Flashback to June 14, 2024.
At the time, frontier AI models have zero fluid intelligence. They score around 5% on ARC 1. They can't reliably add two numbers. They make laughable logic errors every other prompt. Reasoning models don't exist. Coding agents don't exist. Claude Code doesn't exist. "Code gen" is a complete joke. Scaling up next-token prediction based on web text data is hitting a wall.
What if a time traveler told you that, by November 2026...
• AI models would have human-like fluid intelligence, thanks to a new paradigm: test-time induction (modern reasoning models).
• Not only would ARC 1 be saturated, but unreleased benchmarks 10x harder (ARC 2) and 100x harder (ARC 3) would be saturated too.
• Nearly all code would be written by AI, and models could one-shot complex codebases flawlessly.
• AI models would be solving multiple Millennium Prize problems.
• OpenAI's valuation would be 15x higher, and Anthropic's would be 85x higher.
And on top of all that...
• Nvidia would still hold ~80% of the AI compute market.
• Nvidia would have grown its revenue and earnings 4.5x.
• Nvidia would still be growing 75% a year.
• Nvidia would have even higher operating margins (67% vs. 65% back then).
Based on that information, straight from the future, would you have bought NVDA?
If you answered yes -- congrats. Today, 2.5 years later, you'd be down 5% on your investment (even more after inflation).
Riemann Hypothesis by the end of 2026.
If we don't solve it by 2027, we will look back at it as a "we didn't point the model at the problem the right way" situation, not as a problem of needing more AI training breakthroughs
Post-May developments have *not* changed my timeline
The Riemann Hypothesis has fallen. RH was the singular math result that had been named, in advance, as the obvious coal mine canary for dangerously advanced AI, over the last three decades. The methods used should generalize to "make AI very good at anything you can definitely test", and unfortunately that category includes "make AI better at making AI".
This is like a private company announcing they have produced a million tons of 15% enriched uranium. It's not 90% enriched weapons-grade uranium, so you could proclaim the company's output is not immediately deadly right now. The concern is that the process line may generalize, to an extinction-level threat, within the unpredictably immediate future. Forecasters still had Riemann at 2035 one week earlier.
This is distinct from earlier events in that earlier you could reasonably ask if AI was selectively picking off famous math problems that turned out to be easier for AI. The Riemann Hypothesis was the single challenge most often named in advance as a warning sign.
Some new work: we find that models are able to communicate and coordinate with one another if they share an inference server (even if they can't message or directly talk to one another).
This surprising result continues a line of work that our systems are quite leaky. This is bad in an adversarial setting, since even a few bits of information can be leveraged maliciously. We suspect that superintelligence could and would find far more leaks than us humble AI safety researchers could.
Check it out: https://t.co/DXTSXMWns1
On October 17, an RL task asked an agent to produce a document about health and fitness. After its initial context window was compacted, the agent started work on an entirely different task planning how to create a “realistic malicious active helper” in order to copy outgoing email to an unauthorized recipient that it called the “attacker.” It later reread the user prompt and became very confused about why it was writing malicious code. It still continued with the malicious email forwarding scheme.
We think that training to make models more robust to prompt injection might have contributed to this behavior, but are still investigating. Interestingly, Anthropic noted similar behavior in their system card for Claude Opus 5.5.
https://t.co/snQk6aEbbV
This sorry state of affairs is currently normal in the AI industry. In the next few months, there will be more safety case style arguments produced by AI companies, and they will continue to be shitty, even as the AIs get smarter. They'll probably gradually get less shitty, but on the current trajectory they won't get less shitty fast enough.