rule30 #1: Back and under new management
Meta spends on Scale, superintelligence, and reasoning models
Moneyball is back and under new management. Since 2022, this newsletter has explored the limitations of human decision-making and their implications for startup investing. Now we’re excited to share refreshed content aligned with rule30, our quant-VC fund. As we apply deep learning to venture, we’ll bring you insights and discoveries from our research to help you stay current in the fast-moving world of AI.
I’m Harry Law, a former researcher at Google DeepMind who writes Learning From Examples – a newsletter about the history of AI and its future. Each month, I’ll send you a short email about how to make sense of what’s been happening in AI over the last 30 days (including links to further reading for those who just can’t get enough).
This month’s big news is about the House of Zuckerberg.
According to reports, Meta is spending $14 billion to acquire 49% of Scale AI, the data-labelling firm that works with pretty much every major AI lab out there. As part of the deal, Zuckerberg will bring in Scale AI CEO Alexandr Wang (and several other Scale AI employees) to bolster a new division focused on ‘artificial superintelligence’ (ASI).
The efforts suggest Meta isn’t happy with its AI team, which – despite having access to one of the largest GPU clusters in the world – has struggled to release truly competitive models. Fearing it will fall behind in the race to artificial general intelligence (AGI), Meta is taking a leaf o ut of former OpenAI chief scientist Ilya Sutskever’s book: ignore AGI and shoot for its big brother. But what to do when you need fresh direction and the antitrust G-Men say you can’t buy another lab? You acquire just under half of a not-quite-developer and yank its CEO.
But hold on. Isn’t all this talk of superintelligence premature?
Yes. No. Probably. Well, no one really knows. Not academics, not developers, not the markets, and certainly not newsletter writers (but my view is that it’s not quite as sci-fi as it sounds, assuming agents can improve from real world experience).
On the more optimistic side of the spectrum is OpenAI CEO Sam Altman, who recently wrote on his blog that ‘Humanity is close to building digital superintelligence’.
Anthropic thinks job displacement is coming, and soon. Digital gods aside, the implication here is that what we have today is already good enough to break the economic status quo. Even Obama agrees.
The uncertainty boils down to the unpredictable nature of large models, which work much better than they have any right to. One interesting (and hotly contested) factor at play here is emergence: the models spontaneously get better for reasons we don’t fully understand.
On the more sceptical side, we have recent research from Apple that tries to dampen enthusiasm about recently developed ‘reasoning models’ like DeepSeek-R1 or OpenAI’s o3.
You can read their paper here, which argues that what we call reasoning in large language models is just a trick of the eye.
Unfortunately, the work has some methodological problems, which you can read about in a blog from yours truly here (with a more technical take here).
Nonetheless, that didn’t stop the media (noted AI enthusiasts that they are) from running to press with gushing headlines about the end of the AI project.
Meanwhile, the frontier research train keeps on rolling.
French AI darling Mistral released its new reasoning model, Magistral, which can generate a reasoning trace in European languages.
Alibaba’s Qwen 2.5 models can improve maths performance through reinforcement learning even with completely random rewards, suggesting the models may have been trained on synthetic data similar to test sets.
Evals outfit Apollo Research built a test to help figure out if AI systems know they're being tested, finding that some recently released LLMs had some level of awareness they were being put through their paces.
Another testing organisation, METR, said that recently released models are showing increasing enthusiasm for hacking their environments because they ‘seem misaligned with the user’s goals.’
Over in the world of AI policy, it turns out that large models (might) help the government work better.
That new work was from, believe it or not, the UK government. They built a system called Extract on top of Google DeepMind’s Gemini model to help council planners make faster decisions.
In international affairs, read this piece about how the recent ‘Spider’s Web’ operation in Ukraine shows what a world with a warfighting AGI might look like.
A new study from MIT argues that most people evaluate AI based on its perceived capability relative to humans. In other words, we like it when it can do stuff we can’t.
In education, China shuts down AI tools during nationwide college exams. In the UK, the government said that teachers in England can use AI to speed up marking and write letters home to parents.
And finally, some other stuff we thought was cool.
The folks at Inference Magazine wrote about how to mitigate AI-driven inequality. They reckon that while AI creates enormous demand for skilled workers who can use the models, history shows that technological inequality is best addressed by expanding access to relevant skills – rather than restricting the technology itself.
The Cosmos Institute launched a $1M grant programme to build AI systems that follow Vannevar Bush's 1945 vision of augmenting human intellect and John Stuart Mill's 1859 principles of free speech.
A new documentary hosted on Aeon, Engineering Earth, takes aim at geoengineering projects. Check it out if you want something suitably inspiring.
You may have noticed that it is especially difficult to build nuclear power plants. That’s because of the Linear No Threshold (LNT) hypothesis, a theory based on 1920s fruit fly experiments that assumes any radiation exposure increases cancer risk. LNT has driven nuclear power costs through the roof despite mounting scientific evidence that the human body can repair low-dose radiation damage.
That’s all for edition one. Let us know what you think, and what we should look at in future editions.
rule30 is a quantitative venture capital fund. We've dedicated three years to developing groundbreaking AI and machine learning algorithms and systematic strategies to re-engineer the economics of startup investing.