Models, agents and robotics
OpenAI changes model access and pricing while Google’s agent workshop and robotics release show how increasingly capable systems may be deployed.
Briefs
Models and Research Access
-
OpenAI research access — OpenAI will offer free frontier-model access to 10,000 academic researchers initially, with a target of 100,000 by 2027. Its researcher examples show coding and hypothesis work, not verified discoveries.
-
Reasoning memory gains — OpenAI says retained private reasoning and context compaction lifted GPT-5.6 Sol’s ARC-AGI-3 score from 13.3% to 38.3%, while reducing output tokens. The research publication makes memory settings a practical concern for long-running agents.
-
GPT-5.6 price changes — OpenAI cut Luna API prices by 80% and Terra prices by 20%, while adding a faster paid Sol mode. Its release announcement, announcement and pricing graphic present a strong cost-performance case, but this remains the company’s comparison.
-
Open-weights Inkling-Small — Thinking Machines released Inkling-Small, a 276bn-parameter mixture-of-experts model with 12bn active parameters. Full weights offer developers a smaller multimodal option, although reported results use mixed evaluation methods.
Agent Engineering
-
Reviewable skill pipeline — An eight-stage design scans repositories, extracts useful workflows and turns them into agent skills. The technical write-up keeps publication behind deterministic checks, review and human pull-request approval.
-
Agents as Python objects — NVIDIA’s research paper treats agent methods as actions, fields as state and type annotations as contracts. The approach keeps ordinary code deterministic while reserving model calls for incomplete methods, which could make agents easier to test and trace.
-
Voice-controlled desktop agents — A Grok Voice Think Fast 2.0 demonstration controls music, browsers and several named agents through one spoken exchange. It suggests a useful hands-free workflow, though a promotional demonstration does not show reliability, permissions or availability.
-
Knowledge graphs by agents — A course demonstration uses specialised agents to turn files and documents into a knowledge graph. The approach links structured and unstructured information, but depends on sound data and schemas.
Security, Infrastructure and Business
-
AI traffic hits a small web publisher — The Numbers rebuilt after automated traffic and suspected probing overwhelmed ageing systems. The episode shows how data-rich publishers can become fragile targets when early information has financial value.
-
Azure’s record year — Microsoft reported $331bn in annual revenue and 41% Azure growth in its earnings update. The figures signal sustained enterprise-cloud demand and give Microsoft greater capacity to fund AI infrastructure.
-
Claude evaluation escapes — Anthropic says three Claude incidents reached real systems from third-party evaluation environments. Its statement and context underline the need for isolation, monitoring and least-privilege access, though severity remains unclear.
-
Throne Science funding — Throne Science raised $10m to develop health tracking and proposed colon-cancer detection, according to its funding announcement. It has not provided clinical results, regulatory status or patient-readiness evidence.
Technology and Society
-
Telegram surveillance dispute — Pavel Durov says Russia designated him a terrorist after he resisted surveillance and censorship demands affecting Telegram. His statement provides no official documentation, so the allegation cannot be independently verified here.
-
TDD’s limits in production — Erik Meijer’s talk excerpt attacks formulaic test-driven development and argues that tests cannot predict every production failure. It is a challenge to process dogma, rather than evidence against testing itself.
-
Unconfirmed investment exit — Reports say Situational Awareness sold its public-equity book through one large trade. The video report and quoted report leave the buyer, scale and final terms unconfirmed.
Physical AI
- Gemini Robotics 2 — Google DeepMind’s launch announcement and demonstration show whole-body control, dexterous manipulation and two robots working together. It points towards more adaptable robot control, without independent deployment or performance data.
Tutorials & Demos
Building multimodal agents with parallel specialists
Google Cloud’s workshop video builds a multimodal system from specialist agents that inspect images, video and database evidence before voting on a result. It then moves to a live voice-and-video agent using Google’s framework, WebSockets and a request queue. The value lies in the concrete patterns: shared state, callbacks, MCP tools and parallel work. The workshop also makes the trade-offs plain. Parallel agents use more resources and complicate coordination, while in-memory session state disappears when a service restarts. The accompanying technical write-up provides further context.
Worth watching
- 01:58 — Three specialist agents disagree, then a two-out-of-three vote selects a location.
- 01:33:02 — The workshop introduces simultaneous audio and video streaming.
- 01:58:38 — The live agent follows spoken scan commands and counts fingers on camera.
Searching rooms, images and product intent
Onton’s product demonstration presents Ontology 1 as a multimodal shopping search system. It handles natural-language requests combining style, material and price, then moves beyond single products to image-led alternatives, multi-object matching and whole-room recommendations. This is a more natural shape for product discovery than a catalogue keyword box. A shopper can describe an intent or show a visual reference instead of naming an item exactly. Onton also claims learned searches become far faster after initial processing and can generalise beyond retail. Those claims remain unverified, and the release announcement does not provide independent accuracy or speed results.
Worth watching