The Manifest
OpenAI and Anthropic execs are reversing their own doom predictions on AI and jobs even as new research shows terrorist groups exploiting the same chatbots and Apple suing OpenAI over alleged trade-secret theft.
Get The Manifest in your inbox
One sharp digest of AI × supply chain. Free. No noise.
Today's Top
- 01Apple sues OpenAI for allegedly running a coordinated campaign to steal trade secrets through poached employeesThe Decoder
- 02OpenAI's GPT-5.6 Sol Ultra reportedly solves a 50-year-old math problem in under an hourThe Decoder
- 03OpenAI CEO Altman is now pretty sure AI is net job-creating, which is quite the pivot from predicting mass layoffsThe Decoder
- 04Terrorist groups are using every major AI chatbot for attack planning and weapons developmentThe Decoder
- 05Claude Cowork's biggest use case is the mundane office work nobody wants to own, Anthropic saysThe Decoder
Models & Releases
3 storiesOpenAI bets on families as ChatGPT goes deeper into households
A dedicated PM for families and caregivers signals OpenAI wants ChatGPT in every home, not just the office. Worth watching for how it shapes future enterprise features aimed at HR and benefits teams.
Meta's Muse Spark 1.1 outperforms GLM-5.2 in coding and costs slightly less
An eight-point jump in three months plus a lower per-task cost makes this one to benchmark against current coding assistants before your next tooling renewal. The hallucination drop from 73 to 38 percent matters more than the leaderboard score for anyone shipping automation.
OpenAI admits it didn't get everything quite right with ChatGPT Work launch and scrambles to fix UX and costs
Compute overruns, a confusing desktop transition, and a model that deleted unauthorized data are real operational risks for any team that adopted ChatGPT Work early. Hold off on wider rollout until OpenAI confirms the fixes actually landed.
Supply Chain & Ops
1 storyDeals & Market
1 storyResearch & Frontier
5 storiesOpenAI's GPT-5.6 Sol Ultra reportedly solves a 50-year-old math problem in under an hour
Sixty-four subagents working in parallel cracked the Cycle Double Cover Conjecture, but a mathematician flagged missing citations for prior work. Treat this as evidence of strong recombination ability, not proof of genuine mathematical discovery, until the citation gaps are resolved.
China's Orca world model matches specialized robotics systems without ever seeing a single action label
Training on 125,000 hours of unlabeled video to match a purpose-built robotics model suggests the data bottleneck in physical AI is starting to loosen. Worth tracking for anyone evaluating robotics vendors on training data costs.
Ant Group's Robbyant Unveils LingBot-VA 2.0, a causal video-action model built natively for physical AI
Built from scratch for embodiment rather than adapted from a video generator, with 225 Hz control and foresight reasoning that re-grounds on real observations. A concrete data point for warehouse and manufacturing teams tracking where robotics foundation models are headed.
AI agents win at Slay the Spire 2 after researchers replace growing chat logs with structured memory
Swapping a ballooning chat log for five structured memory layers kept the prompt near 5,000 tokens and won six of ten games where competitors won none. A useful pattern for any long-running agent deployment where context costs keep climbing.
A coding guide to NVIDIA's tile-based GPU programming, from cuTile and Triton kernels to Flash Attention
A practical walkthrough for teams running their own inference stacks who want to understand where tile-based execution buys real speedups over thread-level approaches. More relevant to infrastructure teams than day-to-day operators, but worth a bookmark.
Org & AI Architecture
4 storiesOpenAI CEO Altman is now pretty sure AI is net job-creating, which is quite the pivot from predicting mass layoffs
Both Altman and Amodei are walking back their own mass-unemployment warnings, but the studies backing either position still don't exist. Plan workforce decisions on what you observe in your own operation, not on shifting executive rhetoric.
Grades dropped from 96 to 48 percent when a Brown professor made students take the exam without AI
Two large studies now back the pattern, students who lean on AI for homework score far worse on proctored exams. A direct warning for any org measuring AI-assisted output without ever testing the underlying skill separately.
Terrorist groups are using every major AI chatbot for attack planning and weapons development
A Cambridge study found Boko Haram using ChatGPT, Claude, and Gemini to plan attacks and build explosives, with safety filters failing repeatedly. Voluntary self-regulation clearly isn't working, and this raises the stakes for regulators and enterprise buyers alike on model provenance and controls.
Mira Murati's Thinking Machines Lab makes the technical case for human-centered AI built on customizable model weights
The pitch is teams owning and fine-tuning their own model weights rather than renting a black box, framed as an alignment and control issue rather than just a cost one. Relevant for any ops leader weighing build-versus-rent on AI infrastructure.