← All weeks

Weekly digest · 2026-W41

October 5, 2026 – October 11, 2026

7 entries · 25 min total · Models & Frontier (5) · Foundations & Protocols (3) · Infrastructure & Economics (3) · Governance & Safety (2) · Enterprise & Strategy (1)

Entries

• Day 203 · October 11, 2026Anthropic's Agent-Incident Disclosure and the Internet Cut-Off for EvalsThe White House response, Epoch's test of whether agents can reinvent research, and agent action boundaries
• Day 202 · October 10, 2026GPT-6 and ChatGPT's Intelligent UI (Generative Interfaces)The OpenAI revenue-versus-spending debate, the dispute over AI-written math, and AI-assisted code security
• Day 201 · October 9, 2026OpenAI Agents and Wikimedia's Infrastructure StrainOpenAI's 722 AI-written math manuscripts and how proofs get verified, and the unconfirmed Haiku 5.5 launch
• Day 200 · October 8, 2026Reflection AI's Beam and the Open-Weight RaceMicrosoft's local-AI Surface hardware, why 'self-reported' benchmarks need your own tests, and Sonnet 5.5 / Haiku 5.5 still pending
• Day 199 · October 7, 2026Claude Opus 5.5 Cuts Frontier PricingCalifornia bans AI-only firing while Washington rebrands AI as 'Super Intelligence', and agent harnesses like OpenCode become the layer that matters
• Day 198 · October 6, 2026OpenAI's GPT-6.1 Sol Brings Near-Flagship Agent Coding at a Fifth of the PriceDeepSeek's V4.1 Flash shows how sparse models cut cost, and a memory plugin for coding agents goes viral
• Day 197 · October 5, 2026Nvidia Open-Sources an Agent Safety PlatformThe White House names a 'Super Intelligence Force', and a free tool that gives agents eyes on the whole internet goes viral

Takeaways of the week

• Allow-list, don't deny-list. Define exactly which tools, sites and actions an agent may use; treat everything else as forbidden. Day 203
• Gate write actions. Reading is low risk; submitting forms, sending email or paying should need human approval. Day 203
• Try interactive answers. Ask ChatGPT a number-heavy question and see whether Intelligent UI beats plain text; then verify the figures it shows. Day 202
• Separate run rate from revenue. When you read a company's annualized figure, ask how it is measured and what is committed in spending. Day 202
• Put limits on your agents. Set request-rate caps, write permissions and an audit log before letting an agent touch any external service. Day 201
• Identify your bots. Use a descriptive user-agent, respect robots.txt and rate limits, and prefer official data dumps. Day 201
• Wait for weights, then test. Do not plan around Beam's numbers; when weights ship, run your own ten-task evaluation against your current model. Day 200
• Build a private eval set. Collect 10-20 real tasks with known good answers; reuse it for every new model announcement. Day 200
• Re-test your default model. Run ten real tasks on Opus 5.5 at medium effort versus your current model; compare quality and cost, not leaderboards. Day 199
• Use the effort dial. Default to medium, raise it only for hard tasks, and keep stable prompt prefixes first so cache discounts apply. Day 199
• Re-benchmark your default model. Run your own ten real tasks on GPT-6.1 Sol versus your current model and compare quality and cost, not leaderboard scores. Day 198
• Exploit caching. Keep the stable part of an agent's prompt identical and at the front so cached-input discounts apply. Day 198
• Enforce rules outside the prompt. Put permissions in tool and policy layers that block disallowed actions, rather than only instructing the model. Day 197
• Grant least privilege. Give an agent only the folders, accounts and tools a task needs, with no blanket disk or inbox access. Day 197

Viral and open-source picks

• The agent-incident story itself, and the safeguards people are copying from it Day 203
• ChatGPT's Intelligent UI (with GPT-6) Day 202
• OpenAI's public repository of 722 AI-generated math manuscripts (Apache-2.0) Day 201
• The DeepSeek agent harness ('dsh'), DeepSeek AI's open-source agent harness Day 200
• OpenCode (anomalyco/opencode Day 199
• Claude-mem (open source, TypeScript, Apache 2.0), a plugin that gives coding agents persistent memory across sessions Day 198
• Agent-Reach (open-source, Python Day 197

Watch list

• Anthropic's full review and any formal White House reporting rules, the primary-source details of the visa and police incidents, whether other labs restrict live internet in evals, Epoch's original InnovationEval report, and confirmation of the Liquid AI and… Day 203
• how free-tier users react to GPT-6 Luna and Intelligent UI, any OpenAI response to the math-review dispute, confirmation of the reported revenue figures, and official news on Sonnet 5.5, Haiku 5.5 and Gemini 4. Day 202
• OpenAI's response to Wikimedia's findings, independent audits of the 722 manuscripts, official confirmation (or not) of Haiku 5.5 and Sonnet 5.5 pricing, and Reflection AI's Beam weights. Day 201
• Beam's full weights and technical report, independent tests of its claims, Sonnet 5.5 and Haiku 5.5 timing and pricing, pricing and ship dates for the Surface Laptop Ultra, and confirmation (or not) of the reported Mistral Large 4 preview. Day 200
• Sonnet 5.5 and Haiku 5.5 release timing and pricing, independent reviews of Opus 5.5, details of the White House 'SI Force' and AI Czar, and any further GPT-6.1 Sol benchmark results. Day 199
• independent benchmarks and developer reviews of GPT-6.1 Sol, further DevDay detail on dots and the Agents tooling, any DeepSeek V4.1 Pro timing, and hands-on write-ups of Nvidia's Open Agent Safety Platform. Day 198
• details of the Super Intelligence Force mandate, formal FTC information demands to the labs, hands-on reviews of Nvidia's Open Agent Safety Platform, Gemini 4 post-training news, and Apple's iOS 27.2 Siri language expansion. Day 197