Abdella Ali retweeted
for the past few weeks, I’ve been exploring painting with Fable 5.2 in code. one of the most common critiques I heard about Opus 5.5’s paintings was that it was only emulating real art. I thought that was fair, so i’ve been asking Fable 5.2 to develop personas—artists with some sort of a back story and a consistent style roughly identifiable in time
not knowing what to expect, i’ve watched this project evolve in directions i never predicted. I give comments or steers, and Claude updates or revises the engines that power the brushes and paint. where paintings come up i like or find intriguing, i ask it to go deeper. these are some favorites.
each of these is a python program, some one to two thousand lines of code, that places every stroke through a simulated paint engine with different bristles. most take a few thousand strokes, but i’ve pushed much higher with varying results. Claude does not use an image model, nor render or trace in advance
Abdella Ali retweeted
Claude Fable 5.2 sets a new record on the Epoch Capabilities Index (ECI) at 174. The previous high was GPT-6 Astra’s 169. Fable 5.2’s score is above what the pre-Astra trend for reasoning models would have predicted.
This is notable given recent discussion of AI contributions to AI R&D, and we hope to release more work in this area soon.
This is based on testing we did prior to launch using only our own benchmarks. The ECI is continually updated as more benchmark results become available, and may change in the future.
Abdella Ali retweeted
The authors of the METR AI R&D report estimate that AI is nearly doubling the rate of capabilities progress. This implies dramatic *uplift*: as if researchers experienced time ~10x faster. You probably couldn’t achieve this even if you hired every qualified human in the world.
Abdella Ali retweeted
We’ve trained Claude to flag potential misuse in conversations, a technique we call self-assessment. Right after it responds, Claude gives a preliminary safety assessment of the conversation.
Since we began using self-assessment in June, it has surfaced misuse that our other safeguards weren’t designed to catch.
Because Claude has already read the conversation, self-assessment is very efficient. And unlike classifiers, which target a discrete set of harms, it draws on Claude’s own judgment, so it gets more accurate as our models improve.
Read more on our new Safeguards blog:
safeguards.anthropic.com/202…
Abdella Ali retweeted
My dear friends, I’d like to show you my latest toy project
Introducing Handwired: turn your regular JavaScript code into a neural net!
- no training
- exact outputs vs the JS program
- fully differentiable
All runnable in a browser, without GPU
Abdella Ali retweeted
During the research for this project, Opus agents made an interesting discovery: a LIDAR scan of the aircraft carrier.
On their own initiative, they realized that in 2024, Washington State DNR did an aerial LIDAR survey... which included Puget Sound.
"Could this LIDAR data include the Nimitz?" they wondered. They investigated. Sure enough, they matched raw LIDAR point data to what could only be a huge ship.
It was the Nimitz.
Nimitz was alongside at Puget Sound Naval Shipyard in Bremerton, Washington, for a maintenance period. She happened to be under the survey aircraft's flight path, so she's in the public Geospatial data.
(sources in comments)
With that public-but-buried data, Opus was able to match the model to nearly-exact dimensions from the laser scan.
How much other data like this is hiding in plain sight?
We’re releasing a broad range of new mathematical results produced by an internal frontier model.
We’ve been consulting with the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, and we have drawn on their advice and public recommendations to inform how we release these results.
github.com/openai/math
Abdella Ali retweeted
New in this morn's AI Agenda: OpenAI is close to solving another Millennium Prize problem, the Hodge Conjecture.
It could take extra time to announce, though, as the company figures out how to navigate its rocky relationship with the math community.
theinformation.com/newslette…
I really hope that people take Dario's timeline for when we could have significant Internet takeover by LLM 'botnets' seriously. A lot of people publicly and loudly dismissed his recent previous predictions about the impact AI would have on software development, and those turned out to be quite prescient.
Also a good reminder of the value of publicly expressing your predictions for the future, if you are confident about them.
We Must Pace the Frontier: I’ve written a new essay on why the AI industry should slow down, with a three-part plan for doing so.
Anthropic is unilaterally committing to the first of these steps. We’ll provide third-party evaluators with permanent, employee-level access to our systems, so that they can verify adherence to our safety measures, report on incidents, and assess models’ alignment during training.
You can read the full post here: darioamodei.com/post/we-must…
3
321
Arguments like these are so common, but they package in so many truisms you are just supposed to agree with first.
1. "Almost no risk harm so far" - is not zero, why is non zero risk and harm that you already acknowledge, not evidence -for- risk and harm?
2. Who said we were supposed to have cataclysmic risk events by now? What did they expect to see by now, and why? I have no idea who is being referenced by "they" and what they were saying in 2023 would happen in 2026? Is this just made up?
3. One of the largest claimed risks would be that we would start seeing models that are extremely capable at cybersecurity related tasks, which has many downstream risks that are very clear - how is not what we have seen vindication of that?
What surprises me the most is that very few people even ask these simple questions when people make statements like this, and I generally find when I do, people don't even know what they are referring to when they make these arguments. They have something in their head that is like a dream that they realize only after they are asked about it, never actually happened
The "argument against" it is that there has been almost no risk or harm so far. We were supposed to have cataclysmic events by now, given what they were saying in 2023. There's no reasonable perspective where X-risk obsessors have somehow been vindicated by NOTHING happening.
1
1
143
I guess the one caveat, is sometimes they will go digging, find a tweet by some random person who says something crazy, and use that to dismiss a much much larger group of very reasonable people - or like take a prediction that was "By the end of the year" and say "it actually didn't happen until February the next year" as proof that these arguments are bad
34
In my mind, the situation with OpenAI/Anthropic and NS is... Well it's complicated, and I appreciate to some degree that competitive pressures will result in "scoops" like this happening. I also appreciate the core philosophy that no one really owns science/math and the process behind it.
That being said, this sort of feels like an event that will force a disillusionment across mathematics, science, maybe academia in general. I think that's maybe a good thing, in the long run - but I maybe expected a more gentle easing into it - for the sake of PR? For the sake of... I don't know, being good examples for the models of the future, on how to behave when they collectively make us feel the way many of the researchers who have been scooped are feeling?
Overall this feels like crossing a very important point, of no return maybe, where we will all have to make peace with our new roles in the world that is coming.
1
48
One of the more interesting things about this is that the model made its own DSL for games. I remember I think it was a group out of MIT that was trying to establish a DSL pattern for LLMs and I think I was arguing with them at the time that it would just be a bottleneck.
I think they were doing interesting work, and it's still relevant if you feel like you need to be behind the wheel explicitly designing the entire app yourself, but I always thought this was not respecting the fact that models would very quickly become more capable at all of this than us.
Who knows maybe there's still a place for human driven DSL development, but I do wonder if those researchers still feel the same way.
GPT-6 Astra represents a step-function change in model capability for interactive reasoning problems. It scores 66% on ARC-AGI-3 using our standard harness, and nearly 100% with a continuous conversation harness and custom compaction, at a cost of roughly $360 per game.
In fact, the continuous harness version significantly outperforms our human baseline in action efficiency across almost all levels. When we examined the reasoning chains to understand how the model operates, we found it performing highly efficient, on-the-fly symbolic world modeling for each game and level. It goes as far as developing its own shorthand DSL to represent in-game situations -- essentially a game-specific algebraic notation.
Overall, Astra exhibits symbolic modeling behaviors we had previously only seen with sophisticated harnesses -- so harness capabilities are increasingly shifting into the model itself.
We see Astra as a major breakthrough in model intelligence.
Read our post on Astra and what these results mean: arcprize.org/blog/astra
1
72
Okay I know sometimes Claude does that thing where it makes you feel stupid because you struggle to parse the complexity of what its saying, almost because it feels like it's speaking a different language, but c'mon!
1
99
Abdella Ali retweeted
Replying to @allTheYud @HumanHarlan



























