AI Radar #33
Is there something it is like to have a J-space?
It’s been a good week for getting things done. We finally got access to both Fable and Sol, and many of us celebrated by building amazing things for the sheer joy of it. This week’s newsletter is full of saturated benchmarks and charts that go straight up (as well as a few that stubbornly don’t).
Keeping up, however, feels increasingly daunting. My workflow now is completely different from what it was six months ago and even so I’m painfully aware of how much more capability I could get from the models if I just had time to figure out how.
Nobody told me the singularity would be this busy.
Top pick
A global workspace in language models
Anthropic’s latest interpretability paper is a big deal. It brings new insight into how LLMs think, provides strong evidence that they can access some of their own thoughts, gives us a powerful new tool for seeing what LLMs are thinking about, and might make possible a new type of alignment training.
The paper introduces the J-space, a newly-discovered set of patterns that appear to play a central role in LLM cognition. The J-space is in some important ways analogous to the global workspace in humans, serving as a place where important concepts are broadcast throughout the model.
While the paper itself is long and technical, the accompanying blog post is clear and accessible and there’s even a five minute overview video. Of course Zvi is here with a deep dive. Highlighting the importance of the paper, Anthropic solicited commentary from some notable external experts in cognitive neuroscience, AI consciousness, and mechanistic interpretability.
This is important work in its own right and serves as an important data point about our understanding of AI. LLMs are in many ways mysterious—we don’t fully understand how they work, and we don’t fully understand how to control them. But they are no longer impenetrable black boxes: we have powerful tools for studying them, and theoretical models that generate significant, testable predictions about how they behave.
News
Fable is back
Fable is back, with even more hair-trigger classifiers than previously. Zvi has full coverage of what happened.
This episode was a chaotic mess that shows the critical need for clear, well-calibrated regulation of frontier models. The current ad-hoc system imposes immense costs on everyone who uses frontier AI while doing very little to actually manage risk.
I’m cautiously hopeful that we’ll look back on this incident as the trigger for a sensible regulatory framework, rather than as the beginning of a period of chaotic and unhelpful government flailing.
It’s fashionable to show off what you’ve built with Fable—I banged out a pretty great image library manager in a day, complete with a sloptastic promo site.
GPT-5.6
GPT-5.6 was just released following government review. It’s clearly a highly capable model, but it’ll be a few days before we have useful reviews. I’m hearing very good things and can’t wait to try it out.
Claude Sonnet 5 Is Not Frontier But Has Its Uses
Zvi reviews Sonnet 5. Short version: it’s a solid model in an awkward place: less capable than Opus, but not radically cheaper.
It’s best for fast iteration that doesn’t require maximum intelligence, use as a sub-agent (perhaps in combination with Fable), or price-sensitive applications at scale. I don’t see myself having much use for it.
Grok 4.5 is also here
SpaceX just released Grok 4.5. Let’s get the obvious part out of the way first: it simply isn’t a frontier model and I don’t know why you’d use it for anything except searching Twitter. But with that said, it appears to be more capable than expected—perhaps it isn’t quite dead yet?
Using AI
AI art as curation
Never in human history has there been an interesting conversation about whether a particular genre is “real” art or not, and debates about AI art are no exception.
I am, however, interested in the different ways that individuals find meaning in the process of creating art, AI-assisted or otherwise. For Andy Masley, creating AI art is a way to curate a worldview and build bridges to other people.
Capabilities and forecasts
AI doesn’t get better at this board game with practice
Epoch brings us another data point on continual learning: they have the models repeatedly play an obscure board game and see if they get better over time. Which, unfortunately, they don’t:
It’s a good benchmark, both for continual learning and for a particular kind of problem solving. With current models, however, there’s some ambiguity about exactly what limitation the benchmark is measuring. The models only improve modestly when given a human-generated strategy guide, which raises the question of how much of the problem is that the models aren’t learning good strategies over time and how much is that they can’t execute on good strategies even if they have them.
Sol is very good at legal research
It's useful to remember what it means when I say that GPT-5.6 Sol Pro has saturated prinzbench. These are the hardest legal research questions I could think of, in my specialty area. Many involve tricks and/or require very painstaking research through a large volume of legal authorities. I doubt that there is more than a dozen practicing lawyers in my specialty area (of any seniority) who would be able to beat GPT-5.6 Sol Pro on my benchmark.
There will doubtless be legal specialties where this isn’t true yet, but I expect we’re within a year of AI being truly superhuman for many types of legal research.
Scaling works. These researchers are betting billions it isn’t enough
Transformer reviews the case for alternative AI architectures.
There’s a range of opinions on this topic: the mainstream view is that current LLM architectures will go all the way to AI, with enough scaling and iterative technical improvements.
At the far end of the spectrum, some people argue that LLMs have fundamental limitations and will need to be replaced or heavily augmented by a completely different architecture. That’s certainly possible, but I have yet to see a compelling argument for why it’s likely.
As the Adam Marblestone points out in the article, that’s probably a good thing. Because LLMs are pre-trained on a vast amount of human text, they have a natural head start on understanding human values. Aligning LLMs is a hard unsolved problem, but it’s considerably easier than aligning alternate architectures that don’t have that head start.
The AI superforecasters are here
Scott Alexander brings us a deep dive on the state of AI forecasting. Current AI forecasters are getting very good—they aren’t quite at the level of the best humans yet, but they’re getting close.
I appreciate his skepticism about how much this will improve public discourse:
Every day, I see smart, tech-savvy people on Twitter voice opinions which a moment’s consultation with an AI - sometimes an AI that they themselves are building or investing in - would reveal to be definitely false and stupid. We all have geniuses in our pocket willing to advise us on everything, and instead we’d rather repeat inane conspiracies without consulting them. I’m pessimistic that the rise of AI superforecasters will change this too much.
This fits the broader pattern we see with AI. For a relatively trivial amount of money, everyone now has access to an AI that can build sophisticated software, produce professional-quality research on any topic, and predict the future with remarkable accuracy. For a small number of highly capable and motivated people, those capabilities will be transformative.
Most people, however, will find that to be too much work.
A significant increase in digital labor automation
The Remote Labor Index (RLI) attempts to measure how well AI can complete complex, real-world tasks. Models are scored on what percentage of their work is judged to be equal or superior to the work of human professionals. Scores are low but rising fast:
I wouldn’t think of this as “Fable is 16% of the way to replacing any knowledge worker”, but rather “Fable is 16% of the way to being a powerful productivity multiplier for a wide range of professions”.
Notably, they found that success rates were equal or slightly greater for long-horizon tasks. That’s very different from the pattern we see with coding—I’d love to see more work on which pattern applies to different task domains.
Alignment and interpretability
Studying AI welfare empirically
Eleos and the Center for Mind, Ethics, and Policy bring us a new report on Studying AI Welfare Empirically (summary thread here).
Our ability to study AI welfare is hampered by the fact that many of the phenomena we need to understand in AI are not well understood in humans. The report lays out some of the key questions and proposes a framework for approaching them. A good example of the complexity of the task is that when we talk about AI welfare, it isn’t at all obvious whose welfare we’re even talking about:
Strategy and politics
The case for (fixing) the Great American AI Act
Ever since Mythos, there’s been broad (though not universal) agreement that the US urgently needs federal AI regulation. Writing for Lawfare, Charlie Bullock argues that with some fixes, the Great American AI Act (GAIAA) is the best proposed framework for that regulation.
Like many other commentators, Charlie’s biggest concern is with the bill’s preemption clause. While the language of the bill suggests a narrow scope of preemption, he argues that a court would likely interpret it much more broadly.
Building out a solid federal regulatory framework is the most important consideration: I’d rather have an improved bill, but I’d take GAIAA as written over nothing at all.
What would actually reduce AI risk
Benjamin Todd presents a list of 11 priorities for helping the AI transition go well. It’s a solid list and I appreciate his emphasis on interventions that make sense from a range of perspectives.
My biggest disagreement is around how much international coordination is feasible in the current climate. A grand bargain that includes the middle powers would be great, but I don’t see signs of us moving in that direction.
Be careful about relying on Chinese open models
Prinz summarizes a Reuters article on CCP policy discussions (paywalled original):
The CCP has held meetings with top Chinese AI labs (including Alibaba, ByteDance and Z.ai) about potentially restricting overseas access to China's most advanced AI models, including both closed and open-source models.
If your plans for “AI sovereignty” rely on access to near-frontier Chinese open models, you should consider China’s strategy of building foreign dependence on rare earth minerals and then using that dependence as a strategic cudgel.
The robots are here
Robots are clearly going to play a critical role in the AI future. While the US holds a significant lead over China in AI itself, China has a massive and deeply entrenched lead in robotics.
ChinaTalk takes a deep look at robots in China, covering the importance of robots as a general-purpose tool, the extent of Chinese control over the entire robotics supply chain, and what the US should do to catch up.
It’s a grim read from an American perspective:
But cut off the component supply chain and we don’t have robots. They’ve already threatened this — CNC-machine exports to us are already dropping, and they’ve put export controls on the rare earths this requires. We might make some neodymium, but we can’t process it, and we have no dysprosium or boron — so where’s that coming from? They can cut us off in more places than we can pivot, especially this early.
It’s widely believed in some circles that when we get to superintelligence, the AI will figure out how to speedrun robotics and mass industrialization. That’s certainly possible, but it doesn’t obviate the critical need to prioritize building robust domestic manufacturing capacity across the entire supply chain.
Beware delaying public deployment
Eli Lifland argues that—counterintuitively—delaying the public deployment of new models potentially increases risk.
There are complex arguments in both directions, and the right answer will depend on specific circumstances. My best guess is that he’s largely right, although it sometimes makes sense to delay the public release of specific dangerous capabilities. Releasing Fable publicly but restricting Mythos to Project Glasswing participants makes a ton of sense.
Risks
Rare viruses, vaccines, and more in over 1 trillion metagenomic reads
The SecureBio Detection program (formerly the Nucleic Acid Observatory) is doing vital—and fascinating—work monitoring wastewater for evidence of novel concerning pathogens. Their latest post reviews some of their recent detections, all of which ultimately turned out to be benign, but which illustrate their impressive capabilities.
Synthetic biology reaches a new milestone
The New York Times reports on a major advance in synthetic biology:
Blending together dozens of ingredients, the researchers have synthesized simple cells that feed, grow, reproduce and compete with one another for food. If these cells are not yet fully alive, they have most of the hallmarks of life.
This is interesting work in the abstract and has enormous practical potential. Congratulations to Dr. Adamala and team.
But much like mirror biology, synthetic biology has serious x-risk implications. Synthetic organisms have obvious potential for weaponization as well as catastrophic accidents. I’m glad to see that Biotic is prioritizing safeguards at their first meeting, but I’m deeply skeptical that open-sourcing the technology is the best way to do that.
People and data
The history of Claude Code
The history of Claude Code as told by the team, with a great observation by Ben Mann:
you have to build something that works 20 or 30 percent of the time now, so that when the next model comes out, it works 80 percent of the time. And that’s enough to get market traction. And then the next model after that is 90-something percent, and then you really make progress.
Cyber Vulnerabilities | Epoch AI
Bringing welcome rigor to what we already knew, Epoch charts the abrupt rise in reported cybersecurity vulnerabilities.
AI futurism reading list
As part of a recent strategy fellowship, Redwood Research’s Alexa Pan put together a reading list on AI futurism. It’s very well curated: this is my new favorite place to send people who are serious about learning more.
Side interests
Saving Gemini: The 9-min road to recovery
Hey, do you guys remember that one time when Gemini 2.5 Pro spiraled into psychosis and we all had to stage an intervention?
