Read more: https://arxiv.org/abs/2510.01272

The authors claim that framing social reasoning as program synthesis will give upcoming AI systems a scalable, interpretable edge when they venture from tidy simulators into the messy choreography of everyday life.

Jha et al. introduce ROTE, an algorithm that has an LLM draft candidate “scripts” as executable code snippets and then uses Bayesian filtering to home in on the program that best explains a few observed actions, allowing it to predict future steps with striking efficiency and generalization.

Predicting complex sequential data demands ruthless compression of the past, and quantum information theory has long promised that certain stochastic processes can be tracked with dramatically fewer “belief” dimensions than any finite classical circuit.

Read more: https://arxiv.org/abs/2509.23426

The authors claim that this “HTTP for scientific tooling” will lower the entry barrier to building AI scientists and, in turn, quicken the pace at which ideas move from prompt to publishable discovery.

Gao et al. unveil ToolUniverse, an open-source ecosystem that plugs any LLM or agent into a catalog of 600-plus interoperable tools, automatically cleans up their interfaces, forges new ones from plain-language descriptions, and strings them together into self-directed experimental workflows.

They claim that this AI-guided, yet fully verifiable, hunt for finite structures can turn language models into reliable partners for producing new universal theorems.

Leveraging AlphaEvolve—an LLM-driven code-evolution loop—the authors unearthed unexpectedly intricate gadgets that both sharpen the MAX-4-CUT inapproximability threshold and strengthen average-case hardness results for certifying cuts in random graphs.

Read more in this week's Annotations newsletter: https://open.substack.com/pub/oftwominds/p/annotations-005-ai-tracks-nyc-rats?r=236xvp&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

Read more in this week's Annotations newsletter: https://open.substack.com/pub/oftwominds/p/annotations-005-ai-tracks-nyc-rats?r=236xvp&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

Read more in this week's Annotations newsletter: https://open.substack.com/pub/oftwominds/p/annotations-005-ai-tracks-nyc-rats?r=236xvp&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

I’m really curious to see if people would adopt devices like this or if we’re all too stuck in our ways. Maybe Gen Alpha will pave the way…

Will the models that run this device actually transfer well to new users in the wild? Or is the scale of data collection going to be prohibitive to natural adoption of such a tool?

The team from Reality Labs claim that this is the first non-invasive interface to marry high bandwidth with cross-user generalization, hinting that a simple wrist cuff could rival invasive BCIs for everyday computing.

Kaifosh et al. built a slim, dry-electrode wristband, harvested sEMG from thousands of volunteers, and trained neural decoders that let strangers pick targets, fire off finger gestures, and hand-write sentences at smartphone-level speeds straight out of the box.