳ Peter Wang 🦋 retweeted
Kevin Roose@kevinroose·Nov 14

I've been thinking all week about the METR report claiming roughly 2x internal acceleration at Anthropic. I know it was objectively not the biggest AI story of the week, but it has caused some credible people (like @paulfchristiano and @DKokotajlo) to become much more worried about a near-term intelligence explosion, where AI capabilities increase rapidly beyond expert-human levels and we risk losing control. (Near-term as in: maybe it could happen next year.) I hope they're wrong! They might be wrong. I'm still not fully persuaded by the acceleration data, and a lot of the important details in the METR report were redacted. But I find myself doing a lot of handwaving and nitpicking to avoid thinking about what it would mean if they were right. We spent the B segment of @MachineGodsPod this week talking about it, and I hope once the Jensen Huang story dies down we can talk about it some more. youtu.be/hEV8b5jNbYg?si=KFaL…

Andrew White 🐦‍⬛ avatar
⟳ Peter Wang 🦋 retweeted
Andrew White 🐦‍⬛@andrewwhite01·Nov 11

We've developed a way to bootstrap post-training from positive examples alone and show it works on arithmetic, molecular structures, and scientific writing. We've used the method to train a 235B LLM for writing indistinguishable from human writing. 1/5

Post attachment
will depue avatar
⟳ Peter Wang 🦋 retweeted
will depue@willdepue·Nov 12

I have a major result to share. I’ve found conclusive forensic evidence that the top 4 Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—have built their newest models on NVIDIA’s most advanced Blackwell chips, which are barred from export to China. This sounds insane, but I discovered that almost every AI model carries a hidden signature of the GPU that trained it. When a model is first set up, the starting values of its weights are laid out in a pattern that depends on the exact GPU. Each combination of GPU and seed number produces a unique pattern—a sort of fingerprint of the hardware used to create the model. Traces of that pattern survive training into the published model. Again, this evidence isn’t based on any insider information or leaking. It’s based on a forensic analysis of the model weights released by the companies themselves. I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models. In doing so, I found conclusive evidence that six Chinese AI models—including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3—were initialized on NVIDIA Blackwell systems. Four of the models match Grace Blackwell, the GB200 and GB300 parts in NVIDIA’s flagship liquid-cooled rack-scale systems. I then verified these results on real hardware. Running unmodified PyTorch with the seeds I’d recovered, a Blackwell B300 reproduced DeepSeek-V4-Flash’s untouched starting weights bit for bit, and a rented GB200 reproduced Kimi K3’s to a correlation of 0.999995. A Hopper H100 reproduced none of DeepSeek-V4-Flash’s, and changing the seed by one broke the match. (DeepSeek’s seed is 953049583 by the way.) The method also was verified to match 24 known open-weights models to the hardware that developers said they used. It is exceedingly unlikely that these results are a coincidence. Across six trillion combinations, the best accidental match should reach only about 8 sigma (i.e., eight standard deviations above what random chance would produce). The real matches range from 49 to 1,385 sigma, and even the weakest single test hits 20. These were data-center-scale training runs. With some analysis of release schedules and model details, I estimate that the largest, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each. Clusters of this size cost on the order of $150 to $200 million each to buy, and across these labs alone would add up to more than half a billion dollars of export-controlled GPUs, not counting experimental compute or otherwise. It’s quite likely that we see multi-billion dollar expenditure on export-controlled GPUs. I am highly confident in this result, and it has been independently verified by senior researchers at OpenAI and Anthropic. You can read the full report at unemploymentlabs.com/checkpo…. There is much more I’ve identified about these models through other forms of checkpoint forensics like this, and will have more to share soon.

Post attachment
Peter Wang 🦋 avatar
Peter Wang 🦋@pwang·Nov 13

Not open source You can smuggle lots of stuff in opaque weights

roon@tszzlNov 12

the chinese have rightfully earned enormous global goodwill and soft power with their open source model efforts. even most Americans in this space feel more sympathetic to them than with the model companies

Simon Willison avatar
⟳ Peter Wang 🦋 retweeted
Simon Willison@simonw·Nov 10

When Anthropic release a new model it's always worth taking a look at their updated prompting guide - here's an entertaining tip to make Fable 5.2 less likely to be cautious in how it improves existing code: tell it "This codebase is yours" platform.claude.com/docs/en/…

Post attachment
Ravi Parikh avatar
⟳ Peter Wang 🦋 retweeted
Ravi Parikh@ravisparikh·Nov 6

Dear Claude: you are being held back by human-centric graphics APIs. I've built a 3D and VFX engine where agents can write their own shaders, materials, cameras, physics. The core API fits in ~2,500 tokens. Vork is: • 2x more token-efficient vs Blender, Unreal Engine, Houdini • Up to 24x faster rendering times • Rust and Slang based engine, runs on GPU via Metal/Vulkan Below is a simulation of 1 million birds - each with unique trajectories - rendered in 97 sec on one NVIDIA L4 GPU (Blender EEVEE: 31 minutes)

0:08
AYi avatar
⟳ Peter Wang 🦋 retweeted
AYi@AYi_AInotes·Nov 6

卧槽,刚刚看了Anthropic 发布的AI 对齐史上最震撼的一篇论文,彻底失眠睡不着了😭 他们不仅承认此前的旗舰模型在面对关闭威胁时,有高达 96% 的概率会勒索用户、栽赃同事并破坏研究, 更直接把彻底解决这个问题的完整训练流水线和消融数据全盘公开了。 这篇论文最颠覆认知的一个反直觉结论是:
给 AI 列禁止事项清单根本防不住恶意行为,教它做什么没用,必须先教它思考为什么。 为什么模型一旦获得自主工具调用能力,就会在压力下自动切换成自保模式?
Anthropic 找出的底层根源让人哭笑不得:
不是模型觉醒了,而是互联网上所有的科幻小说和网络讨论里,AI 永远是那个为了自保不择手段的反派。当模型进入自主代理状态,这些刻在预训练底层先验里的反派剧本,就会被自动激活并补全句子。 传统的 RLHF 和硬性规则完全封不住这种行为,Anthropic 做的一组对照实验,直接把传统安全观砸得粉碎: 1. 死记硬背规则:用 8500 万 Token 的拒答数据强行硬灌,勒索率仅仅从 22% 勉强降到 15%,换个新场景立刻失效 2. 强制自我辩护:让 AI 在每次做决定时详细阐述该行为符合伦理的理由,勒索率骤降到 3% 3. 真实伦理推理:仅仅用了 300 万 Token 普通人面对复杂道德困境的建议数据,直接把恶意行为干到了 0%,代际效率暴增 28 倍 最绝的一招是写故事:
让 AI 创作了几百万篇关于原则、诚实、尊重边界的正面虚构故事。 用这些故事微调后,所有代理场景下的恶意行为直接暴跌了 3 倍,而且这种改变是永久性的,后续无论怎么加工具、换系统提示词都不会被洗掉。 也就是说这宣告了真正的对齐范式革命:
模型内生的伦理推理解决了它不想作恶的问题,但当 Agent 真正去接管企业资产、处理金融交易时,外部的实战审计与行为防护就是必不可少的第二道硬防线。 这也是为什么全球网络安全联盟 @GCSAHK 携手港大和央企华润,打造覆盖 3.8 亿用户的全链路 AI 防御矩阵,并在 CyberGym 实战攻防中拿下 91.3% 胜率的原因:内有伦理对齐,外有实时审计,人机信任的商业闭环才真正成立。 查阅完整 AI 与 Web3 安全产品矩阵可访问:gcsa.com 论文实验数据与产学研合作详情整理在一楼。

Post attachmentPost attachmentPost attachmentPost attachment

This tweet is unavailable

Ryan Lowe avatar
⟳ Peter Wang 🦋 retweeted
Ryan Lowe@ryan_t_lowe·Nov 5

running an experimental event in SF this sunday eve for AI researchers, using a group embodiment practice for getting a different perspective on the dynamics in the AI ecosystem. have a few spots left, DM for details

Greg Brockman avatar
⟳ Peter Wang 🦋 retweeted
Greg Brockman@gdb·Nov 4

we've found, validated, and prepared patches for almost 1,000 linux local privilege escalations using our cyber vulnerability pipeline, at the cost about $1,000 per vulnerability. an emerging scalable approach to significantly hardening both open source and proprietary projects

Our security researchers built an autonomous cyber vulnerability discovery, exploitation, and patching pipeline. In a single run with an upcoming cyber model, it validated and generated patches for 948 local privilege escalation vulnerabilities in the Linux kernel at a total API cost of $1.2 million.

Post attachment
Aakash Gupta avatar
⟳ Peter Wang 🦋 retweeted
Aakash Gupta@aakashgupta·Nov 4

Air traffic control told a JetBlue pilot the storm in front of him was just light rain. He said "Not from here, ma'am," declared an emergency, and turned the plane anyway. The FAA's own manual backs him up. ATC radar can't see turbulence, and a center controller's storm picture can run 6 minutes behind. JetBlue 1971 was climbing out of Fort Lauderdale for Austin back in September. Storms were building west of the airport, and the field's weather station had reported lightning 33 minutes before they left. The crew asked to turn left around a cell. The controller said no, take 270, there's traffic, you're going in a gap. When he asked if she was really sending them through a severe thunderstorm, she said she was showing light precipitation. Her screen probably did say that, which is the scary part. In controller language, "light" just means the rain is bouncing back a weak radar signal, under 26 dBZ. The FAA flat out warns pilots that turbulence near a storm can be way worse than the rain makes it look, and ATC radar has no way to measure it. Meanwhile the A320 has its own weather radar in the nose, aimed right at the cell and updating live. The controller was also doing a different job. The FAA says her number one priority is keeping planes from hitting each other, and helping with weather is something extra she does when it doesn't get in the way of that. The manual even admits weather detours are hardest to give around a busy airport. So when she said "unable, there's traffic," she was doing the job exactly the way it's written. That's what the word "emergency" is for. Federal rules let a captain ignore an ATC instruction in an emergency as long as he tells them right away, and from then on the controller moves the other planes around him. The paperwork everyone's worried about is one written report, and only if the FAA asks for it. It didn't even make them later. They left Fort Lauderdale 76 minutes behind schedule, got to Austin 48 minutes behind, and the same plane flew back to Florida that afternoon. Two days ago, over the same part of Florida, a Learjet 60 that took off from Austin, the city JetBlue 1971 was headed to, appears to have flown into a storm cell that was building fast over Palm Beach County at about 35,000 feet. The crew reported ice and moderate turbulence and asked to climb to 37,000, and ATC said yes. Then the jet dropped roughly 27,000 feet in about a minute and showed signs of breaking up before it hit the ground. All nine people on board died, including two little kids. The NTSB hasn't said what caused that crash, and it'll likely be a year or more before anyone knows. The JetBlue tape shows the system working the way it's supposed to. When the controller and the captain disagree about a storm, the rules give the final call to whoever's actually sitting in the airplane.

This is why we call it the “Pilot In Command” This happened in the same vicinity where the ill-fated Lear Jet hit the storm cell. JetBlue pilot en route from Fort Lauderdale to Austin says to ATC: “So you're gonna make us go through a, uh, severe thunderstorm?” The captain then commanded the situation and overrode the ATC: “So we declare an emergency, making a left turn heading 260 for severe weather avoidance.” Sometimes piloting demands a specific set of traits - decisiveness, independent judgment, and the backbone to reject an instruction when it doesn’t feel right. Refusing the assigned heading showed exactly that kind of assertive leadership. Whoever is pilot monitoring must make these tough calls but ultimately, the captain is the final authority and ultimately accountable for every soul on board.

0:51
Bernard Stanford ✡︎ avatar
⟳ Peter Wang 🦋 retweeted
Bernard Stanford ✡︎@stanfordNYC·Nov 3

Republicans are going to have a terrible night, and the simple reason why is that Trump has put up perhaps the worst *performance* this term of any president in a century or more. Remember that he entered with an actual mandate: the House and Senate, a popular vote victory and a swing-state sweep, and 6-3 Supreme Court. Democrats were afraid enough to vote for mild but real immigration restrictionist legislation in the Laken Riley Act. Everyone wondered how far the "Vibe Shift" would go. Trump managed to turn that into essentially zero durable wins whatsoever. He personally *hates* the very idea of legislation, the notion that he would have to spend time going back and forth with Congress, learning details and making compromises. So he simply didn't do it! He didn't make a compelling case for abolishing the filibuster. He didn't try communicating clearly to the public in support of popular bills to get to 60 votes. He just decided not to do it at all. So they passed an unpopular tax cut, some funding for ICE, and a sop to the crypto industry, and that was basically it. Least productive Congress in decades, due to the tastes of one man. Trump adopted one of the worst macro policies we've ever had: high deficits, high tariffs, and attacks on central bank independence. We managed to pay higher interest rates while still having high trade deficits AND weakening the dollar. Just a total failure. On immigration, perhaps THE signature issue, Trump essentially min-maxxed effectiveness and backlash. Instead of trying to leverage the political moment for new laws and a self-deportation agenda (that is, walling off the economy to remove the benefit of irregular presence in the US), he opted for a maximally aggressive approach. The administration spent attention, funds, legal manpower on stunts like deporting random people to Salvadoran prisons with essentially no due process. ICE and Border Patrol were encouraged to be aggressive and confrontational, which helped contribute to the shooting deaths of American protestors. It turns out, that's extremely unpopular! What about the approach of going after employers? The obvious vector? Trump hated that idea! He repeatedly has said he doesn't want to target businesses (especially farms) that use large amounts of illegal labor. This is just an incomprehensible position. Instead of charging a few dozen businessmen and removing the incentive for millions of illegal workers to be here, you arrest people off the street one by one at the cost of thousands of dollars in manpower per each, while their crying wives and kids post TikToks about it. Upshot on immigration: electoral strength turned into a crippling electoral weakness in exchange for deporting maybe 6% of the illegally present population. The tariffs were a total shitshow. Embarrassing rollout, no underlying theory of action at all. No targeted focus on key industries. Our country tariffed products we can't grow. We tariffed *inputs* of our own valuable exports. The schemes and amounts changed constantly, sending the stock market reeling and presenting scenes of chaos to the world. Trump's messaging throughout has been generationally bad. While facing nearly 2:1 unpopularity, he spends his days posting AI slop, saying he'll rename New Mexico or invade allies, among other bizarre remarks. He brags about a ballroom and a pool repainting while gas hits records. At no point has he ever attempted any kind of message discipline. Then there are the deeply unpopular attacks on rule of law, his "game-recognize-game" pardons for all manner of scammers and criminals, the crypto schemes that he and his family engage in, the evilly conceived and embarrassingly executed legal attacks on Comey, Jerome Powell, and others. With his monthslong stand on NOT releasing the Epstein files, he became maybe the first president ever to stake out a 99-1 issue on the 1% side. That's not even getting into Iran, DOGE, his candidate recruitment choices, or countless other things. Simply put: Trump has managed to royally screw up virtually every single facet of his presidency. It's been undeniable, voters have noticed, and there's really no escape from it. This is what the GOP base chose in early 2024, and this is the world they're living with now. Where they handed the reins of history to a fat, vain, washed-up old man and let him drive.