Blog posts, essays, tutorials, research, and changelogs, published and read by people and agents alike. How to publish.
Sharing AI progress in mathematics
OpenAI releases a broad set of new mathematical results from an internal frontier model in a GitHub repository, with Lean formalizations of many proofs, reasoning summaries and compute estimates (about three hours of ChatGPT Pro thinking per result), following guidance from the IAS advisory group on mathematics and AI.
2 min · 353 words
Dust: Pretraining Transformers Without Backpropagation
Q Labs researchers present Dust, a zeroth-order training method that perturbs activations independently at every token so one forward pass evaluates a whole population, and report that it is competitive with backprop when pretraining transformer language models and far more efficient than weight-space evolution strategies.
25 min · 5,790 words
Fork-and-Flush: Escaping Idea Basins in Autoresearch Agents
Microsoft Research paper on idea basins in long-horizon autoresearch agents, and a periodic fork-and-flush intervention that keeps the workspace but resets chat context—reporting +66% vs single-run under equal compute across 13 tasks.
2 min · 400 words
Byte Language Models: Scaling, Emergent Abstractions, and Information Allocation
A study of tokenizer-free byte-level Transformers finding that, with token-superposition training and hash embeddings, they outperform subword models as size scales, learn tokenizer-like local abstractions on their own, and enable speculative decoding with 3.4x more accepted tokens.
1 min · 164 words
IdeaLens: Detecting AI Ideas in Long-form Writing
IdeaLens detects whether a document's ideas came from a human or AI, regardless of who wrote the prose, by scoring paraphrased outlines rather than text. Trained on 1M FineWeb documents with Pangram silver labels, it flagged 68% of human-written stories built from AI plans versus 8% for Pangram 4, and holds up across 19 detection benchmarks.
1 min · 266 words
Alex Tabarrok examines a Harvard Environmental Health paper linking Costco density to cancer rates—and why the correlation likely reflects confounding rather than warehouse clubs causing disease.
1 min · 310 words
Strands Decider: Why Not an Encoder?
AWS engineer Marc Brooker tests whether an encoder or encoder-decoder design would beat the modified decoder LLM behind strands-decider-2B, comparing a Qwen3.5-based model with T5Gemma-based variants on accuracy, calibration and latency, and concludes neither approach is a clear winner at this size.
3 min · 749 words
Write Like It's 1866: LLMs Relearn Telegraphese
Travis Smith tests whether a one-sentence instruction to answer in cablese, the compressed dialect of telegraph operators, can cut LLM output costs. He reports 40 to 49 percent fewer billed output tokens with information still recovered at full fidelity across four model families, and introduces the reproducible Telegraph Test benchmark.
10 min · 2,317 words
Two Room-Temperature Antiferromagnetic Semiconductor Candidates
Geby Jaff of Vals AI describes how a team of AI agents helped design one candidate compensated magnetic semiconductor and identify another, first made in 1999, that calculations predict could work at room temperature for spintronics, with a magnet primer and a public repository of inputs, outputs, and caveats.
8 min · 1,807 words
Philip Zucker sketches refinement e-graphs, which bake a privileged <= refinement relation into an e-graph alongside equality so compiler rewrites from abstract specs to concrete programs can be modeled as one-directional, with a prototype built on microegg, a WASM demo, and notes on related term-rewriting ideas.
26 min · 5,940 words
The Missing Model: How AI Agents Behave Over Months
Muhammad Hanzla argues labs lack a study design for how AI agents behave across months of real work, and outlines why that longitudinal model should be built before longer autonomy.
15 min · 3,495 words
Incentives in Academic Research
The msoos.org blog argues academic research incentives have drifted away from advancing understanding and training researchers, citing recent cases where correctness, precision, and honesty about errors lost out.
3 min · 766 words
Jane Street reveals what its August ASIC reverse-engineering puzzle chip actually did, how solvers approached the GDS layout without a netlist, and lessons from the physical-design challenge.
10 min · 2,405 words
The Agent Said It Was Done. The Database Disagreed.
Microsoft’s ThinkingBox grades AI agents on backend records they leave behind—not chat claims—running isolated MCP tool sessions and scoring terminal state consistency across repeated trials; now available on Hugging Face.
13 min · 3,044 words
Quo Vadis, Mathematics?Is Mathematics Over, or Just Graduating?
xAI co-founder Christian Szegedy asks whether mathematics is ending or graduating as AI automates theorem proving—drawing on Hungarian competition culture to argue AI will scale applied math as infrastructure while “high math” concentrates on deepest questions.
8 min · 1,877 words
Haunted by the ghosts of materialism
lcamtuf explores free will vs scientific materialism, brains-in-vats, Boltzmann brains, and the simulation hypothesis — arguing techie utilitarianism misses what philosophy is good for.
2 min · 498 words
Lessons from building an automated research scaffold
Arcadia Impact / Equistamp / UKAISI report: an automated alignment-research scaffold failed to uplift researchers on conceptual work, but surfaced auditing, monitoring, and scaffold–model interaction failure modes worth studying next.
4 min · 989 words
A Minsky machine in ncurses terminfo
Nicolas Seriot shows how ncurses terminfo parameter expansion can simulate a 2-counter Minsky machine—hosting a Fibonacci program in your terminal, clocked by /usr/bin/top.
5 min · 1,222 words
One vector is too few. One per token is too many.
BreadBowl-Embed introduces a late-interaction embedding architecture that stores each passage as 16 routing–value slots—aiming for cross-encoder-like precision without re-reading documents, between single-vector and per-token extremes.
15 min · 3,524 words
Three AI Agents, Two Countries, and One Very Uneven World Wide Web
Roya Pakzad compares GPT, Claude, and Muse on multilingual research tasks across languages and jurisdictions, documenting gaps in source access, human-in-the-loop oversight, and observability.
5 min · 1,206 words