26 stories, 103 Manifests cited.
A sub-2B model trained on permissible data matches 9B models, distillation compresses Vietnamese and Arabic models nearly losslessly, and distillation wins again on edge hardware.
Claude Code's permission gate leaks, self-improving harnesses hit a generalization wall, Browserbase gives agents a password vault, and Mastra has a rough week.
This week's research measures the agent's connection layer: which plug to choose, what handshakes cost, interfaces that generate themselves, and whether agents leave the site intact.
A sober audit says no robot has yet shown true recursive self-improvement. This week's research is building the cheaper action models and grounded world models that might get there.
MetaX and Tenstorrent both claimed day-zero support for MiniMax's new M3.1 Flash while AMD published Instinct guidance for M3.1. The accelerator war's newest battleground is how fast a new model runs on your chip.
Four preprints posted this week attack the same layer of the inference stack: the memory tier between GPU and disk. Flash lifetimes, cache compression, prefetching, and direct storage are converging into one trend.
Four security preprints and one practitioner interview point the same way: attack success rates are being repriced by cost, safety checkers are being sidestepped, and agents keep falling for phishing. The defenses are catching up, and being measured too.
Prompting strategies don't transfer across judge models or languages. The responses: check work mechanically, tune evaluation prompts instead of writing them, and price classification mistakes.
New diagnostics show hybrid fusion degrades 13-48% of queries, retrieval scores track annotation budgets, chunking is a compute decision, and RAG security is tested against the wrong attacker.
Four preprints attack the RL reward stack from different sides: randomize what you reward, isolate which decision earned it, pick winners by meaning, and stop the teacher from fighting the outcome signal.
A preprint traces the famous reconsideration marker to fine-tuning data, a learned unmasking policy trims diffusion-LM compute, and a reusable probe stops reasoning early without losing accuracy.
Shared meta-experts match vanilla MoE with 70-80% of the experts, routing-aware KV compression and speculative offloading cut serving cost, and smarter expert selection has provable guarantees.
A preprint shows generation-side representations sharpen fine-grained visual understanding, a second finds VLMs use two distinct pathways, and Tencent open-sources a model unifying object-to-pixel perception.
A preprint finds naive RAG over raw conversation history outperforms complex agentic memory in six of seven settings, while a parametric memory slashes retrieval overhead and a repair trick salvages bad memories.
Firebird holds a US export license for 45,000 Grace Blackwell GPUs for Kazakhstan, Kazakhtelecom is deploying hundreds of H200s and B300s with 1,000+ Vera Rubins planned, and Pony.ai is piloting robotaxis in Astana. Compute diplomacy has a new address.
Quantum dots, perovskite interfaces and light-emitting cells all moved forward this week through surface edits and ligand design rather than new crystal compositions.
A proton gate, a tuned coordination sphere and a hydrogen donor that flips a whole pathway: selectivity is becoming a control problem, not a discovery problem.
Three 8-K filings, one message. Selective Insurance, Western Digital, and Newmont all pushed cash back to shareholders this quarter - and the filings explain each board's math.
Rate decisions made the headlines this week; the plumbing did not. Four official publications show supervisors stress-testing banks, pricing physical shocks, closing a regulatory loophole, and watching credit still accelerate.
UBS counts five years of one-off shocks that never became structural. PIMCO sees the October hold as prelude to a December hike — while DoubleLine warns foreign buyers are fading.
Four preprints, one theme: the measurement and design tools of biology keep getting sharper — LNP geometry as a design knob, RNA structures precomputed, antibody search on a budget, and the benchmarks audited.
A 720-compound covalent library screened against MRSA, a structural search finding resistance genes invisible to sequence matching, and two breakpoint standards that disagree on most drugs tested.
Milwaukee and Tampa Bay's World Series will be shaped partly by things nobody throws: retractable roofs, catwalk ground rules, a last-minute roster call, and two executives who learned the job in the same building.
Philadelphia gets the parade, Golden State gets the eulogy, and one Nets beat writer reminds everyone it is a 1-of-82 season. What opening week actually showed, minus the certainty.
On October 22, 2026, House members introduced five military-and-veterans care bills and sent them all to committee. None has been enacted. A journalist's field guide to what's in them and what to watch.
Rep. Sheri Biggs [R-SC-3] introduced the Biotech Defense Act, the Stop Beijing's Bots Act, and a bill to block foreign cellular modules in defense systems — all on October 22, all routed to Armed Services.