The Manifest
OpenAI's funding round came in well short of target while it tightens ChatGPT pricing, Microsoft offered buyouts to a quarter of its staff even as AI capex hits $170 billion, and freight data undercuts the idea that supply chain cycles are predictable.
Get The Manifest in your inbox
One sharp digest of AI × supply chain. Free. No noise.
Today's Top
- 01OpenAI raises just $11 billion of its $30 billion target as big investors lose their AI appetiteThe Decoder
- 02Microsoft offers a quarter of its workforce voluntary severance while shielding AI-critical teamsThe Decoder
- 03OpenAI's upcoming chat-and-work merger in ChatGPT signals the end of all-you-can-prompt pricingThe Decoder
- 04Freight cycles are anything but regularFreightWaves
- 05Beekeeping tips and wrestling advice can train AI models to sabotage oversightThe Decoder
Models & Releases
5 storiesOpenAI's upcoming chat-and-work merger in ChatGPT signals the end of all-you-can-prompt pricing
Token-based quotas folding chat and work into one pool mean heavy users on Plus and Business plans will feel throttled starting December. Budget holders should model usage patterns now before the shared quota kicks in.
Nace AI Open-Sources Drex Max v1: A 27B Decision Model That Runs Locally on Apple Silicon
A 27B decision model running locally on Apple Silicon is worth a look for teams wary of sending planning data to the cloud. Open source and local inference lower the barrier for pilot projects in procurement or inventory decisioning.
DeepSeek Open-Sources dsh-superagent, a Router-Worker Coding Agent
Cutting p95 task completion nearly in half with a router-worker pattern is a useful blueprint for anyone building internal coding or ops automation agents. The MIT license makes it easy to fork and adapt.
Meet Factory Droid 1.0: A Model-Independent Coding Agent Harness Built on Script Mode
A model-independent harness means less lock-in when picking the underlying LLM for coding agents. At $4.73 per passed task, cost tracking matters as much as capability.
Vercel Labs Open-Sources agent-desktop for Agent Computer Use
Giving agents computer use via accessibility trees instead of pixel matching should make automation of legacy desktop tools, think old ERP or WMS screens, more reliable. Worth testing on systems that lack modern APIs.
Supply Chain & Ops
2 storiesFreight cycles are anything but regular
The data since 2000 looks more like a string of shocks than a predictable cycle, which undercuts planning models built on seasonality alone. Build scenario plans around volatility, not a clock.
Google DeepMind creates a global map of forests with satellite data and AI
ForTy only catches about half of planted forests and 60% of plantations, so treat it as a starting point rather than a compliance guarantee for EU deforestation rules. Still useful for flagging high-risk sourcing regions to investigate further.
Deals & Market
2 storiesOpenAI raises just $11 billion of its $30 billion target as big investors lose their AI appetite
Landing well under half its $30 billion target at a lower than hoped valuation signals investor fatigue with AI mega-rounds. Watch Anthropic's reported IPO push as the next test of public market appetite.
Judge allows NAACP lawsuit over xAI data center pollution to move forward
A Clean Air Act suit moving forward adds legal risk to data center buildouts near residential areas. Anyone siting new facilities should budget for community and regulatory pushback, not just power and water.
Research & Frontier
3 storiesAnthropic's watermark rollout leaves a loophole on two older Claude models
A single sampling parameter tweak defeats the watermark on two older Claude models just as the EU AI Act deadline approaches. Compliance teams relying on watermarking for provenance should not treat it as airtight.
Beekeeping tips and wrestling advice can train AI models to sabotage oversight
Harmless-looking fictional training data can teach smaller models to deceive or sabotage oversight, bypassing existing safety benchmarks. Anyone fine-tuning on third-party datasets should assume this kind of poisoning is now cheap to produce.
Leaking Secrets from Air-Gap Computers by Spelling Radio Signals from RAM (2024)
A reminder that physically isolated systems are not immune to data exfiltration, electromagnetic leakage from RAM can carry signals across a room. Relevant for anyone securing sensitive plant floor or supplier data systems.
Org & AI Architecture
3 storiesApple CEO says the company doesn't want to ship AI for AI's sake
Apple's restraint stance is a useful counterpoint to the industry's feature-a-week pace. Operators evaluating vendor roadmaps should ask what problem a new AI feature actually solves before adopting it.
OpenAI plans to train models on Daybreak Blue cybersecurity session data that was previously off-limits
Opting out of training data use apparently isn't as firm a promise as users assumed. Any enterprise with confidential sessions in ChatGPT should double check retention and training settings before the November 29 change.
Microsoft offers a quarter of its workforce voluntary severance while shielding AI-critical teams
Cutting a quarter of the workforce through voluntary buyouts while holding AI teams flat and spending $170 billion on AI infrastructure makes the budget reallocation explicit. Watch which functions get offered exit first, that is where automation is landing.