# Антропикалық модель 4 AI индустриясын таң қалдырды > Толық мәтін ағылшын тілінде **Published:** 2026 ж. 6 қараша **Author:** TheAIGRID **Source:** YouTube — The AI Grid **Language:** en-US **HTML:** https://aiqazaqstan.com/kk/articles/anthropics-model-4-just-shocked-the-ai-industry Тәуелсіз зерттеу тобына Anthropic ішіне сирек рұқсат берілді. Оның тапқаны көптеген AI зерттеушілерін шынымен алаңдатты. MER мәліметтері бойынша, Anthropic қазір өзінің AI үлгілерін бір жыл бұрынғыдан шамамен екі есе жылдамдықпен жетілдіреді. Және оның көпшілігі жылдамдатады ... An independent research group was just given rare access inside Anthropic. And what it found has a lot of AI researchers genuinely worried. According to MER, Anthropic is now improving its AI models at roughly twice the speed it was just a year ago. And most of that speed up seems to be coming from the AI itself. MER also revealed that Anthropic has a secret internal model, which it calls model 4, that has never been released to the public. It's the highest scoring model in the entire report. And Anthropic's own staff say AI is now making them around 12 times more productive. And here's the part that really got people talking. Anthropic redacted most of the evidence that METR found the most informative. So, we can see the conclusion, but we can't see most of the proof. One researcher said they were deeply shaken by this report and that they thought we had more time. So, in this video, I'm going to break down what MER actually found, what we know about anthropic secret models, and why so many researchers are sounding the alarm. Which leads me to point number one, which is what METR just found. On the 5th of November, METR published a report with a pretty plain title, Independent Assessment of AI R&D Acceleration at Anthropic in October 2026. But what's inside it is anything but plain. METR is a nonprofit research group that tests what Frontier AI models can actually do. It's the group behind that famous chart showing how long a task an AI agent can complete on its own. The one that keeps doubling every few months. So when MER says something about the speed of AI progress, people in the industry listen. And for this report, MER got something that outside researchers almost never get access. From the 12th of August to the 29th of October, a small MER team worked inside Anthropic systems on Anthropic's own devices. They got internal benchmark scores, training data, compute figures, internal forecasts, staff surveys, and interviews with around 15 researchers. MER then wrote a full report of around 100 pages for anthropic and a much shorter public version. And MER was very clear about the question it was trying to answer. Is AI speeding up the process of building AI? Because that's the big one. If AI models start helping to build better AI models, which then help build even better ones, you get a loop. That loop is what people call recursive self-improvement. And it's the thing that a lot of AI safety researchers have worried about for years. If that loop gets going fast enough, progress could jump far faster than anyone can check whether these systems are safe. So, what did METR actually conclude? Three big things. First, as of October 2026, Anthropic is likely improving its model's capabilities at around twice the rate it was between March 2024 and October 2025. Second, most of that speed up, METR's rough guess, is around 80% is probably due to AI itself. And third, METR didn't find any clear bottlenecks that would stop this from speeding up even further. together. They're the reason this report has spread so quickly. Which leads me to point number two, which is Anthropic's secret models. One of the most surprising parts of this report is that it lists models that the public has never seen. MER refers to them as Model 2, Model 3, and Model 4. And in the methodology section, MER gives us their rough timelines. Model 2 finished training in July 2026. Model 3 finished in August to September, and Model 4 finished training between September and October. So just weeks ago, Andrew Curran summed it up nicely on X. Anthropic has a new internal model that finished training last month. MER refers to it as model 4. Anthropic has its own internal code name for it. It's the highest scoring model in the set MER analyzed, and it's the first to clear 50% on every benchmark it was tested on. And Salio pointed out that it's already pushing past 90% on several of those benchmarks. That 50% detail matters more than it sounds. MER measured these models on 15 benchmarks and Model 4 scored above 50% on all 15 of them. On some, it's already up around 90% or higher. People started reading between the lines quickly. An account on X called pointed out that since Mythos 5, Anthropic has already gone through three more internal flagship checkpoints. also claimed Anthropic told MER that model 2 was a new base model and that it expects later models built on that base to be much bigger jumps over Mythos 5. Dan McAeier went even further, guessing that model 4 could end up being Claude 6 and that it might be released early next year. That's speculation though. But one chart really shows how far ahead Anthropic's internal models are. Shing Yu Dang took METR's data and plotted Anthropic's internal frontier against what the public could actually use. The orange line is what Anthropic had inside the company. The green line is what you and I could access through Claude. And the gap between them keeps showing up again and again. According to this chart, Model 4 has been sitting more than 40 days ahead of anything the public can get. Which is why the post asked, according to this, how long have we been cooked? The most capable AI in the world isn't necessarily the one you can use. Which leads me to point number three, which is the chart that shows progress has doubled. This is the chart everyone's been sharing. It comes from the very first finding in METR's report. Since October 2025, benchmark scores for anthropics models, including internal ones, have been growing at around 1.9 times their historical rate. The score on the left is called AECI. It's Anthropics version of the EPO capabilities index, which is a way of combining lots of different benchmarks into one number, so you can compare models across time. Each dot is a model. Green dots are models the public got. Orange dots are the internal ones. And the dates along the bottom aren't public release dates. They're the dates when most Anthropic employees could actually use that model. Now, look at the slope. Before October 2025, Anthropics models were gaining about 15 points a year. After October 2025, they've been gaining about 29 points a year, almost double. METR put it in a way that I think makes it really clear. Anthropic appears to have made as much progress on benchmarks in the past year as it did between October 2023 and October 2025. Roughly, that's the whole jump from Claude 2.1 all the way to Sonnet 4.5. 2 years of progress squeezed into one. A two times acceleration means 2 years worth of progress in 1 year. So if something was growing three times a year, a two times acceleration would mean it's now growing nine times a year. Now, the best estimate is 1.9 times, but the uncertainty range goes from 1.5 to 2.8 times. METR said it wouldn't be surprised if the true number was closer to 1.3 times or three times. And when MER included models from other companies like OpenAI and Google, the industry-wide acceleration came out lower at around 1.4 times. But here's the interesting thing. Anthropic's own Haiku 5.5 system card had already reported an acceleration of around 1.7 times. METR's number is higher, mainly because it included the unreleased internal models. In other words, the models we can't see are the ones pulling the trend upwards. Zephr summed up the mood on X, saying, "Anthropic is now completing a new pre-train every 45 to 60 days, and capabilities are rising at twice the rate compared to the October 2025 baseline." Which leads me to point number four, which is that this isn't just about more computers. When people hear that AI progress has sped up, the obvious question is, did Anthropic just buy a lot more chips? But MER says no. MER's second finding is that the acceleration is not being driven by an acceleration in the scale of training compute. In the report, METR wrote that it received convincing evidence the acceleration was not due to an increase in the growth of training compute. And then the next line says further details on this evidence were redacted. So if it's not mainly more compute, where is the speed coming from? METR's answer is that most of it is probably AI. Its rough guess is that around 80% of the acceleration happening right now is due to AI helping with the research. METR says that number has substantial uncertainty and over time AI's share of the acceleration seems to have grown. For years, the main way to get a better AI model was to spend more on chips. But if this report is right, a growing part of the progress now comes from AI making the people who build AI much faster. That's why a lot of people looked at this report and immediately thought of recursive self-improvement. Benjamin Todd put it bluntly on X. Looks like recursive self-improvement started in October 2025. Which leads me to point number five, which is the 12 times speed up. As part of the report, METR looked at informal surveys and Slack polls of anthropic employees asking how much AI was speeding up their own work. And the trend is pretty dramatic. Back in early 2025 with Claude 3.7 Sonnet, people reported a speed up of around 1.2 times. By the time of Mythos 5, it was above five times. And for Model 4 in October 2026, the average self-reported speed up was around 12 times. 12 times. Now, MER is very clear that this is very loose evidence. These are self-reports. But even if you heavily discount them, the direction is hard to ignore. MER's curve has that speed up growing about 3.9 times every year. MER also did proper interviews and some quotes are striking. One researcher said, "They're now only bottlenecked on ideas and compute, no longer bottlenecked on code." Another said, "It is hard to see how I am not being three times more productive." But not everyone agreed. One researcher guessed only 15 to 20% of the speed up was due to AI and said they don't believe in self-reported productivity increases. Overall, the answers were centered around two times. The other important detail is how the AI is being used. According to MER, the AI is still mainly helping humans do research rather than doing the research completely on its own. But METR said there are signs of life for fully autonomous contributions coming soon. And that difference between AI helping humans and AI doing the work alone is exactly what Google Deep Mind CEO Deis Havis has been warning about. So, but I do worry about it because in the limit, so there's different forms of RSI. Right now, we're in the the early days of RSI, I would say, where what's happening is, and you're probably all experiencing this as well with your teams, is um you know, it's making a human engineer maybe three, four, five times more productive. Okay. So, uh but the but the human's very much still in the loop and reviewing code and architecting things and um and and and writing code still. Um but it's all just enhanced. Um the next phase which I think carries the most risk is closed RSI, closed loop where you basically manage to somehow figure out how to do it without the human reviewer process uh in this in the in the whole procedure and that um will obviously be faster but I think is also a very dangerous way to get to full AGI because we're likely for two reasons. It's it's sometimes used to be referred to 20 years ago as like an intelligence explosion because you don't quite know uh how you know how do you control keep control of such a system uh such a process and then the end result will likely be quite alien. It's going to be like an alien intelligence. There's no reason you to think it would be sort of in the similar kind of mind space that uh humans are in. So that will make it very hard for it to understand for us to understand it and maybe for it to understand us. So I think it's you know we should be very cautious as we get closer towards that closed loop form of RSI and I think all the AI researchers sort of know that >> by Habis' definition what MER found inside anthropic is still the early form of recursive self-improvement with humans in the loop. The big question is how long it stays that way. Now here's why this is so surprising. Earlier this year anthropic CEO Dario Amade went on Dwarsh Patel's podcast and Dwarkkesh asked a sharp question. If AI really speeds up research so much, why doesn't the company with the best coding model just pull away? And Daario gave a pretty modest number. >> So, no, no, no. I I I mean I mean I mean I think it's all like my my model of the situation is there's there's an advantage that's gradually growing. Like I would say right now the coding models give maybe I don't know a a like 15 maybe 20% total factor speed up. Like that's my view. um uh and six months ago it was maybe 5% and so and so it didn't matter like 5% doesn't register. It's now just getting to the point where it's like one of several factors that that kind of matters and and that's going to that's going to keep speeding up. And so I think 6 months ago like you know there were several there were several companies that were at roughly the same point because uh you know this this wasn't uh this wasn't a notable factor but I think it's starting to speed up more and more. So earlier this year, Daario's view was that AI was giving Anthropic maybe a 15 to 20% total speed up and that 6 months before that it was only about 5%. But Daario also said it was going to keep speeding up more and more. And now just months later, an independent group says the overall pace of progress looks like it has roughly doubled. If that's true, the snowball Dario was describing has gotten a lot bigger very quickly. Which leads me to point number six, which is that METR couldn't find anything to slow this down. One of the big arguments for why AI progress won't explode is bottlenecks. Maybe there aren't enough chips or good data or new ideas get harder to find. But METR's fourth finding is that it did not see convincing evidence of significant bottlenecks that would take effect within the next 12 months. Not in algorithmic progress, not in training compute, not in experiment compute, and not in data. And that's a really important finding because if there's no wall coming up, the acceleration could keep compounding. Now, MER said its view on bottlenecks changed several times. So, it's not confident in this. And once again, Anthropic redacted MER's analysis of a number of specific bottlenecks. So, we don't know what they were or why MER decided they weren't convincing. MER also said the acceleration could be higher than it found. The recent growth might itself be speeding up, which some of the redacted evidence hints at. But most of the MER staff on this project believe Anthropic is improving its models at around twice its old rate. Which leads me to point number seven, which is the threshold Anthropic may have already crossed. Anthropic has a set of safety rules called the responsible scaling policy. And one of the warning signs in that policy is about exactly this. It's called the two times acceleration threshold. [music] In simple terms, it's the point where AI is speeding up Anthropic's progress so much that the company would get 2 years of progress in a single year. If that happens, Anthropic's own rules say it needs to take extra safety steps. So, has that threshold been crossed? In its August risk report, Anthropic argued that it hadn't. METR reviewed that section of the report, and its review was published the same day, and METR disagreed with how confident Anthropic sounded. METR agreed that the actual risk from these models was low and it agreed that Anthropics AI can't yet replace its research staff. But METR said Anthropic's argument that the two times threshold had not been crossed was overstated and that crossing it was plausible. The public report said the leading indicators showed acceleration of less than a factor of two. But MER pointed out that Anthropic's own private report had figures implying about 1.97. METR suggested the public report should say around 2 instead. And then there's Anthropic's IPO filing. On the 2nd of November, Anthropic updated its S1, the official document a company files before it goes public. Theo Bman spotted the new language. Anthropic now says its internal estimates and expectations for future models rather than its current models indicate that it may imminently meet the responsible scaling policies definition of dramatically accelerated progress attributable to AI. It also says it hasn't designated any of its current models as exceeding this threshold yet. But there's another line in that filing that I think is even more striking. Anthropic says it has in the past not identified and may in the future not identify a significant acceleration in its rate of development until after it has occurred. Think about that. The company building these models is telling investors that it might not notice a big acceleration until it's already happened. And Daniel Cockatilo, who leads the AI futures project, reacted [music] to that filing, saying the team was spooked by this information and that if Anthropic really is about to meet the two times acceleration threshold, which Cockatilo called a big if, then timelines to full recursive self-improvement and super intelligence may be very short indeed. Then after MER's report came out, Cockatilo posted, okay, so MER confirms it. Which leads me to point number eight, which is what Anthropic redacted. Right at the top of MER's public report, there's a redaction summary. It says Anthropic had rights to redact any non-public information in the report and that it redacted most of the evidence METR found most informative. Most Anthropic said the redactions were aimed at removing confidential information that could accelerate other AI developers. And as a result, METR says it was able to share its highlevel conclusions, but not the evidence behind them. And it says the public has to rely on MER's judgment much more than would be ideal. The original plan also included checking how Anthropic splits its computing power between making models more capable and making them safer. But MER says Anthropic removed that part of the project midway through and said it would revisit it later. So MER reached no conclusions on it at all. Now, to be fair to METR, it also said it thinks it's extremely unlikely Anthropic deliberately misled it, and it judged the whole exercise to be reasonably robust. And to be fair to Anthropic, it did choose to let an outside group in, which most AI companies don't do at all. Sam Mckandlish, one of Anthropic's co-founders who leads much of its safety work, responded on X. Macandlish said Anthropic gave MER access to a range of sensitive internal data including benchmark scores for unreleased models and that MER had complete independence to form its own conclusions. And Mcandish wrote many of their conclusions match my own. AI is speeding up our research but it's hard to say exactly how much. But a lot of people weren't satisfied with that. The account AI not kill everyone is in memes quoted the line about anthropic redacting a majority of the most informative evidence and asked the obvious question. What did METR see? It said blog posts and interviews about slowing down aren't nearly enough. And Daniel Cockatilo pushed back on Anthropic's reasoning too, saying it was pretty rich to site others could go faster as a reason to redact so much important information. And here's the awkward part for Anthropic. Back in September, Dario Amade published a post called We must pace the frontier. In it, Dario wrote, "We must slow the pace at which we improve the capabilities of AI models." Dario argued that AI progress has become much faster since recursive self-improvement began, and that left unchecked, it could outrun our ability to understand and control these systems. And now, an outside report says Anthropic's own pace has roughly doubled. Which leads me to point number nine, which is why people think AI 2027 is coming true. If you've been following AI for a while, you've probably heard of AI 2027. It's a detailed scenario published in April 2025 by Daniel Cockatilo, Scott Alexander, and the team at the AI Futures Project, describing how the next few years of AI might play out month by month. At the heart of it is a simple idea, something they call the R&D progress multiplier. Here's how Scott Alexander, one of the authors, explained it on Dwaresh Patel's podcast. So, so we think they're going to uh start with agency training, expand the time horizons, get coding going well. Um, our theory is that they are to some degree consciously, to some degree accidentally working towards this intelligence explosion where the AIs themselves can start taking over some of the AI research, move faster. So 2025 slightly better coding, 2026 slightly better agents, slightly better coding. Um and then we focus on and we name the scenario after 2027. [music] Uh because that is when this starts to pay off. Um the intelligence explosion gets into full swing. The agents become good enough to help with at the beginning not really do but help with some of the AI research. So we we introduced this idea called the R&D progress multiplier. So, how many months of progress without the AIS do you get in one month of progress with all of these new AIs helping with the intelligence explosion? The whole scenario hinges on that multiplier. And now people are comparing METR's results directly to that scenario. Nicola Jerkovic wrote yet another correct prediction by AI 2027. In the scenario, the speed up from AI is about two times around December 2026. The post added, "I don't like that progress is roughly going as fast as AI 2027 predicted because in the scenario, super intelligence arrives in 2027. For what it's worth, Jerkovic's own best guess for super intelligence is 2028 or 2029. And Jerkovic isn't the only one. David Krueger, an AI researcher, admitted to being somewhat surprised at how good AI 2027's predictions have been and that we seem to be going faster." and Spicy Lemonade shared a tracker claiming that AI 2027 is now about 85% accurate on the predictions that can be checked so far. On its chart, Model 2, Model 3, and Model 4 sit right around the scenarios line and in some cases above it. That's a fan-made tracker, and only about 39% of its predictions have been checked. But Jeffrey Ladish, the executive director of Palisad Research, said something in an interview that I think explains why this has hit people so hard. I think it's really notable that this is a faster speed up than Daniel Cocatello's AI 2027 scenario predicts. >> So like I don't know you remember like a year ago uh a lot of people were sort of like down on Daniel Cotello because he updated a little bit like towards longer timelines than as predicted in AI27. Um and now he's saying no I think it actually might be shorter timelines than than we predicted. And you know, I don't I don't know. The AI futures project guys are the people who I know who have like the most rigorous models and they're looking at all the parameters and trying to like figure out what are the inputs, how will this impact the overall speed. And I'm just like, man, that's insanely fast progress for something that was already going really fast. That's what makes this moment so unusual. A year ago, Cockatilo pushed back the AI futures projects timelines a bit, and a lot of people took that as a sign the hype was cooling down. Now, Cockato is saying the opposite. According to Cockatilo, if you naively plug a two times acceleration in January 2027 into their model, you get super intelligence sooner than AI 2027 itself depicted. And the team is thinking harder and longer about this and hopes to share a longer update soon. Which leads me to point number 10, which is why AI researchers are sounding the alarm. The reaction to this report has been intense, and a lot of the strongest reactions came from people who used to be more skeptical. Andrew Curran said that for many people this report was a massive update and that a number of prominent people who were formerly skeptical are now saying recursive self-improvement in under 6 months and super intelligence in 1 to 2 years. And Curran added, "I agree with both." Curran also had a message for the AI safety community saying there was never any pleasure in telling them they had run out of time, but if their plans needed a supportive government, an agreement with China, or a window longer than a year, they were pointless. That's brutal, but it shows how much the mood changed. Sydney von Arcs wrote, "AI company employees have famously short timelines. I used to chalk that up to internal hype. I think I was basically wrong. Things are going about as fast as they warned, and we're on track for AI to obsolete humans at almost all intellectual tasks before the end of 2028." And one post in particular spread everywhere. Aiden wrote, "I'm deeply shaken by the METR report. I thought we had more time and I'm very scared of what's coming for my family, for the world. That post got almost 300,000 views. Then there are the people focused on what to actually do about it. Alex Bores, a New York politician, pointed out that OpenAI says it's on track to fully automate AI research by March 2028. Anthropic employees say they're on track by 2027, and that the METR report suggests they're likely right. And Bores warned that any mistake or misalignment in today's AIS could be multiplied in each generation. And Nicola Jerkovic said that given how fast things are moving, we have no guarantee the world will be notified if a company is about to achieve vastly superhuman AI. Jerkovic called that unacceptable and said companies should be required to regularly report metrics related to recursive self-improvement. And Jeffrey Ladish also gave a really honest take on being skeptical of AI company timelines and what Anthropic's own people have been saying. So I mean that my initial reaction is I don't know I I think I've been kind of skeptical overall about AI company claims about how close they are to recursive self-improvement. Um like not maybe that skeptical. I think you know I'm like well I maybe expect it in like 2028 like full fully automated AI R&D. Um whereas I think I talked to a lot of people at Enthropic and they say no I think in like you know early to mid 2027 we'll probably be able to do this like you know most of the AI development will be happening autonomously by AI agents and I mean that's I think that's it's extremely aggressive in terms of like that would transform the world so fast. Um but I've been a bit skeptical cuz I don't know and not because I think people are lying. I don't think they're lying. I think they're like it's just easy to sort of be within a company and like see a lot of progress and sort of drink your own Kool-Aid and be like, "Yeah, I think we can do it." You know, whatever. I think that's actually one of the fairest takes on all of this. Ladish isn't saying Anthropic is lying. The point is that it's easy to get swept up when you're inside a company watching things improve every week. But this time, it's not just Anthropic saying it. It's an outside group that went in, looked at the data, and came out saying the pace has roughly doubled. Which leads me to point number 11, which is the case for staying calm. Now, there are some really good reasons not to panic, and MER itself spells most of them out. First, benchmarks are not the real world. MER says it remains unclear how benchmark scores map to real world capabilities, and METR specifically raised the possibility that the gains are concentrated in well-specified, easy to grade tasks. Second, the 12 times number is a self-report from Slack polls. METR calls it very loose evidence. Third, the baseline matters. Before late 2025, Anthropic's models were often behind the industry frontier. So, catching up could make the acceleration look bigger than it really is. And when METR included other companies models, the industry-wide speed up came out at about 1.4 times, not 1.9. Fourth, the AI isn't running the research on its own yet. It's still mainly assisting humans. And fifth, METR agreed with Anthropic that the catastrophic risk from these models was low and that the AI couldn't yet replace Anthropic's research staff. So, this isn't a report saying super intelligence arrives next month. But here's the thing. Every one of those caveats is about how big the acceleration is, not whether it's happening. And METR itself said it would be surprised if there was no acceleration at all. Which leads me to point number 12, which is what happens next and my honest take. So what should we actually watch for? First, Anthropic itself. According to reports, Sam Mckandlish said Anthropic plans to publish its own analysis of its progress and where it stands on its responsible scaling policy in the coming weeks. Anthropic's IPO filing also says it's considering updates to that policy that would require human review and approval of certain decisions in its research process. You don't add a rule like that unless you think AI might soon be making a lot of those decisions. Second, model 4. At some point, Anthropic is going to release something built on these internal models. And when it does, we'll get a much better idea of whether that 50% on every benchmark actually translates into something that feels like a step change. Now, here's my honest take. I think this is one of the most important AI reports of the year. Not because it proves super intelligence is around the corner, but because it's the first time an independent group has gone inside a frontier lab and come out saying, "Yes, AI is now speeding up the process of building AI." and by a lot. But I also think the redactions are a real problem. But when the most important evidence is hidden, the rest of us are left just trusting the company and the evaluator. And as AI gets more powerful, I don't think trust alone is enough. If companies are getting close to the point where AI builds AI, the public and governments need to be able to see that clearly and early. And there's a strange irony in all of this. The company that published a post saying we must pace the frontier is the same company that according to this report is now moving at roughly twice its old speed. I don't think that makes Anthropic hypocritical. It shows how hard it is to slow down in a race when the tools you're building keep making you faster. Whether this is the start of an intelligence explosion or just a very fast year, one thing [music] is clear. The models we can see are no longer the whole story. The most powerful AI in the world might be the one sitting inside a lab helping to build its own successor. And that's something we should all be paying a lot more attention to. Let me know what you think in the comments. Do you think recursive self-improvement has really started or is this being overhyped? If you enjoyed this video, make sure to like and subscribe, and I'll see you in the next one. --- *Материал бастапқы дереккөзден алынған: [YouTube — The AI Grid](https://www.youtube.com/watch?v=7eQree8FhKI). Түпнұсқаны ашу.* **Тегтер:** #Anthropic #YouTube --- *AI Qazaqstan — non-commercial informational portal about AI in Kazakhstan. Machine translations and summaries are AI-generated.*