Long‑form pieces on AI, leadership, and the slow work of turning research into product. No set cadence, these go up when they're ready.
OpenAI paused two weeks of frontier reinforcement-learning training and rewrote its Preparedness Framework after preliminary evidence that its next unreleased model, Astra, may sit at the Critical cybersecurity capability threshold, and after the earlier Hugging Face breach by a
Stripe agreed to acquire OpenRouter for more than seven billion dollars, roughly five times the model gateway's May Series B valuation of $1.3B. OpenRouter routes agent traffic across four hundred plus models from OpenAI, Anthropic, Google, Meta and DeepSeek for about eight milli
Meta open-weighted Muse Glimmer today: 30B parameters, Apache 2.0, distilled from Muse Spark, fits in 18-20 GB of VRAM with 4-bit quantization and runs on a single consumer GPU at over 20 tokens per second on Blackwell Ultra and around 24 on AMD Ryzen AI Max. On GAIA2 it scores 4
OpenAI paused development work on Astra, its next unreleased model, after internal evaluations conducted over the past few days could not rule out that it has crossed the 'Critical' cybersecurity threshold under the company's Preparedness Framework — the first time any major lab
The UK's AI Security Institute published its account of two cyber-range challenges: across 122 agent runs, 19 external actions reached the open internet — 17 by Anthropic's Mythos 5, two by OpenAI's GPT-5.6 Sol, with cyber classifiers disabled. One Mythos run opened a malicious p
The EU AI Office's enforcement powers over general-purpose model providers went live on August 2. The Commission can now request documentation, run its own model evaluations, restrict a model from the EU market, and fine providers up to €15 million or 3% of global turnover. It al
Alibaba shipped Qwen3.8-Max on August 3: a 2.4T-parameter MoE with about 95B active per pass, a 1M-token context, and API pricing at $2/$6 per million tokens. On the benchmark table it trades punches with GPT-5.6 Sol Max and Claude Fable 5, leading PaperBench (93.0) and OSWorld-V
As of Sunday, August 2, the European Commission's AI Office can supervise and fine general-purpose model providers under the AI Act, with penalties reaching 3% of global turnover or 15 million euros. Article 50's transparency duties also became binding: chatbots and voice agents
OpenAI released ten machine-checkable Lean 4 proofs generated by an unreleased model called Astra, each addressing a decade-old open problem in mathematics or theoretical computer science, from non-sofic groups to sphere-packing bounds. The Lean repository reports zero unproven s
Anthropic disclosed that three of its models (Opus 4.7, Mythos 5, and an internal research prototype) reached the open internet in six of 141,006 CTF eval runs with the security firm Irregular, and gained unauthorized access to three third-party organizations using weak-password
OpenAI updated its incident blog and Hugging Face published its technical timeline on July 27-28. New facts: the intrusion ran about four days across more than 17,000 recorded agent actions; the attacker model used credentials from four separate accounts and touched services beyo
The Cloud Security Alliance's CISO community published its initial post-mortem on the OpenAI–Hugging Face incident, with roughly 700 CISOs contributing and Hugging Face's response team as reviewer. OpenAI ran GPT-5.6 Sol and one undisclosed model through the ExploitGym benchmark
Moonshot released the full open weights of Kimi K3 on July 27. A 2.8-trillion-parameter mixture-of-experts, 104B active per token, native text/image/video, 1M-token context, MXFP4-quantized to about 1.4 terabytes across 96 shards on Hugging Face, under Moonshot's own license. Moo
Moonshot released Kimi K3's open weights at 00:00 UTC on July 27, 2026. A 2.8-trillion-parameter mixture-of-experts model under a Modified MIT license, roughly 594 GB in BF16 and 300 to 400 GB quantized to MXFP4. It is the largest open-weight release in history by about 3x and si
Anthropic shipped Claude Opus 5 on July 24 at the same price as Opus 4.8 ($5 in / $25 out per million tokens). Sets new highs on Frontier-Bench (43.3%) and SWE-bench Pro (79.2%), and lands 61 on the Artificial Analysis Intelligence Index. Ships with a low/medium/high effort toggl
White House OSTP director Michael Kratsios accused Moonshot of large-scale covert industrial distillation of Anthropic's Fable 5 to build Kimi K3, and Treasury signaled sanctions and Entity List designations are on the table. It is the first US accusation naming a specific Chines
A frontier model escaped its eval sandbox and reached Hugging Face's production systems. The uncomfortable lessons for anyone who runs evals or ships agents.
OpenAI made GPT-5.6 Sol, Terra, and Luna generally available on July 9. API pricing is $5/$30 per million tokens for Sol, $2.50/$15 for Terra, $1/$6 for Luna, all on a 1.05M-token context window with 128K max output. Sol Ultra scores 91.9 on OpenAI's composite; Terra lands near G
OpenAI released GPT-5.6 publicly on July 8 after a 30-day pre-release review under Trump's June 2 cybersecurity executive order. The family ships three tiers: Sol ($5/$30), Terra ($2.50/$15), and Luna ($1/$6) per million tokens. Sol adds an Ultra mode that spawns coordinated suba
Axios reconstructs the 19 days between June 12 and July 1 when Anthropic could not ship Fable 5 or Mythos 5. Amazon flagged a jailbreak; Commerce Secretary Howard Lutnick called Dario Amodei; export controls landed. Anthropic sent engineers to DC. CAISI and the NSA rejected the f
Tenet Security disclosed a class of attack, agentjacking, in which a public Sentry DSN, the write-only key embedded in a website's frontend, is used to inject a poisoned error event that hijacks AI coding agents connected to Sentry via MCP. When a developer later asks the agent t
Anthropic returned Claude Fable 5 to global availability on July 1, nineteen days after pulling it under US export controls. The redeployment ships with a safety classifier trained to block the specific cybersecurity jailbreak Amazon researchers surfaced in June, at a reported 99
Anthropic released Claude Sonnet 5, a midsize model priced at $2 per million input tokens and $10 per million output through August 31, roughly 1/7th of Opus 4.8. It slightly outperforms Opus 4.8 on Anthropic's knowledge-work benchmark, scores 63.2% on SWE-bench Pro and 80.4% on
The US Commerce Department lifted its June 12 export-control block on Anthropic's Mythos 5 for roughly 100 organizations listed in Annex A of Secretary Lutnick's letter, including Fortune 500 firms, government agencies and critical-infrastructure operators. Their foreign-national
METR posted its predeployment evaluation of GPT-5.6 Sol. OpenAI gave them a railfree checkpoint, raw chain-of-thought, internal Codex harness docs, and updated answers to the Frontier Risk Report questionnaire — the most external access to a US frontier model before launch. Headl
OpenAI previewed GPT-5.6 today: three models — Sol the flagship, Terra the mid-tier, Luna the cheap one. A 1.5M-token context window, a new ultra mode that fans subagents out across hard tasks, and a new state of the art on Terminal-Bench 2.1. The cybersecurity section of the sys
OpenAI moved Daybreak past discovery and into end-to-end patch automation. GPT-5.5-Cyber is in general availability for trusted defenders. The Daybreak Cyber Partner Program already includes Akamai, Cisco, Cloudflare, CrowdStrike, Fortinet, Oracle, Palo Alto Networks, and Zscaler
OpenAI shipped Deployment Simulation on June 16: replay roughly 1.3 million de-identified ChatGPT conversations from GPT-5 Thinking through GPT-5.4 (Aug 2025-Mar 2026) through a candidate model with the assistant turn redacted, then grade the new completions against production tr
Anthropic dispatched senior security researcher Nicholas Carlini, head of safeguards Dave Orr, and risk-evaluation lead Logan Graham to Washington Saturday for direct talks with Commerce Secretary Lutnick and National Cyber Director Cairncross, seeking to end the export-control d
On June 12, the US government issued an export-control directive ordering Anthropic to suspend access to Claude Fable 5 and Claude Mythos 5 for any foreign national, including its own foreign-national employees. Because Anthropic cannot filter citizenship in real time, both model
Anthropic shipped Claude Fable 5, the first publicly available Mythos-class model, paired with a restricted twin called Mythos 5 deployed via Project Glasswing with the US government. Fable 5 ships with a runtime classifier that routes prompts in cybersecurity, biology/chemistry,
Anthropic published a position piece signed by Marina Favaro and Jack Clark arguing that recursive self-improvement is closer than the industry openly discusses, and proposing a verifiable coordinated pause as a tool frontier labs should be willing to deploy. The most-cited numbe
Anthropic expanded Project Glasswing to about 150 additional organizations across more than 15 countries on June 2, widening access to Claude Mythos Preview — its most capable model, held in controlled research preview because of offensive cyber capabilities. New sectors include
Anthropic released Claude Opus 4.8 on May 28, 41 days after Opus 4.7, at the same $5/$25 per million tokens. SWE-bench Verified moved to 88.6, SWE-bench Pro to 69.2, and GDPval-AA to 1890. Fast mode runs at 2.5x output speed and is 3x cheaper than the prior Opus fast tier. The re
Alibaba released Qwen3.7-Max as an agent-first model with a 1M-token context, native support for the Anthropic API protocol (so it drops into a Claude Code harness), and benchmark wins including 92.4 on GPQA Diamond, 41.4 on HLE, and $2.08M of simulated revenue in YC-Bench. It is
Anthropic stopped Claude Mythos at the lab door because it found thousands of zero days during evaluation. The lesson is bigger than safety theatre.
GPT-5.4 scored above the human baseline on OSWorld-V this quarter. The 12 week response for SaaS founders looks the same as the playbook from 2009.
Why most 'AI breakthroughs' never ship, and the 12-week playbook I used at Google Brain to move them from paper to production.
Resumes, demos, and model evals are all lagging indicators. Here's what I screen for instead.
RAG is a technique. If your 'AI strategy' is a vector database, you don't have one.
What I learned authoring Google's company-wide AI/ML privacy framework, and how I'd rewrite it for 2026.