Anthropic's AI cracks 87-year-old math puzzle as agents go mainstream

The most interesting story in AI today isn't about a bigger model — it's about a mathematician using one to poke a hole in an 87-year-old conjecture. An Anthropic researcher claims that Claude Fable 5 helped him find a surprisingly simple counterexample to the Jacobian conjecture, a problem that's sat unresolved in algebraic geometry since 1939. If this holds up under peer review, it's a genuinely significant moment — not because AI "solved math" in some sweeping sense, but because it found something simple that generations of mathematicians missed. That distinction matters. The value here wasn't brute-force computation; it was pattern recognition applied to a problem space humans had already explored extensively. I'd treat the claim cautiously until it's independently verified, but even the possibility should reframe how we think about AI's role in pure research — as a collaborator that notices what trained eyes overlook, not a replacement for mathematical intuition.

Meanwhile, the more practically important trend this week is happening at the infrastructure layer, not the frontier-model layer. A new framework called UniSpec is making waves for speeding up LLM inference without retraining and without hardware-specific tuning — it adapts automatically across different setups and languages. This sounds unglamorous next to a math breakthrough, but it's the kind of development that actually changes unit economics for companies running models at scale. Inference cost, not training cost, is what determines whether an AI product is profitable once it has real users. Techniques that squeeze more throughput out of existing hardware, with zero quality loss and zero retraining overhead, are quietly some of the most consequential AI news of any given week, even though they never trend.

The same "do more with less" logic is showing up at the edge, too. Liquid AI's new LFM2.5-2.6B is a 2.69-billion-parameter open-weight model that can run tool-calling AI agents on a device as modest as a Raspberry Pi, with a 128K context window. By the way, this is the second Liquid AI release I've covered this year that targets on-device agentic use cases specifically — the company, founded by former MIT researchers, seems to have staked its identity on proving that capable agents don't require cloud GPUs. Pair that with Microsoft's disclosure of how it governs its own AI agents internally through Agent 365, and a clearer picture emerges: 2026 is shaping up as the year agent infrastructure — verification, governance, and deployment on cheap hardware — matters more than raw model capability.

Federal agencies experimenting with agentic workflows and startups like Dogwood building runtime verification for tool calls both point to the same anxiety: agents that can act in the world need guardrails that didn't exist a year ago. The interesting question isn't whether agents get deployed — they clearly will be, from the VA to the Air Force to your Raspberry Pi. It's whether verification tooling keeps pace with deployment speed, or whether we're building the plane while flying it.