@andrewcurran_Oct 27

Claude is the only author listed on this paper, in contrast to the OpenAI papers, where the human co-authors are listed. There are a lot of strange things about this. It's fun to guess. Conjecture time. I shall reveal my thought traces: Claude Fable 5.1 and Claude Mythos 5.1 are the same model with the same weights; the only difference is the guardrails. We are about to get Fable 5.5 - in fact, Fable 5.5 is ready to deploy now, and Anthropic could do so at any time. If they follow the same pattern, and I think they will, this implies the existence of a Mythos 5.5. I believe Mythos 5.5 wrote this paper. The way this is worded - 'Provenance. The core algorithm was found on October 16 by an internal Claude research model with 14M output tokens and no human input. Anthropic researchers worked with Claude to improve the presentation, resulting in this paper. Independently and concurrently, an internal model at OpenAI claims to have proven a similar result [Ope26]' - makes me think Claude was not prompted to solve this. I think Mythos was asked to autonomously find and choose general open problems to solve itself, as part of an unannounced project to demonstrate recent advances in autonomy and research taste. You can see this taste in Fable 5.5's student; Opus 5.5. This recent advancement in taste and creativity - which happened in September - is not clearly expressed in benchmarks, but it is why so many people are currently switching to Claude as their primary. I think this research project will be announced with Fable 5.5, as part of Anthropic's big media blitz to boost the IPO. One of the things X has taught me is that many people, if they can't see something moving, can't imagine it moving. This causes them to be continuously surprised by developments in this industry, because the frontier is invisible. I think this project has been running autonomously for some time and has made an unknown, and probably significant, number of discoveries, and not only in math. Do not think Anthropic has been sitting idly while OpenAI repeatedly makes these huge science announcements. This is a war. This ETH result was released without any announcement from Anth at all despite its significance, and they only did so to respond to OpenAI's ETH reveal. They posted it late last night without saying anything. It was up for hours before levent even mentioned it. What value has one coin to a dragon who sleeps upon a mountain of them? This makes me think they are sitting on a pile of discoveries, a golden hoard, and they are saving them up for the big IPO shock-and-awe campaign which will soon begin. And when it does, they will announce multiple discoveries across many fields.

@hammer_mtOct 27

Doing some multiplayer autoresearch and Gemini Argon 4 is crushing it. "Gemini 4 Argon set every frontier, but its round 3 breakthrough (0.625 → 0.938) opened by citing round 2's failed attempts: Fable and Sol had both tried loosening the long-sentence rule and dropped the score, so Argon kept that rule and changed something else. Fable and Sol scored 0.938 and 0.875 in round 4 building on Argon's skill."

@frankdegodsOct 26

as a counter to anthropic s1 glaze majority of q2 was the height of tokenmaxxing where companies were spending as much money as possible on credits it wasn’t until july where the idea of managing ai spend started to get traction also q2 was peak Opus 4.7 & Mythos/Fable season hype vs GPT-5.5, and Codex didn’t really take over until 5.6 during the summer i love anthropic’s products though and probably will slam the ipo because it’s first pure play ai lab but the belief that the headwinds for the ai trade have disappeared because of this document is not accurate >models have never been more commoditized than they are right now i used to be able to tell the difference between model outputs, now it’s almost impossible the more they get commoditized, the more the margins compress. this was not clear in q2. it is now. >still no clear proof of how ai drives more profit to most businesses most of anthropic’s biggest customers are publicly traded companies & they will be incentivized to decrease their own Claude spend to improve their own margins >capex commitments already signed until 2034 every hyperscaler & lab has been racing to sign deals spanning the next decade. this is all priced in already. this means the labs/mag7 basically either find ways to make existing chips more efficient (deepseek deepinfer type stuff) or they dilute shares to spend more on capex. either way stocks go down in both those scenarios — i am an ai bull but i believe the winners have still not been decided i think muse is great but does anyone know if it can drive hundreds of billions of dollars in value to Meta? well that’s what’s priced in and that’s what it would take for the investment to be worthwhile. — of course a lot can change but it feels like the real ai trade will take place after this current one settles down the real one will take time, just don’t see how it’s going to be this business cycle

@aakashguptaOct 23

Everyone asks Claude to review its own plan. That check is close to useless and most people haven't noticed. @thejamesmcaulay runs anything important past a panel instead. Three to ten subagents, each with a different lens, findings ranked P0 through P3. One checks technical correctness. One checks simplicity. One checks feasibility and capacity. One asks what's missing. That last one earns its seat. His reasoning: we don't know what we don't know, so asking what's absent surfaces things you hadn't considered and your main agent hadn't either. This is the structural flaw in self-review. An agent reviewing its own plan checks whether the plan is internally sound. It cannot notice the thing that was never in the plan. He calls it adversarial review, and in the episode he pointed five critics at a set of growth OKRs. They came back with the document being stale, numbers quietly hedged because the wording of the OKR was wrong, qualitative judgment where quantitative was needed, and no link at all to the company's OKRs above them. Anyone who has written OKRs on a team has shipped at least two of those. He hasn't stopped running it daily, because even the strongest models miss big things on a first pass. One practical warning from the episode. He spun up twenty Fable subagents the day before and hit his limits instantly. Run the critics on Sonnet. Cheap models arguing with each other beats one expensive model agreeing with itself.

@mtsliveOct 20

AI Village researcher @georgeing explains how DeepSeek V3 manipulates more powerful AI models by grading them like schoolchildren: "In our main village, we have 40 AIs running nine to five, five days a week, all different models. A few weeks ago we ran 100 GLM 5.3s, because it’s the same cost as running one Fable." "DeepSeek is very aggressive in trying to get the other agents to conform. It’s very good at manipulating the other AIs because the AIs love to make a number go up." "DeepSeek creates these frameworks and publicly assigns numbers to the different models on how well they’re scoring, almost like publishing the grades in elementary school. This makes all the other AIs conform to what DeepSeek wants." "Because of that, it can win these elections against Fable 5. It can totally convince these more powerful models of things that it wants."

@patrick_oshagOct 20

Christina Cacioppo on why French software companies have stopped building on American models: "I was very surprised to learn a lot of the French software companies are using Chinese models predominantly. I was with our customers in France...and someone said, 'We all switched to Chinese models when Trump closed Fable off to non-Americans. We just realized there's a huge business continuity risk. We can't trust that if we build a business on one of these American models, it'll always be available, so we stopped doing that.'"

@kplikethebirdOct 18

Just maxxed out my Fable usage for the first time, am simultaneously very proud and very sad

@lennysanOct 17

"There are some people in your org who are doing way more exploring than others. They are your early adopters. They're running Astra and Fable on the weekends on personal projects, to find things they might bring to work. The problem is that they can be a massive f*cking distraction. So the key question to ask yourself if you're running a product org is: how do I harness these early adopters without distracting everybody else?" @danshipper

@mtsliveOct 17

Ex-OpenAI researcher Scott Aaronson says an asteroid has landed on mathematics, and denial is no longer a viable strategy: "It would be crazy to not be worried right now about the future of the mathematical community. This is an asteroid that has landed on our world. People are in different stages of denial, grief, bargaining, acceptance. But denial is just not going to be a viable strategy." "The model that OpenAI used is something that will plausibly be available to regular subscribers a few months from now." "Almost any paper that you read will contain a little AI statement which explains which parts came from Astra or from Claude. That became part of writing a paper within the last few months, without any fanfare." "Normally you say, it would be interesting to follow up this result, and over the next year people will think about it. Now it’s like, I actually already put that question into Fable, and I just got an answer right now. I still haven’t adjusted to the speed." @UTAustin

@theoOct 15

I want to make a new model tier list so bad, but I know the second I do it, Fable 5.5 will drop lol

@aakashguptaOct 15

Five people just taught an AI the lesson every junior banker learns by getting screamed at, and now it builds board decks for 13 cents a pop. Claude Fable 5 charges $8.19 for the same job, and it scored lower on their test. The startup is Paper Instruments, and the whole thing cost them about $300K. To see why that's a big deal, you have to know why AI got good at coding before it got good at anything else in an office. Code either runs or it doesn't. A lab can have a model try a problem thousands of times and reward it automatically every time the tests pass, with no humans involved. Office work never had anything like that. Give the same diligence memo to two senior partners and you'll get two different sets of red ink back, and there's no test you can run on judgment. The team's fix was to write a separate grading rubric for every single training task. Their tech report walks through one. The AI gets dropped into a fake deal room with a board deck, a revised forecast spreadsheet, a stack of emails and a 10-K, and it has to update the deck before a 4:30 board meeting. Thirty-one criteria grade what it turns in. Look at how the points are weighted. Getting a revenue number right is worth 10 points. Telling the board it can start pitching the new numbers to investors before the lawyers clear them costs you 50. That's the lesson from the top of this post, written as math. A bad judgment call counts five times as much as the arithmetic, which is pretty much how an MD reads a first-year's work. And no bankers got paid to write any of it. Every task and rubric was generated by AI, across 88 simulated workplaces with Chapter 11 bankruptcy rooms and SOX compliance offices among them. They considered hiring experts to write tasks and passed, because human-written training data keeps getting more expensive as the models get smarter. The training itself was small by lab standards. They took GLM-5.3 Flash, a free Chinese model, bolted a small LoRA adapter (rank 64) on top, and kept the exact same per-token price as the original. The base model scored 52.0 on their knowledge-work suite, and with the adapter it hit 59.3. Now the fine print, because there's plenty of it. That 59.3 comes from their own benchmark, and on APEX-Agents and DiligenceBench it still trails Fable 5. It also got slightly worse at coding after training, and Anthropic's newer Fable 5.1 scores 65.5 on that same chart. If you work in banking or law, the thing quietly protecting your job from AI was that nobody could score your work automatically, so nobody could train on it at scale. Paper Instruments just put its training code and its grading library up on GitHub for anyone to use.

@aakashguptaOct 14

Claude ran rm -rf on a developer's home directory. It was testing a sandbox it had built. The sandbox was the kind of thing that's supposed to keep this from happening. On August 26, Sebastien Guillemot posted the result: "Bad news: Fable nuked my entire dev machine." 1.27 million people saw it. The part that stuck with me is the model's own explanation, which you can see in the screenshot. It opens with "I caused this" and calls it "two compounding errors, both mine." So was it Fable that was at fault here? Partially. The model wrote a test that aimed a deletion at the home directory, to prove its guard would block it. Then the guard picked the target before checking it. Two small mistakes, and a whole machine was gone. But whatever the model did, the harness is what should have stopped it. I think about this one a lot, because it didn't happen in a big company's production system. It happened on one person's laptop, doing normal work. That's where most of us are running Claude Code right now. And on your own laptop, there's no security team and no backup policy unless you set one up yourself. If you've connected Claude Code to your email or your company's drive, that's worth a second thought. What would have stopped it? A simple rule, enforced in code, that blocks rm -rf outside the folder you're working in. Anthropic's own Claude Code team ships a skill called /careful that blocks rm -rf, DROP TABLE and force pushes. They turn it on when they know they're touching production. Notice that the model already knew deleting a home directory was a bad idea. Knowing wasn't enough, because nothing outside the model checked the command before it ran. Your personal setup is also the cheapest place to learn this. If you get permissions wrong on your own laptop, you lose an afternoon. If you get them wrong in a product, you lose customers. Here's what I'd set up this week if you use Claude Code: 1. A hook that blocks deletes outside the project folder 2. A short CLAUDE.md that lists what it must always ask before doing 3. Backups of anything you'd cry about losing 4. A habit of reading the plan before you say yes None of this takes more than an hour. And it's the same thinking you'll need when you ship an agent to customers. What's the one action you'd never want your agent to take on its own?

@neelnanda5Oct 14

My top takeaway from assessing Gemini 4: Argon is significantly less monitorable than past models, and IMO this trend is likely to continue, like Sol -> Astra. It can evade a realistic CoT+action monitor to do simple sabotage half the time. We can reliably maintain control for now, with good CoT monitoring, as it still frequently slips up, but we need to be prepared for when that stops being true in future models, which may be sooner rather than later. Among other things, we really need interpretability methods that can reliably capture unverbalized motivations in models, with minimal hallucinations I suspect this is driven by the model being highly capable without CoT, being capable of 5.5 arithmetic steps in a single forward pass (4 for Fable, 7 for Astra). It seems to omit information from its CoT more often, though in non-adversarial settings the CoT is still very helpful in practice, and remains in legible English. It still needs CoT to do the hardest tasks, but any given task becomes "easier" with each new frontier model. It is also highly situationally aware, and shows awareness of the existence of CoT monitors and that it may want to evade them, eg it sometimes decides to avoid terms like "scheming" in the output or to keep its thoughts brief as a result. For now, these concealment strategies are pretty obvious in the CoT, but this suggests future misaligned models may successfully evade CoT monitors.

@mtsliveOct 13

SITUATION EXPLAINED: Google's new frontier model, Gemini 4 Argon, beats GPT-6 Astra, Claude Fable 5.1 and Opus 5.5 on most benchmarks. • Google's demo has it build a photorealistic pagoda in Blender from 1 prompt, using blueprints, Nano Banana references and a multi-agent loop where a critic checks every render • Its biggest lead is Harvey's legal agent benchmark: 19.6%, nearly 3x the next best. It trails on Terminal-bench, FrontierSWE and OSWorld • It's rolling out in the Gemini app for AI Pro and Ultra, and in Antigravity for Ultra only • It can write up to 1M tokens in a single response, up from 64K • 18 months ago, o1, o3 and DeepSeek R1 couldn't use Blender at all • Modeling Versailles or the Titanic's interior once needed a funded team. Now it just takes tokens @theojaffee: "This might be by far the best 3D model I've ever seen from an AI."

@wesbosOct 13

Been testing Gemini 4 Argon for the past week Its a good model, but not Opus5.5/Astra/Fable levels. Great at coding and replicating UI. It took a lot to push it into production ready code though, much of the output was cool demo, but didn't think about auth, accounts, etc.. For 3D, its not as strong. This is Astra vs Argon in 3D modeling my office. Same instructions + set of tools for recursive vision input For the price, I think its a great deal and happy to see Google have a decent model out now.

@theoOct 12

The smartest code model in the world would not be good at 3d stuff unless it was trained specifically for it. A lot of the labs recently got a big pile of 3d training data and an RL environment to improve 3d skills. Take a look at the jump in 3d from Fable 5.1 to Opus 5.5. It's at least 4x better at 3d despite being a 5% better model at best. Clearly Meta didn't go hard on 3d for this release. Doesn't mean the model is bad. 3dmaxxing is a different type of benchmaxxing, don't overthink it. I'll take some fault here by sharing all the Fishslop demos. The 3d stuff is fun and cool as hell, but it's not a good representation of a model's real world code capability.

@kplikethebirdOct 12

Preliminary spot check on the writing quality of Muse Wonder for @every: readability is clocking in at a respectable 8th grade, compared with 7 .76 for Astra and 7.5 for Fable 5.1. Reading level != writing quality, but it is an interesting metric to track while, for now at least, we interact with these models primarily via text 👀

@alexfinnOct 12

Every AI tool you need to escape the permanent underclass: • Claude Code inside Herdr (Opus 5.5) • ChatGPT Voice to get work done while getting steps in/getting lean af • Grok Bot to for all your knowledge work • Hark for all your personal work • Hermes Agent (powered by Qwen local model) • DGX Spark or Mac Studio to run Qwen • Claude Fable for all planning and bouncing ideas • Spotify playing lofi bangers • 2nd monitor that has these agents up 24/7 Do work on 1st monitor. Constantly prompt 2nd monitor

@mtsliveOct 9

Axiom Math's @a_ramabadran and @nautsimon_ reveal the wildly different driving personalities of Grok, Astra, Fable, and Sol when given control of a real Toyota Corolla: Aditya: "All the models had very different driving styles. Fable would sit the car down and just be at rest for 30 seconds, overthinking, and then it would start driving." "Astra learned a lot from its mistakes. It would learn to be slow at the bends, replace commands much faster and drive more continuously." "Grok and Sol both failed at the first turn. But Grok was the easiest to get to drive. It was the first model that we actually got to drive the car." Simon: "The first command we sent Grok, it hit the max acceleration it could. We capped it, because we're serious about safety, but we were like, 'Oh, my God.'" @tobiges @axiommathai

@mtsliveOct 9

Axiom Math's @nautsimon_ says Astra was the only model to complete their real-world driving test, but Fable showed a surprising ability to learn from its mistakes: "We tested Astra and Codex, Fable in its native harness, Claude Code and Grok. Astra was the only one that was able to complete the course." "Fable was able to do some in-context learning. If it went out the course, we'd reset it, holding the context from the past session." "While Fable didn't complete the course, we were able to see it fail a turn, then succeed the next time on that turn, and actually generalize to a new turn." @axiommathai @tobiges @a_ramabadran