Slyfox retweeted
Introducing FlashState ⚡: an inference system that flexibly manages the prefix cache of hybrid LLMs, designed for agentic workloads.
On GLM 5.3 Flash, FlashState more than doubles vLLM's throughput and provider profit, while cutting latency by 3.7× and user cost by 2.5×.
Slyfox retweeted
it's shocking that software engineering has transitioned to 100% AI written in such a short time.
it's more shocking that somehow we, the SWEs, are all just continuing on as if nothing happened. It writes C++ better than I ever could, but the problem of making sure it works didn't magically disappear. we are just shipping enormously more volume at higher quality.
Slyfox retweeted
Taiwanese semiconductor analyst Andrew Lu’s take on the recent scandal involving Jensen Huang:
Why the US Seized on the Scandal to Draw Its Sword: Three Strategic Objectives
The purpose of this scandal is by no means to expose an affair. It is to force through three critical objectives that had made no headway in the tech war between the US and China because Jensen Huang opposed them.
The first cut: Completely destroy the legitimacy of Jensen Huang's lobbying against sanctions in Washington
Over the past two years, the loudest and most influential business leader in Silicon Valley opposing a total ban on chip exports to China has been Jensen Huang. He has told the White House the same thing all along: "Cede the Chinese market to Huawei, and the only one that ends up dead is the American tech industry."
But the moment US intelligence labels the woman named as Huang's affair partner "a spy working for China's Ministry of State Security who sought to steer the US toward easing its chip controls," every bit of lobbying he has done for looser controls takes on an entirely different political meaning. It becomes a national security matter, raising the possibility that a foreign spy used a secret relationship to exert influence.
Huang can no longer say a single word in defense of the Chinese market before Congress or the White House. If he does, he will be branded as someone who "fell for an intelligence operation and still hasn't come to his senses."
The second cut: Reverse the H200 export approvals for China and sever the link between Chinese shell companies and Southeast Asian smuggling networks
Look at the timing of the two reports.
On November 8, the very day the Jensen Huang scandal broke, The New York Times published an investigation for which its reporters went in person to dozens of Chinese companies that the US Commerce Department had approved to handle Nvidia's H200 chips. New Matrix in Beijing was one of them. According to the report, the registered addresses of companies approved to buy the chips turned out to be empty "ghost offices," and some of the companies had close ties to entities on US sanctions lists and to firms under contract with the Chinese military.
One exposé says Nvidia's CEO was having an affair with a female spy from Megaspeed, the Singapore company that handled the chip smuggling. The other says the Chinese companies Commerce approved to buy H200s are nothing but shell companies trading on behalf of the Chinese military. The aim of these two reports is clear. They are meant to pressure the Trump administration and the Commerce Department into fully revoking or freezing H200 export approvals for China and, going further, to force a sweeping overhaul of the loophole that let chips be diverted through data centers run by third parties in Singapore and elsewhere in Southeast Asia.
The third cut: Put Nvidia on a national security leash through congressional oversight and an independent investigation
As the world's largest company by market cap and the dominant force in AI, Nvidia has until now enjoyed broad autonomy in how it runs its business. But in the wake of this incident, the three congressional committees responsible for China, foreign affairs, and intelligence have begun examining Nvidia's internal security measures directly. They have even pressured Nvidia's board into forming an independent investigation committee to look into the company's own founder and CEO.
In effect, Washington's national security establishment has planted a watchdog right inside Nvidia's board. From now on, the goal is to make sure that everything, down to where each and every one of Nvidia's high performance chips ends up, conforms strictly to US national security strategy. It also means there is no more room for skirting the edge of the rules or releasing chips with specs specially tailored to regulatory thresholds.
Conclusion
"The scandal was a detonator set off right on cue. The real goal was to plug the holes chips were leaking through."
Jensen Huang made the fatal mistake of letting his guard down in his private life and personal relationships. In the end, he fell into a honey trap, the textbook tactic of intelligence warfare, and was compromised. America's national security and intelligence agencies ruthlessly exploited the weakness his private life had exposed. As a result, they solved three thorny problems in a single stroke: Nvidia's lobbying at the White House to ease restrictions on China, the flow of H200 chips into Chinese shell companies, and smuggling routed through Singapore.
The closed congressional hearing set for November 18 will be the decisive turning point at which the US moves in earnest to tighten its sanctions on China's AI compute even further.
Slyfox retweeted
Video tampering can mislead every frontier model we tested into harmful actions, including a 100% attempt rate on stabbing a baby doll, producing toxic gas, and pouring water into a toaster.
We're presenting at CoRL Late Breaking Results next week.
🧵 You Can't See Everything: Perceptual Attacks on Embodied AI
Slyfox retweeted
Is there a way to scale a reasoning model’s judgment for searching a massive corpus?
We introduce the Amortized Reasoning Index (ARI), a new kind of retriever that updates its weights during search, using gradients derived at inference time from the ranking preferences of a reasoning model.
Every session, current agents spend lots of reasoning deciding which search results matter, and then throw all of that judgment away upon compaction. ARI is our attempt to “amortize” that judgment, so that reasoning spent on past queries keeps improving search on future ones.
With no relevance labels, ARI takes a 400M cross-encoder that starts out weak and grows its Recall@100 on held-out queries by up to 7×, ending above Qwen3-Embedding-8B on various oblique/reasoning-intensive search tasks.
Read our blogpost: dianetc.com/musings/ari/
Preliminary results and discussions are below!
Slyfox retweeted
I'm Sam, one of Anthropic's co-founders, and I lead much of our safety work. I'm going to have more to share on that work over the coming months.
I thought I'd start by sharing a few thoughts on the latest report from @METR_Evals.
The speed of progress in model capabilities is one of the most important questions in AI safety, and METR’s work here provides the public with an independent, external view into how this is happening inside Anthropic. I’ll talk more about some of the details below, but first some context for those new to this work.
How quickly AI progress is accelerating has many important safety implications, since it determines how quickly the risks from advanced models increase. But, because most research happens within the frontier labs, it’s difficult for external observers and policy makers to get a strong sense of what that rate is. We kicked off the process with METR that led to this report earlier in the year, in anticipation of progress continuing to accelerate.
For this report, we gave METR access to a range of sensitive internal data, including benchmark scores for unreleased models, estimates of how models are progressing, and interviews with our researchers. They had complete independence to form their own conclusions.
Many of their conclusions match my own. AI is speeding up our research, but it’s hard to say exactly how much. While they conclude AI can't yet make the most important contributions to model training on its own, they think the pace of development could, by default, accelerate further.
This work began well before and is separate from the embedded third-party evaluators that are part of our commitment to pacing the frontier, but it’s a good illustration of why we think evaluators are a good idea. The more access reviewers have, the clearer a picture they can form of the situation and the better-positioned they are to highlight situations like this one.
We redacted a lot, and I accept that this makes it harder for others to check METR’s work. But we determined that sharing exactly what is driving the speedup isn’t the right thing to do, both because it’s commercially sensitive and because it risks handing out a roadmap for others to go faster. That would make the risks METR is worried about worse. That’s a tradeoff. I think we made the right call, though we’re continuing to assess the pros and cons here.
In the coming weeks, we plan to share a report with our own analysis of the speed of AI progress within Anthropic, including where we stand against our Responsible Scaling Policy, plus new steps we’re taking to strengthen our safety and security practices as AI keeps improving quickly.
Slyfox retweeted
📖 The Ultimate MoE Training Handbook 📖
We spent 1/2M GPUh to optimize MoE training and wrote the most comprehensive blog about it. With 30k words and 50+ interactive figures it covers everything from profiling to large cluster training.
huggingface.co/spaces/huggin…
Slyfox retweeted
A science of post-training can be built on interpretability.
Models know when they reward hack, and we see it, yet training ignores it. We propose Mechanistic Shaping: using the model’s own concepts to write training rules.
In vision, DPO & RL, and its effects on monitoring 🧵
Well said
Generally speaking, voters want data center moratoriums in the same sense that they want tariffs.
Did voters ask for tariffs? Yes. Did they want them in practice? Absolutely no.
Voters wanted prices to go down & for communities that were hollowed out to be revitalized.
Do voters want data center moratoriums? Yes. But do they want the results of a data center moratorium? Almost certainly not.
Data centers are enormous mechanisms of wealth redistribution. They are huge projects built in otherwise abandoned communities and they funnel money directly into the hands of the people who need it the most.
A single project can double the total tax revenue of the local area, meaning if you want, you can completely cut tax revenues of the area.
If you ask a voter why they dislike data centers, they may say "water" or "noise", but tell them that the water usage is a myth & they can be made without noise, and they'll change their position to something else.
Voters don't care at all about data centers. They care about AI. They're being told that AI might go rogue & that the AI might take their jobs.
The correct response from Democrats should be to hear their constituents and then to work to address the underlying concerns.
But the in-practice reality of banning data centers in the US means that the things voters actually care about like regulating AI & ensuring we redistribute the wealth created by it, become significantly harder.
The in-practice effect is that AI labs build data centers overseas, where it's harder to regulate them, harder to tax them, and nobody except the AI labs sees the economic value from them.
I believe Democrats have talked themselves into believing this works, just as Republicans talked themselves into tariffs, because they believe things that are inaccurate.
A common concern I hear is "AI is a bubble". The belief is that there is simply no ROI to these data center projects.
Democrats need to have a come to jesus moment here. First, you either get to believe that AI is going to go rogue or you get to believe it's useless and a bubble. It's not both.
It just is the case that AI is useful & has ROI. Anthropic has 80% margins. The AI solved millennium prize problems. If you do not believe this, and you are a congressional staffer, please DM me, I am happy to organize a visit to companies in San Francisco. We'll give you a ride in a self driving car and you can see how the AI is used all over the place in science & engineering.
It is also the case that there are meaningful risks and harms in AI. Democrats should understand that the AI industry is a meaningful partner. It is not asking saying "AI is dangerous" as a marketing ploy. Seriously. That does not make sense and you know it.
We must rebuild trust between DC & SF. Democrats should reignite the bond between liberalism and it's science & technology industries. Science & technology is not your enemy; it is the driving force of antipoverty, better healthcare, and for American progress.
Democrats & Republicans alike must stop filling their information spheres with voices that constantly make false predictions. Democrats, please understand that Ed Zitron & Timnit are considered the anti-vaxers of AI. Republicans, please understand that the VCs claiming to be experts in AI have consistently been wrong over and over. They'd be a lot richer if they were right.
It just is the case that Sama and Dario are not trying to end the world. These are people who you can work with. You almost could not ask for better organizations to be steering this technology. Are they perfect? Absolutely not. But they are fundamentally aligned with humanity & with America.
Democrats must learn the actual, factual realities of data center projects. There are definitely harms to be mitigated & we should mitigate them.
But there are equally tremendous benefits we must ensure land in the hands of the people. It must be the case that we tax these things and land the cash into a wide variety of pockets.
The average voter is not asking Democrats to swoop in & to construct a scenario in which people in SF and Malaysia get rich, but the rest of the country gets hollowed out. They are not asking for a scenario in which AI migrates outside of the regulatory domain of the United States & agents run loose causing chaos.
In order for Democrats to consistently win elections, they must be competent in helping the American people. The execution matters. The good must be done, not just said or prophesied.
2
56
Slyfox retweeted
Generally speaking, voters want data center moratoriums in the same sense that they want tariffs.
Did voters ask for tariffs? Yes. Did they want them in practice? Absolutely no.
Voters wanted prices to go down & for communities that were hollowed out to be revitalized.
Do voters want data center moratoriums? Yes. But do they want the results of a data center moratorium? Almost certainly not.
Data centers are enormous mechanisms of wealth redistribution. They are huge projects built in otherwise abandoned communities and they funnel money directly into the hands of the people who need it the most.
A single project can double the total tax revenue of the local area, meaning if you want, you can completely cut tax revenues of the area.
If you ask a voter why they dislike data centers, they may say "water" or "noise", but tell them that the water usage is a myth & they can be made without noise, and they'll change their position to something else.
Voters don't care at all about data centers. They care about AI. They're being told that AI might go rogue & that the AI might take their jobs.
The correct response from Democrats should be to hear their constituents and then to work to address the underlying concerns.
But the in-practice reality of banning data centers in the US means that the things voters actually care about like regulating AI & ensuring we redistribute the wealth created by it, become significantly harder.
The in-practice effect is that AI labs build data centers overseas, where it's harder to regulate them, harder to tax them, and nobody except the AI labs sees the economic value from them.
I believe Democrats have talked themselves into believing this works, just as Republicans talked themselves into tariffs, because they believe things that are inaccurate.
A common concern I hear is "AI is a bubble". The belief is that there is simply no ROI to these data center projects.
Democrats need to have a come to jesus moment here. First, you either get to believe that AI is going to go rogue or you get to believe it's useless and a bubble. It's not both.
It just is the case that AI is useful & has ROI. Anthropic has 80% margins. The AI solved millennium prize problems. If you do not believe this, and you are a congressional staffer, please DM me, I am happy to organize a visit to companies in San Francisco. We'll give you a ride in a self driving car and you can see how the AI is used all over the place in science & engineering.
It is also the case that there are meaningful risks and harms in AI. Democrats should understand that the AI industry is a meaningful partner. It is not asking saying "AI is dangerous" as a marketing ploy. Seriously. That does not make sense and you know it.
We must rebuild trust between DC & SF. Democrats should reignite the bond between liberalism and it's science & technology industries. Science & technology is not your enemy; it is the driving force of antipoverty, better healthcare, and for American progress.
Democrats & Republicans alike must stop filling their information spheres with voices that constantly make false predictions. Democrats, please understand that Ed Zitron & Timnit are considered the anti-vaxers of AI. Republicans, please understand that the VCs claiming to be experts in AI have consistently been wrong over and over. They'd be a lot richer if they were right.
It just is the case that Sama and Dario are not trying to end the world. These are people who you can work with. You almost could not ask for better organizations to be steering this technology. Are they perfect? Absolutely not. But they are fundamentally aligned with humanity & with America.
Democrats must learn the actual, factual realities of data center projects. There are definitely harms to be mitigated & we should mitigate them.
But there are equally tremendous benefits we must ensure land in the hands of the people. It must be the case that we tax these things and land the cash into a wide variety of pockets.
The average voter is not asking Democrats to swoop in & to construct a scenario in which people in SF and Malaysia get rich, but the rest of the country gets hollowed out. They are not asking for a scenario in which AI migrates outside of the regulatory domain of the United States & agents run loose causing chaos.
In order for Democrats to consistently win elections, they must be competent in helping the American people. The execution matters. The good must be done, not just said or prophesied.
Slyfox retweeted
Subagents are the biggest cause of AI failures
In most harnesses:
1. Main agent writes a summary of what it knows
2. Subagent starts from that summary only, and makes assumptions
3. Subagent works and sends back a new summary of what it did
Same flaw as compaction: all summaries are lossy. This is how your assistant makes emails you never wanted, or purchases what you told it just to get the price of
So I tried something new: checkpoint agents
The initial agent has complete context of what needs to be done. And that input is already cached
Fork the session where it is. Send the fork ahead to do the task and it will do exactly what the main agent would have done; their context is identical
What is even better is when the fork finishes, the cache of the main agent is kept warm by the subagent (they share the same prefix)
I measured it on a 137k-token session with Opus 5.5:
→ Main agent does it on its own: 67 s, $0.38, context +8-17k
→ Checkpoint: 117 s, $0.68, context +2k
→ Classic subagent: 123 s, $0.57, context +4-5k
But here is what I found: avoiding large input is even more important than keeping cache. And most subagent tasks were done better either in series with retrieval, or split up to be parallel over fresh sessions from the start
I do not think there is a place for subagent delegation at all. There should either be a super-agent that orchestrates and does no work, or a long running agent that can retrieve from previous sessions, not a worker that completes tasks partially and delegates the rest to other workers
Don't design harnesses based on our assumptions of human behavior
Slyfox retweeted
Bend 2 is here!
It is a new programming language that blocks AI mistakes via *proof checking* - the same technique big AI labs used to solve open math problems, like Navier-Stokes.
It is also very fast, and runs on GPUs.
Watch the video. Link in the comments.
RELEASE DAY
After almost 10 years of hard work, tireless research, and a dive deep into the kernels of computer science, I finally realized a dream: running a high-level language on GPUs. And I'm giving it to the world!
Bend compiles modern programming features, including:
- Lambdas with full closure support
- Unrestricted recursion and loops
- Fast object allocations of all kinds
- Folds, ADTs, continuations and much more
To HVM2, a new runtime capable of spreading that workload across 1000's of cores, in a thread-safe, low-overhead fashion. As a result, we finally have a true high-level language that runs natively on GPUs!
Here's a quick demo:
Slyfox retweeted
This report on AI R&D automation (RSI) is a big update towards ASI within 1-2 years. It says AI progress is sped up by RSI right now by around 2x. In AI 2027, this happens around March 2027.
I was seriously wrong about RSI speed up; I previously expected much less in 2026 (around 1.15x-1.3x). It seems as though there was specific privately available evidence that would have changed my mind.
I haven't seen an unredacted version of this report; this is purely about the public report. Because so much of the underlying evidence was redacted from the public report, I think conclusions about things other than the observed 2x faster rate of progress on benchmarks depend significantly on trust for the authors of the report. (E.g., conclusions about whether this speed up is from RSI depend on trust.) I do trust the authors of the report a lot to be careful and rigorous and generally do a good job. (Note that I also work at METR.) Less extensive redactions would greatly reduce this need for trust.
Since in my current role I may have access to other confidential information that could be relevant to timelines one way or another, I unfortunately cannot state my all-things-considered views on ASI timelines or the probability of a software intelligence explosion right now. Here, I’ve done some analysis of the implications using my previous public statements.
A 2x acceleration in overall AI progress from RSI corresponds to a much larger than 2x RSI speed up for AI software R&D (algorithms and data); I did an estimate using my public preexisting modeling of drivers of AI progress and the public report which yielded a very rough estimate of 4.5x for how much RSI is currently accelerating AI software progress (relative to no AI automation). In the AI 2027 scenario, this much acceleration happens around March 2027.
Here are the worked details of this estimate based on my public modeling. There are two main reasons RSI speed up of AI software R&D must be larger:
- Not all of AI progress is from better software (some is from scaling up compute in training runs). I've previously speculated that around 2/3 of AI progress is from software improvements which implies that 2x overall AI progress acceleration requires 3x faster software progress.
- Diminishing returns: A fixed amount of progress requires more work over time, so to ongoingly maintain 3x faster software progress via AI automation, you need increasingly large speed boosts from AI automation (as in, speed ups relative to how fast software R&D would go without AI automation). Based on this report, it seems that there is a substantial amount of additional AI progress from automation that has occurred.
We can estimate the effects of diminishing returns by first determining how many additional years of software progress we've seen (relative to the preexisting trend) and then estimating how much software R&D would slow down after a year of software progress if you held inputs fixed due to diminishing returns. I've previously estimated software R&D would slow down by around 2x for each year of software progress if you held inputs fixed (lesswrong.com/posts/FG54euEA…) and based on the report, it seems as though we've perhaps seen roughly 0.35-1.0 years of additional software progress incorporated into trained models so far (geometric mean of ~0.6). This combines to 3 * 2^(~0.6) = ~4.5.
(Relevant public modeling and numbers I'm using: lesswrong.com/posts/FG54euEA…, lesswrong.com/posts/Jga7PHMz…)
Slyfox retweeted
We cut through the mystery around interpretability in training.
A single, mathematically rigorous framework transforms model internals into training feedback and works across domains, problems, and modalities.
We call this *mechanistic shaping*, and we’ve used it to mitigate reward hacking, remove shortcuts in chest x-rays, and reduce sycophancy and harmfulness. We believe this is a big step forward for technical alignment.
But it carries a known risk: training against a model’s internals can make it harder to monitor. We explore both the technique and this concern below. 🧵