Exploring schemas and propositions about language models of all kinds on https://xitter.twitterapi.xyz/t.co/GU07uzaA4F and on https://xitter.twitterapi.xyz/t.co/VDN6KwY9h8.
UK
Joined April 2009
- Tweets17.5K
- Following830
- Followers2K
- Likes11.3K
Pinned Tweet
10 Things that are not really* a concern for AI in 2026:
1 Hallucination
2 Training on your data
3 Bias in models
4 Dependence on proprietary models
5 Running out of pre-training data
6 Jail breaking
7 Model collapse
8 Sycophancy
9 Knowledge cut off dates
10 Models trained "on the Internet"
*Not that they don't exist or may never cause problems but because they are not a systemic concern worth worrying about.
I don't know if folks remember how we thought the single forward pass was a problem and considered solutions like a backspace token. Bitter lesson strikes again.
(1/10) Today, we’re sharing new research from @OpenAI: How does CoT monitorability (no-CoT capability and CoT controllability) evolve over the model training lifecycle?
We emerged with unsurprising but sobering findings: as we scale toward more intelligent systems, models are becoming harder to monitor through their CoT. 🧵
38
"Intelligence isn't magic."
It's amazing how many people on both sides of the AI debate seem to forget this basic truth.
I am pretty confused about how people can look at the pace at which machines are becoming more intelligent and not draw some basic inferences from the simple facts that:
1) Humans don't have plot armor.
2) Human intelligence has transformed the world so much.
3) Intelligence isn't magic.
to conclude that truly insane things could happen relatively soon. I don't think this relates at all to how quickly companies adopt or find uses for AI as it currently exists.
1
1
84
This should get some love, too. Cheaper and as good as Haiku 5.5 - that's a good model to have around.
Releasing Muse Ember today -- our smallest model that delivers higher intelligence per dollar than others in its class. Built for token-heavy work that agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing. Available today on the Meta Model API and in Muse Code.
53
Worth knowing: Run Fable on High rather than Opus on Max (if you're full enough to do that):
"With high and xhigh effort, Claude Fable 5.2 scores 62, ahead of every other model at any effort, using 98k output tokens per task against 173k for Opus 5.5 (max). At max effort Fable 5.2 costs ~2.6x more than Opus 5.5, but at lower effort settings it can outperform Opus 5.5 at lower cost per task"
Anthropic has released Claude Fable 5.2, taking the #1 spot on the Artificial Analysis Intelligence Index v5 at 65, led by gains in science, knowledge, and coding
At max effort it scores 65, 5 points ahead of Gemini 4 Carbon (high, 60) and 6 ahead of Claude Opus 5.5 (max, 59). Gains are driven by scientific reasoning, knowledge and FrontierCode 2.0, and roughly matching Opus 5.5 on AA-Briefcase and Terminal-Bench 4.0.
Priced at $8/$40 per 1M input/output tokens, a 20% reduction from Fable 5.1, Claude Fable 5.2 averages a Cost per Task of $26.5 with max effort, a ~2.6x increase compared to Opus 5.5 (max, $10.0) and ~2x compared to Fable 5.1 (max, $12.99)
Key takeaways:
➤ Claude Fable 5.2 leads 5 of the 9 evaluations in the Intelligence Index: At max effort, Fable 5.2 scores 65, ahead of Gemini 4 Carbon (high, 60), Claude Opus 5.5 (max, 59) and GPT-6 Astra (max, 57), and ~15 points above Claude Fable 5.1 (max, 49). It sets the highest scores on HLE-Diamond (80%, previous best 69% from Gemini 4 Carbon), AA-Omniscience v1.1 (67, previous best 54 from Gemini 4 Carbon), Terminal-Bench-Science (73%, previous best 65% from Gemini 4 Carbon), CritPt v2 hard subset (77%, previous best 72% from GPT-6 Astra) and FrontierCode 2.0 (49%, previous best 46% from Claude Opus 5.5)
➤ Less progress in agentic knowledge work and terminal coding: Claude Fable 5.2 improves on Fable 5.1 across capabilities, but sometimes only to Opus 5.5 levels. It reaches 1785 Elo on AA-Briefcase, within the confidence interval of Opus 5.5 (1800) and just behind Claude Sonnet 5.5 (1813). On GDP.pdf it reaches 24% all-pass against 26% for Opus 5.5. On Terminal-Bench 4.0 it scores 60%, level with Opus 5.5 and behind Kimi K3.1 Pro (67%), Gemini 4 Carbon (65%) and Sonnet 5.5 (64%). AutomationBench-AA is an exception, at 76% against 70% for Opus 5.5, narrowly behind Gemini 4 Carbon, 78%
➤ Lower efforts still take leading positions: With high and xhigh effort, Claude Fable 5.2 scores 62, ahead of every other model at any effort, using 98k output tokens per task against 173k for Opus 5.5 (max). At max effort Fable 5.2 costs ~2.6x more than Opus 5.5, but at lower effort settings it can outperform Opus 5.5 at lower cost per task
Other model details:
➤ Context window: 1 million tokens with image and text input
➤ Pricing: $8/$40 per 1M input/output tokens; cache writes $10 with 5-minute TTL, $16 with 1-hour TTL, and cache reads $0.20
➤ Effort settings: five (low, medium, high, xhigh, max). Intelligence Index evaluations were run at all five with Anthropic's default fallback enabled
Congratulations to @AnthropicAI on this impressive release!
87
Goal and budget engineering: "Set the goal, the budget, and how much checking you need before walking away."
BREAKING:
@AnthropicAI just dropped Claude Fable 5.2. It’s a paradigm-shifting model if you have a project ambitious enough to give it.
We’ve been testing it at @every, and I’ve seen this crazy explosion of creativity on the team. People are making their own creative software, games for their kids, and documentaries instead of slide decks.
Here’s your day zero vibe check:
• Coding: @kieranklaassen built Everycut, a video editor with the features we use from Premiere, Descript, and other tools—all in one app, with an agent built in. The first version was live in 3 days, and we immediately used it for our internal show-and-tell. This feels like a new era of professionals building their own professional tools, tailored to exactly how they work.
• Games and design: @jackcheng gave it a photo of his toddler’s messy playroom. It turned the actual toys into a playable racing game, with fun details like a stray shoe as an obstacle and a sippy cup as a power-up.
• Knowledge work: @hammer_mt turned his research into a documentary. The model came up with interview questions, he answered them on camera, and it cut the footage together. He called this the most productive 3 weeks of his career.
• A different way of working: Give it a big job and leave it running overnight or over a weekend. But tell it what “done” means: Mike found it could finish the build, then spend another day or two testing. Set the goal, the budget, and how much checking you need before walking away.
• The catch: It’s slow and burns through tokens. Our early-access usage was free, and these projects would have been expensive at API rates. It’s 20% cheaper than Fable 5.1, but for everyday writing, emails, or quick back-and-forth coding, I’d usually reach for something else.
The clearest signal: When our early access ended, @TylerNishida stopped working on some projects until he could get this model back. He didn’t want to build them with anything else.
The best measure of a new model is whether it makes you try things you wouldn’t have tried before.
Fable 5.2 absolutely does that.
Here’s our full day-zero Vibe Check:
53
Breaking! Teens less suicidal in 2025 than in 1995. Thank Gods for Social Media.
U.S. teens saw notable declines in self-reported sadness and suicidal ideation from 2021 to 2025.
Relative to 2021, they were also more likely in 2025 to:
1) play sports (49% to 58%)
2) do strength training (45% to 51%)
2) eat fruit >=twice/day (19% to 28%)
3) eat at least one vegetable/day (55% to 63%)
65
Swwweet...
We're giving Ultra and Team Premium users Fable resets based on how long you've subscribed:
<6 mo: 1
6–12 mo: 2
>12 mo: 3
Use them anytime by March 8.
Pro and Team Standard get a week to try Fable 5.2, up to half your weekly limit, plus one reset. support.claude.com/en/articl…
1
178
There goes my early night...
Claude Fable 5.2 is available today on all paid plans, on the Claude Platform, and on all major cloud platforms.
It’s $8 per million input tokens and $40 per million output tokens, down from $10 and $50.
Read more: anthropic.com/claude-fable-a…
58
"Lazy people invented almost everything worth having."
“AI is for lazy people.”
Yes. Exactly.
Lazy people invented almost everything worth having.
Someone was too lazy to carry things, so we got the wheel. Too lazy to wash clothes by hand, so we got the washing machine. Too lazy to do math on paper, so we got the calculator. Too lazy to walk, so we got the car.
Every great invention in history started with someone saying “I really don’t want to do this anymore.”
Hard work was never the goal.
It was the price we paid because we didn’t have anything better.
Now we do.
Call me lazy all you want.
I’m in very good company.
55
Great way to illustrate both benchmaxxing and LLM variability. And to show why Claude 5.5 Opus is so great.
Nice DeepSWE score! Does it survive a lowercase filename?
Introducing MaxxBench: 89 DeepSWE tasks with small changes to the instructions or starting code. Same engineering problem, slightly different requirement.
Models clustered around 74% on DeepSWE end up 23 pp apart.
Good at coding, or good at DeepSWE? 🧵
73
Dominik Lukes retweeted
The bitter lesson in 26 words:
Don’t be distracted by human knowledge, as AI has been historically.
Instead focus on methods for creating knowledge that scale with computation, like search and learning.
"RL post-training in its current form, in my opinion, is also not particularly elegant."
This is exactly the wrong bitter lesson lesson. The main message of the bitter lesson is - optimise for results - make computer faster, more data efficient and learning feedback loops tighter.
This may mean temporarily leaning into a bit of "how we think we think" and "how we think we learn" but this can only ever be used as a heuristic. Your aim is never to make Machine Learning "look" like a perfect picture of learning (human or animal).
Your aim is to enable it find ways of learning that are machine learning. And that can only be done by compute/data optimisation - this could be bigger or better. But it cannot be fully designed in one go based on any one idea of how learning works.
The most un-bitter lesson objections to Machine Learning are "but that's not how we learn" or "that's not how animals learn". Richard Sutton has not learned his own lesson in this and that's why ultimately it was not his direction in research that prevailed.
I've seen "the Bitter Lesson" stifle academic research many times, including my own. Too often, promising ideas are dismissed as "not bitter lesson–pilled enough". This is, of course, not Sutton's fault, but a misinterpretation of his original message.
Yes, it would be nice to have very clean ideas that scale with computation instead of relying on human inductive biases. But it is important to note that research is circumstantial; it happens within the constraints of our time.
Time and time again, we've seen ideas that, in hindsight, look "not bitter lesson–pilled", be eventually replaced by cleaner, more general approaches. But that doesn't mean those earlier ideas were misguided. Often, they were exactly what allowed us to make progress at the time. And I wish we, as a research community, did not dismiss those ideas/works at the moments they transpired.
RL post-training in its current form, in my opinion, is also not particularly elegant. Much of its success relies on extensive environment construction and data engineering to patch missing capabilities in frontier LLMs (or, in other words, "benchmaxxing"). Yet it's incredibly timely! It takes advantage of a time where building these data flywheels is increasingly feasible for well-resourced labs. We'll eventually see cleaner, more general approaches that replace parts of this machinery.
RLMs are another great example. Giving agents a REPL environment to programmatically manage their own context was a good step forward. I would argue it has also changed the way labs trained their model. And because of that, agents become better in managing their own context and that enables "cleaner" ideas like Context Language Models.
This applies more broadly. We're beginning to see a shift toward methods that "go where the agent goes", and this building on the foundations laid by foundation model pre-training research, be it ugly or not.
If I were to add to Sutton's tweet, I think we need more clever methods that don't necessarily have to be "bitter lesson–pilled", but challenge existing conventions and unlock new capabilities that overcome today's bottlenecks. An idea doesn't necessarily have to survive scaling forever to be an important contribution to AI progress.
1
104
