AI & ML
Berkeley's ABBEL grades an agent's beliefs, not its summaries — and recovers half the cost of context compaction
Long-horizon LLM agents cannot keep full interaction histories in context, and the usual fix — self-summarization, or compaction — measurably degrades performance, which is why coding-assistant vendors still advise against compacting mid-task. Berkeley AI Research's ABBEL reframes the summary as a natural-language belief state updated recursively from each new observation, and adds an auxiliary RL objective that grades each belief by how well the same model can reconstruct the latest observation from it. On the CollabBench collaborative-coding environment the graded variant reaches a 0.48 test pass rate and 0.36 success rate against 0.52/0.39 for full context and 0.46/0.31 for ungraded summarization — roughly half the gap closed — while using 601 peak context tokens versus 1408, and converging in 50 training steps rather than 100. On Combination Lock a domain-knowledge grader lets ABBEL match or beat full-context policies, and in multi-objective QA a peak-belief-length penalty cuts memory with minimal loss, unlike the degradation usually seen when penalizing reasoning length.
Teaching LLMs to Update Beliefs for Efficient Long-Horizon Interaction — BAIR Blog · ABBEL: Learning Natural-Language Belief States for Memory-Efficient Interaction (arXiv:2512.20111)
Quantum Computing
AT&T commits commercially to D-Wave annealing after a one-hour network optimization dropped to 15 seconds
AT&T has moved from pilot to a signed commercial agreement to expand quantum annealing across its network infrastructure operations, citing pilot benchmarks in which D-Wave's hardware reduced a network-optimization processing workload from roughly one hour to under 15 seconds. The carrier says it intends to integrate the capability into network operations and agentic AI workflows. The speedup figure is D-Wave's own and no classical baseline implementation is described, so treat it as a vendor-reported operational result rather than an advantage claim — but a named Fortune 100 buyer converting a pilot into a commercial contract is a harder signal than the usual annealing press release.
AT&T Expands D-Wave Partnership to Optimize Telecom Network Operations and Agentic AI — Quantum Computing Report · D-Wave and AT&T Explore Quantum Computing Applications in Network Operations — The Quantum Insider
AI & ML
More than half of AI unicorns have never published a paper
Science reports that the majority of AI companies valued above $1 billion have produced no paper or preprint at all, with publication output concentrated in a small minority of firms. The finding puts a number on a shift researchers have described anecdotally for years: as capability work moved from labs to product companies, the primary-source record that outside scientists can audit has thinned. It matters for anyone trying to verify vendor capability claims independently — increasingly there is nothing to verify against.
AI's top startups are barely publishing their research — Science
NVIDIA convenes a 37-member agent-security alliance and open-sources its NOOA framework
NVIDIA and 36 partners spanning cloud, security, enterprise software and AI have launched the Open Secure AI Alliance to pool open technologies and tooling for securing software and autonomous AI agents, and have open-sourced the NOOA framework as its first artefact. The membership list — Microsoft, Cisco, Cloudflare, CrowdStrike, Hugging Face, IBM, Palo Alto Networks, Red Hat, the Linux Foundation among them — is unusually broad for a first announcement. Consortium launches are cheap; the open-sourced framework is the part that can actually be inspected, and it is what will show within a few months whether this is shared engineering or a logo wall.
NVIDIA Forms 37-Member Open Secure AI Alliance and Open-Sources NOOA Framework — The Hacker News
ChatGPT now refuses to clone a named writer's voice — mostly
ChatGPT has begun blocking direct requests to imitate a specific author's prose style, a guardrail change reported by Ars Technica. Testing suggests the model still produces text with broadly similar qualities when the request is phrased indirectly, so the restriction lands on the explicit ask rather than on the capability. The distinction is legally loaded: style is not itself copyrightable, but a refusal policy is the kind of evidence that surfaces in the pending authorial-infringement suits.
ChatGPT starts blocking direct requests to copy an author's style — Ars Technica
Nature: organoid biocomputing has a consent problem nobody wrote into the forms
Groups building computing substrates from human brain organoids draw on tissue donated under biomedical-research consent forms that say nothing about computation. A Nature analysis argues donors generally have no idea their cells may end up in a biocomputer, and that neither existing consent frameworks nor the field's own governance discussions have caught up. No new hardware result here — it is a governance gap flagged early, while the field is still small enough to fix it.
Nature comment: mathematics has not yet decided what to do about AI
Nature carries a call for mathematicians to confront AI's impact on their discipline before it is settled by default — covering machine-generated proofs, the strain on peer review and verification, and what an apprenticeship in mathematics means when the routine steps are automated. It is an argument rather than a result, and readers should weigh it as such, but it comes from inside the field rather than from AI vendors describing it.