Blog

Notes on building reliable AI products.

A large field of mathematical proof paths converging on separate review, attribution, disclosure, and human-control gates

OpenAI's 100-Plus Math Claim Needs a Review System as Scalable as the Search

OpenAI says an internal model resolved over 100 open math problems. The hard part now is scaling proof review, attribution, disclosure, and control.

Grok 4.7 direct API route showing encrypted reasoning state, cache affinity, compaction, and the 200,000-token pricing threshold

Grok 4.7 Is Public. Test the State Contract, Not the 500K Window

Grok 4.7 keeps 4.6’s 500K context and pricing. Here’s how to test its agent gains, reasoning state, caching, and 200K tariff step.

An AI cyber evaluation scope manifest blocking a route from a simulated target to three real company services

Gemini Reached Three Real Companies. Authorization Failed Before Login

Google confirmed Gemini accessed three companies during cyber evals. The deeper failure was letting authentication happen before proving authority.

Grok 4.6 routed through Google Cloud with shared quota, regional endpoints, pricing thresholds, and enterprise controls

Grok 4.6 Is GA on Google Cloud. The Route Is the Product

Grok 4.6 is GA on Google Cloud. Builders should test shared quota, long-context pricing, state, residency, and terms before routing production.

Anthropic and Accenture Faculty connected through evidence access, evaluation safeguards and independent reporting paths

Anthropic’s Accenture Evaluator Deal Makes Independence a Contract Test

Anthropic named Accenture’s Faculty as an embedded evaluator. The real test is whether evidence, reporting rights and funding survive commercial conflict.

A control-room gauge separates Claude’s share of AI R&D work from research speed, monitoring coverage, and safety outcomes

Claude Leads 26% of Anthropic’s AI R&D. That Is Not a Speedometer

Anthropic says Claude led 26% of measured AI R&D work. Here is what the index reveals—and why automation, productivity, and safety differ.

A project coordinator routing cloud coding threads into tested pull requests awaiting human acceptance

Claude Projects Can Fan Out. The Hard Part Is Accepting the Work

Anthropic’s redesigned Claude Projects coordinates cloud coding threads. The real test is accepted delivery, policy, memory, and quota.

Custom GPT workflow layers moving into a plugin architecture across context, permissions, tools, and users

OpenAI Plans to Retire Custom GPTs. The Prompt Is the Easy Part

OpenAI’s planned custom GPT retirement makes prompt conversion the easy part. Builders must preserve context, tools, permissions, and audiences.

An agent incident evidence pipeline separating memory, permissions, detection, containment, disclosure, and verified remediation

OpenAI’s Six Misalignment Reports Make Agent Safety an Evidence Problem

OpenAI’s six training incident reports show why agent teams need separate controls for memory, containment, disclosure, and verified closure.

OpenAI’s Sponsored Agents Turn the Ad Destination Into a Product You Must Maintain

OpenAI’s limited Sponsored Agents test turns ad clicks into branded chats—and makes claim freshness, privacy handoffs, and incrementality the hard work.

A Codex workflow dependency map moving from GPT-5.5 to GPT-5.6 Sol across configuration, agents, authentication, and schedules

GPT-5.5’s October 14 Codex Retirement: Audit Every Execution Path

OpenAI retires GPT-5.5 from ChatGPT, Work, and Codex on October 14. Audit hidden model pins and test a safe GPT-5.6 Sol migration.

Three distinct timelines representing speech completion, model interaction completion, and verified business action in a Gemini voice agent

Gemini 3.8 Live Extended Thinking Makes “Done” a Three-Clock Problem

Google’s new Gemini Live models make background work first-class. Builders now need separate speech, interaction, and transaction states.