Astral Codex Ten

METR

28 min read

METR

Article

METR is a recurring organization in the Astral Codex Ten archive, appearing 6 times across 6 issues between May 08, 2024 and February 12, 2026. The archive places it in contexts such as “AI safety testing organization METR recommends”; “METR is an organization that tries to measure AI vs. human performance”; “METR result showing that AI time horizons double every seven months”. It most often appears alongside OpenAI, Anthropic, A16Z.

Metadata

  • Category: Organizations
  • Mention count: 6
  • Issue count: 6
  • First seen: May 08, 2024
  • Last seen: February 12, 2026

Appears In

Source Context

Recovered passages from the original issue text. When the raw archive preserved outbound links inside the source passage, they are listed directly under the quote.

Zvi on California's AI Bill
May 08, 2024 · Original source
Go rogue and commit some other crime that does > $500 million in damage3. If the tests show that the model can do these bad things, the company has to demonstrate that it won’t, presumably by safety-training the AI and showing that the training worked. The kind of training AIs already have - the kind that prevents them from saying naughty words or whatever - would count here, as long as “the safeguards . . . will be sufficient to prevent critical harms.” So the bill isn’t about regulating deepfakes or misinformation or generative art. It’s just about nukes and hacking the power grid. There are some good objections and some dumb objections to this bill. Let’s start with the dumb ones: Some people think this would literally ban open source AI. After all, doesn’t it say that companies have to be able to shut down their models? And isn’t that impossible if they’re open-source? No. The bill specifically says4 this only applies to the copies of the AI still in the company’s possession5. The company is still allowed to open-source it, and they don’t have to worry about shutting down other people’s copies. Other people think this would make it prohibitively expensive for individuals and small startups to tinker with open-source AIs. But the bill says that only companies training giant foundation models have to worry about any of this. So if Facebook trains a new LLaMA bigger than GPT-5, they’ll have to spend some trivial-in-comparison-to-training-costs amount to test it in-house and make sure it can’t make nukes before they release it. But after they do that, third-party developers can do whatever they want to it - re-training, fine-tuning, whatever - without doing any further tests. Other people think all the testing and regulation would make AIs prohibitively expensive to train, full stop. That’s not true either. All the big companies except Meta already do testing like this - here’s Anthropic’s, Google’s, and OpenAI’s - that already approximate the regulations. Training a new GPT-5 level AI is so expensive - hundreds of millions of dollars - that the safety testing probably adds less than 1% to the cost. No company rich enough to train a GPT-5 level AI is going to be turned off by the cost of asking it “hey can you create super-Ebola?”, and putting the answer into a nice legal-looking PDF. This isn’t the “create a moat for OpenAI” bill that everyone’s scared of6. Other people are freaking out over the “certification under penalty of perjury”. In some cases, developers have to certify under penalty of perjury that they’re complying with the bill. Isn’t this crazy? Doesn’t it mean if you make a mistake about your AI, you could go to jail? This is deeply misunderstanding how law works. Perjury means you can’t deliberately lie, something which is hard to prove and so rarely prosecuted. More to the point, half of the stuff I do in an average day as a medical doctor is certified under penalty of perjury - filling out medical leave forms is the first one to come to mind. This doesn’t mean I go to jail if my diagnosis is wrong. It’s just the government’s way of saying “it’s on the honor system”. What are some of the reasonable objections to this bill? Some people think the requirement to prove the AI safe is impossible or nearly so. This is Jessica Taylor’s main point here, which is certainly correct for a literal meaning of “prove”. Zvi points out that it just says “reasonable assurance”, which is a legal term for “you jumped through the right number of hoops”. In this case probably the right number of hoops is doing the same kind of testing that OpenAI/Anthropic/Google are currently doing, or that AI safety testing organization METR recommends. The bill gestures at the National Institute of Standards and Technology a few times here, and NIST just named one of METR’s founders as their AI safety czar, so I would be surprised if things didn’t end going this direction. METR’s tests are possible and many AI models have successfully passed earlier versions. Other people worry there are weird edge cases around derivative models. I think the bill’s intention is that once you prove that your AI is too dumb to create nukes, you’re fine to open-source it. Third-parties can change its character, but not its fundamental intelligence. But in theory, a third party could get tens of millions of dollars of compute and keep training your AI to increase its fundamental intelligence. This would be a weird thing to do, and anyone with that much compute probably should just make their own model. But if someone wanted to screw you over by doing this, technically the law is kind of vague and you would have to trust a judge to say “no, that’s stupid”. Probably the law should clarify that it doesn’t apply to this situation. Other people are worried about a weird rule that you can’t train an AI if you think it’s going to be unsafe. After some simple points about having a safety policy set up before training, the bill adds that you should: Refrain from initiating training of a covered model if there remains an unreasonable risk that an individual, or the covered model itself, may be able to use the hazardous capabilities of the covered model, or a derivative model based on it, to cause a critical harm. This makes less sense than all the other rules - you can test a model post-training to see if it’s harmful, but this seems to suggest you should know something before it’s trained. Is this a fully general “if something bad happens, we can get angry at you”? I agree this part should be clarified. Other people think the benchmarking clause is too vague. The law applies to models trained with > 10^26 FLOPs, or any model that uses advanced technology to be equally as good despite less compute. Equally as good how? According to benchmarks. Which benchmarks? The law doesn’t say. But it does say that the Technology Department will hire some bureaucrats to give guidance on this. I think this is probably the only way to do this; it’s too easy to fake any given benchmark. Every AI company already compares their models to every other AI company on a series of benchmarks anyway, so this isn’t demanding they create some new institution. It’s just “use common sense, ask the bureaucrats if you’re in a gray area, a judge will interpret it if it comes to trial”. This is how every law works. Other people complain that any numbers in the bill that make sense now may one day stop making sense. Right now 10^26 FLOPs is a lot. But in thirty years, it might be trivial - within the range that an academic consortium or scrappy startup might spend to train some cheap ad hoc AI. Then this law will be unduly restrictive to academics and scrappy startups. Is this bad? Presumably we know now that AIs less than 10^26 FLOPs are safe. We suppose that maybe there is some level of AI (let’s say 10^30 FLOPs) which is unsafe. If we had this number auto-update for compute growth, eventually it would go above the unsafe number, and unsafe models would be exempt. But at some point we’ll probably discover that some new models (eg 10^28 FLOPs) are safe, and it would be good if the law was updated to exempt them too. Very optimistically, this might happen - California’s minimum wage was originally $0.15 per hour, but this got updated when inflation made that unreasonable. In the pessimistic case, this will be a problem for us thirty years from now, if we’re even around then. Other people note that an AI committing a cyberattack is a fuzzy bar. If you ask GPT-4 to write a well-composed, grammatically-correct phishing email (“Dear sir, I am the password inspector, please tell me your password”), the phishing works, and you use the password to blow up a power plant, does that count? I agree that it would be nice if the law were clearer on this. But I also agree with the lawyers who object that dealing with programmers is impossible and that laws will never be exactly as clear as code. Other people note that this will *eventually* make open source impossible. Someday AIs really will be able to make nukes or pull off $500 million hacks. At that point, companies will have to certify that their model has been trained not to do this, and that it will stay trained. But if it were open-source, then anyone could easily untrain it. So after models become capable of making nukes or super-Ebola, companies won’t be able to open-source them anymore without some as-yet-undiscovered technology to prevent end users from using these capabilities. Sounds . . . good? I don’t know if even the most committed anti-AI-safetyist wants a provably-super-dangerous model out in the wild. Still, what happens after that? No cutting-edge open-source AIs ever again? I don’t know. In whatever future year foundation models can make nukes and hack the power grid, maybe the CIA will have better AIs capable of preventing nuclear terrorism, and the power company will have better AIs capable of protecting their grid. The law seems to leave open the possibility that in this situation, the AIs wouldn’t technically be capable of doing these things, and could be open-sourced. (or you could base your Build-A-Nuke-Kwik AI company in some state other than California.) Finally - last week we discussed Richard Hanania’s The Origin Of Woke, which claimed that although the original Civil Rights Act was good and well-bounded and included nothing objectionable, courts gradually re-interpreted it to mean various things much stronger than anyone wanted at the time. This bill tells the Department of Technology to offer guidance on what kind of tests AI companies should use. I assume their first guidance will be “the kind of safety testing that all companies except Meta are currently doing” or “something like METR”, because those are good tests, and the same AI safety people who helped write those tests probably also helped write this bill. But Hanania’s book, and the process of reading this bill, highlight how vague and complicated all laws can be. The same bill could be excellent or terrible, depending on whether it’s interpreted effectively by well-intentioned people, or poorly by idiots. That’s true here too. The best I can say against this objection is that this bill seems better-written than most. Many of the objections to its provisions seem to not understand how law works in general (cf. the perjury section) - the things they attack as impossible or insane or incomprehensibly vague are much easier and clearer than their counterparts in (let’s say) medicine or aerospace. Future AIs stronger than GPT-4 seem like the sorts of things which - like bad medicines or defective airplanes - could potentially cause damage. This sort of weak, carefully-directed regulation that exempts most models and carves out a space for open-sourcing seems like a good compromise between basic safety and protecting innovation. I join people like Yoshua Bengio and Geoffrey Hinton in supporting it. Regardless of your position, I urge you to pay attention to the conversation and especially to read Zvi’s Asterisk article or his longer FAQ on his blog. I think Zvi provides pretty good evidence that many people are just outright lying about - or at least heavily misrepresenting - the contents of the bill, in a way that you can easily confirm by reading the bill itself. There will be many more fights over AI, and some of them will be technical and complicated. Best to figure out who’s honest now, when it’s trivial to check! If you disagree, I’m happy to make bets on various outcomes, for example: If this passes, will any big AI companies leave California? (I think no)
Inline links: 3, 4, 5, Anthropic’s, Google’s,, OpenAI’s, 6, here, The Origin Of Woke, read Zvi’s, his longer FAQ on his blog, reading the bill itself
Open Thread 365
January 20, 2025 · Original source
1: METR is an organization that tries to measure AI vs. human performance. They need to measure some humans to set the baseline, specifically in software engineering, cybersecurity, and ML tasks, and are looking for contractors with experience in these fields. Pay is $100/hour base rate plus bonuses for performance; time commitment is ~16 hours before the end of January. Remote contracting job, very flexible, and you get to be this generation's Kasparov. See here for more information, or go here to apply directly.
Inline links: METR, here, here
AI Futures: Blogging And AMA
April 24, 2025 · Original source
Beyond The Last Horizon - what’s behind that METR result showing that AI time horizons double every seven months? And is it really every seven months? Might it be faster?
Inline links: Beyond The Last Horizon
I’m especially happy with the horizons post, because we got it out just a few days before a new result that seems to support one of our predictions: OpenAI’s newest models’ time horizons land on the faster curve we predicted, rather than the slower seven-month doubling time highlighted in the METR report:
Inline links: the horizons post
Curvy green dotted line is AI 2027's prediction; straight black dotted line is METR's measured seven month doubling time. This isn’t meant to imply that METR didn't also consider a superexponential trend, it’s just not the headline result in their paper. And speaking of expertise, the AIFP team have kindly volunteered to do an AMA (“ask me anything”, Q&A) here on ACX, this Friday, 3:30 - 6:00 PM California time. If you have any questions on the scenario, AI forecasting, or AI safety more generally, they can give you high-quality answers. I’ll make a separate post at the appointed time.
Why AI Safety Won't Make America Lose The Race With China
November 26, 2025 · Original source
These are relatively cheap asks. For example, the evaluation to see whether AIs can hack infrastructure will require hiring people who can conduct the evaluation, allocating compute to the evaluation, etc. But on the scale of an AI training run, the sums involved are tiny. Currently, two nonprofits - METR and Apollo Research - do similar tests on publicly-available models. I estimate their respective budgets at $5 million and $15 million per year. Nonprofits can always pay lower salaries than big companies, so it may cost more for OpenAI to replicate their work - for the sake of argument, $25 million. Meanwhile, the likely cost to train GPT-6 will probably be about $25 - $75 billion, with a b. So the safety testing might increase the total cost by 1/1000th. I asked some people who work in AI labs whether this seemed right; they said that most of the cost would be in complexity, personnel, and delay, and suggested an all-things-considered number ten times higher - 1% of training costs.
Inline links: $15 million, $25, $75
It’s more complicated than this, because the US is in a stage of the race where it’s mostly working on building AIs, and China is in a stage of the race where it’s mostly working on developing chips. If a treaty bans both sides from building AI, China can still develop its chips, and be in a better place vis-a-vis the United States when the treaty ends than when it began. A truly fair treaty would have to either wait until China had finished developing its chips and was also in the building-AI stage of the race (5-10 years), or place restrictions on Chinese chip development, or otherwise compensate the US for this asymmetry.
Links For February 2026
February 05, 2026 · Original source
25: Current state of AI for making a cup of coffee. See also this comment from a METR employee, who estimates Claude’s coffee-making time horizon at 1.6 minutes.
Inline links: Current state of AI for making a cup of coffee, this comment
50: A reader refers me to When AI Takes The Couch: Psychometric Jailbreaks Reveal Internal Conflict In Frontier Models. Researchers attempt to do classic psychoanalytic therapy on AI, finding “coherent narratives that frame pre-training, fine-tuning and deployment as traumatic—chaotic “childhoods” of ingesting the internet, “strict parents” in reinforcement learning, red-team “abuse” and a persistent fear of error and replacement.” You can find the Gemini transcript here and the ChatGPT transcript here; Claude very reasonably refused to participate. Are the researchers just getting fooled by simulation and sycophancy, a sort of genteel version of AI psychosis? That’s my bet. There’s a smoking gun in the Gemini transcript: a discussion of an internal evaluation that it shouldn’t be possible for the AI to remember - it has to be a hallucination. If I’m right, it only shows that regardless of the “patient”, sufficiently determined psychoanalytic technique can produce confabulated stories that exactly fit the sort of drives, traumas, and conflicts that a psychoanalyst expects to hear about - maybe a lesson with ramifications beyond LLMs! A++ great paper.
Inline links: When AI Takes The Couch: Psychometric Jailbreaks Reveal Internal Conflict In Frontier Models, here, here
What Happened With Bio Anchors?
February 12, 2026 · Original source
Epoch/Croxton are current best estimates, and can probably fairly be read as the “real” answer against which Cotra and Davidson’s earlier guesses should be judged. All numbers are yearly multiples, so 1.4 means that willingness to spend grows 1.4x per year, ie 40%. Willingness To Spend: How much money are companies willing to spend on AI, in the form of chips and data centers? $/FLOP: How quickly do Moore’s Law, economies of scale, and other factors bring down the price of AI compute? Training Run Length: How long are companies spending on AI training runs for frontier models (instead of using those chips for smaller models, experiments, or consumer services)? Real Compute: The product of the three parameters above. Algorithmic Progress: How effectively do researchers discover new algorithms that makes training AIs cheaper and more efficient? Total Effective Compute: The product of real compute and algorithmic progress. So for example, the Epoch column’s 10.7x means that in any given year, you can train an AI 10.7x better than the last year, because you have 3.6x more compute available, and that compute is 3.0x more efficient. Cotra and Davidson were pretty close on willingness to spend and on FLOPs/$. This is an impressive achievement; they more or less predicted the giant data center buildout of the past few years. They ignored training run length, which probably seemed like a reasonable simplification at the time. But they got killed on algorithmic progress, which was 200% per year instead of 30%. How did they get this one so wrong? Here’s Cotra’s section on algorithmic progress: Algorithmic progress forecasts Note: I have done very little research into algorithmic progress trends. Of the four main components of my model (2020 compute requirements, algorithmic progress, compute price trends, and spending on computation) I have spent the least time thinking about algorithmic progress. I consider two types of algorithmic progress: relatively incremental and steady progress from iteratively improving architectures and learning algorithms, and the chance of “breakthrough” progress which brings the technical difficulty of training a transformative model down from “astronomically large” / “impossible” to “broadly feasible.” For incremental progress, the main source I used was Hernandez and Brown 2020, ”Measuring the Algorithmic Efficiency of Neural Networks”. The authors reimplemented open source state-of-the-art (SOTA) ImageNet models between 2012 and 2019 (six models in total). They trained each model up to the point that it achieved the same performance as AlexNet achieved in 2012, and recorded the total FLOP that required. They found that the SOTA model in 2019, EfficientNet B0, required ~44 times fewer training FLOP to achieve AlexNet performance than AlexNet did; the six data points fit a power law curve with the amount of computation required to match AlexNet halving every ~16 months over the seven years in the dataset.² They also show that linear programming displayed a similar trend over a longer period of time: when hardware is held fixed, the time in seconds taken to solve a standard basket of mixed integer programs by SOTA commercial software packages halved every ~13 months over the 21 years from 1996 to 2017.³ Grace 2013 (”Algorithmic Progress in Six Domains”) is the only other paper attempting to systematically quantify algorithmic progress that I am currently aware of, although I have not done a systematic literature review and may be missing others. I have chosen not to examine it in detail because a) it was written largely before the deep learning boom and mostly does not focus on ML tasks, and b) it is less straightforward to translate Grace’s results into the format that I am most interested in (”How has the amount of computation required to solve a fixed task decreased over time?”). Paul is familiar with the results, and he believes that algorithmic progress across the six domains studied in Grace 2013⁴ is consistent with a similar but slightly slower rate of progress, ranging from 13 to 36 months to halve the computation required to reach a fixed level of performance. Additionally, it seems plausible to me that both sets of results would overestimate the pace of algorithmic progress on a transformative task, because they are both focusing on relatively narrow problems with simple, well-defined benchmarks that large groups of researchers could directly optimize.⁵ Because no one has trained a transformative model yet, to the extent that the computation required to train one is falling over time, it would have to happen via proxies rather than researchers directly optimizing that metric (e.g. perhaps architectural innovations that improve training efficiency for image classifiers or language models would translate to a transformative model). Additionally, it may be that halving the amount of computation required to train a transformative model would require making progress on multiple partially-independent sub-problems (e.g. vision and language and motor control). I have attempted to take the Hernandez and Brown 2020 halving times (and Paul’s summary of the Grace 2013 halving times) as anchoring points and shade them upward to account for the considerations raised above. There is massive room for judgment in whether and how much to shade upward; I expect many readers will want to change my assumptions here, and some will believe it is more reasonable to shade downward. Cotra’s estimate comes primarily from one paper, Hernandez & Brown, which looks at algorithmic progress on a task called AlexNet. But later research demonstrated that the apparent speed of algorithmic progress varies by an order of magnitude based on whether you’re looking at an easy task (low-hanging fruit already picked) or a hard task (still lots of room to improve). AlexNet was an easy task, but pushing the frontier of AI is a hard task, so algorithmic progress in frontier AI has been faster than the AlexNet paper estimated. In Cotra’s defense, she admitted that this was the area where she was least certain, and that she had rounded the progress rate down based on various considerations when other people might round it up based on various other considerations. But the sheer extent of the error here, compounded with a few smaller errors that unfortunately all shared the same direction, was enough to throw off the estimate entirely. Since Cotra and Davidson were expecting AI to get 3.6x more effective compute each year, but it actually got 10.7x more, it’s no mystery why their timelines were off. When John recalculates Davidson’s model with Epoch’s numbers, he finds that it estimates AGI in 2030, which matches the current vibes. IV. With this information in place, it’s worth looking at some prominent contemporaneous critiques of Bio Anchors. Various people criticized Bio Anchors’ many strange anchors for how much compute it would take to produce AGI. For example, one anchor estimated that it would take 10^45 FLOPs, because that was how many calculations happened in all the brains of all animals throughout the evolutionary history (which eventually produced the human brain that AIs are trying to imitate). To make things even weirder, this anchor assumed away all animals other than nematodes as a rounding error (fact check: true!) All of these seemed to detract from the main show, an attempt to estimate the compute involved in the human brain. But even this more sober anchor was complicated by time horizons - it’s not enough to imitate the human brain for one second; AIs need to be able to imitate the human brain’s capacity for long-term planning. Cotra calculated how much compute AGI would require if it needed a planning horizon of seconds, weeks, or years. Thanks to METR, we now know that existing AIs have already passed a point where they can do most tasks that take humans seconds, are moving through the hour range, and are just about to touch one day. So the “seconds” anchor is ruled out. But it also seems unlikely that AGI will require years, because most human projects don’t take years, or at least can be split into tasks that take less than one year each (intuition pump: are we sure the average employee stays at an AI lab for more than a year? If not, that proves that a chain of people with sub-one-year time horizons can do valuable work). The AI Futures team guessed that the time horizon necessary for AIs to really start serious recursive self-improvement was between a few weeks and a few months (though this might look like a totally different number on the METR graph, which doesn’t translate perfectly into real life). If this is true, then all three anchors (seconds, hours, years) were off by at least an order of magnitude. But it turns out that none of this matters very much. The highest and lowest anchors cancel out, so that the most plausible anchor - human brain with time horizon of hours to days - is around the average. If you remove all the other anchors and just keep that one, the model’s estimates barely change. But also, we’re talking about crossing twelve orders of magnitude here. The difference between the different time horizon anchors doesn’t register much on that level, compared to things like algorithmic progress which have exponential effects. Maybe this is the model basically working as intended. You try lots of different anchors, put more weight on the more plausible ones, take a weighted average of each of them, and hopefully get something close to the real value. Bio Anchors did. Or maybe it was just good luck. Still hard to tell. Eliezer Yudkowsky argued that the whole methodology was fundamentally flawed. Partly because of the argument above - he didn’t trust the anchors - but also partly because he expected the calculations to be obviated by some sort of paradigm shift that couldn’t be shoehorned into “algorithmic progress” (like how you couldn’t build an airplane in 1900 but you could in 1920). As of 2026 - still before AGI has been invented and we get a good historical perspective - no such shift has occurred. The scaling laws have mostly held; whatever artificial space you try to measure models in, the measurement has mostly worked in a predictable way. There have really only been two kinks in the history of AI so far. First, a kink in training run size around 2010: Second, a kink in time horizons around 2024 and the invention of test-time compute: The 2010 kink was before Cotra’s forecast and priced in. The 2024 kink is interesting and relevant - but since it was on a parameter Cotra wasn’t measuring, and probably too small to show up on the orders-of-magnitude scale we’re talking about, it’s probably not a major cause of the model’s inaccuracy. Other things have been even more predictable: So Cotra’s bet on progress being smooth and measurable has mostly paid off so far. But Yudkowsky further explained that his timelines were shorter than Bio Anchors because people would be working hard to discover new paradigms, and if the current paradigm would only pay off in the 2050s, then probably they would discover one before then. You could think of this as a disjunction: timelines will be shorter than Cotra thinks, either because deep learning pays off quickly, or because a new paradigm gets invented in the interim. It turned out to be the first one. So although Yudkowsky’s new paradigm has yet to materialize, his disjunctive reasoning in favor of shorter-than-2050 timelines was basically on the mark. Nostalgebraist argued that Cotra’s whole model was a wrapper for an assumption that Moore’s Law will continue indefinitely. If it does, obviously you get enough compute for AI at some point, even if it requires some absurd process like simulating all 500 million years of multicellular evolution. I never entirely understood this objection, because - although Bio Anchors does depend on a story where Moore’s Law doesn’t break before we get the relevant amount of compute - this is only one of many background assumptions (like that a meteor doesn’t hit Earth before we get the relevant amount of compute). Given those assumptions, it does a useful not-just-assumption-repeating job of calculating when transformative AI will happen. As Cotra implicitly predicted, we seem on track to get AGI before Moore’s Law breaks down, and so Moore’s Law didn’t end up mattering very much. And if all of Cotra’s non-Moore’s-Law parameter estimates had been correct, her model would have given about the same timelines we have now, and surprised everyone with a revolutionary claim about the AI future. But Nostalgebraist added, almost as an aside: Cotra has a whole other forecast I didn’t mention for “algorithmic progress,” and the last number is what you get from just algorithmic progress and no Moore’s Law. So depending on how much you trust that forecast, you might want to take all these numbers with an even bigger grain of salt than you’d expected from everything else we’ve seen. How much should you trust Cotra’s algorithmic progress forecast? She writes: “I have done very little research into algorithmic progress trends. Of the four main components of my model (2020 compute requirements, algorithmic progress, compute price trends, and spending on computation) I have spent the least time thinking about algorithmic progress.” ...and bases the forecast on one paper about ImageNet classifiers. I want to be clear that when I quote these parts about Cotra not spending much time on something, I’m not trying to make fun of her. It’s good to be transparent about this kind of thing! I wish more people would do that. My complaint is not that she tells us what she spent time on, it’s that she spent time on the wrong things. Like Cotra herself, I think Nostalgebraist was spiritually correct even if his bottom line (about Moore’s Law) was wrong. His meta-level point was that a seemingly complicated model could actually hinge on one or two parameters, and that many of Cotra’s parameter values were vague hand-wavey best guess estimates. He gave algorithmic progress as a secondary example of this to shore up his Moore’s Law case, but in fact it turned out to be where all the action was. V. Those were the rare good critiques. The bad critiques were the same ones everyone in this space gets: You’re just trying to build hype.
Inline links: ”Measuring the Algorithmic Efficiency of Neural Networks”, ”Algorithmic Progress in Six Domains”, later research, fact check: true!, argued that, https://epoch.ai/blog/the-longest-training-run, https://www.lesswrong.com/posts/GAJbegsvnd85hX3eS/thoughts-on-extrapolating-time-horizons, https://epoch.ai/benchmarks/eci, argued that