The Manifest
S&P downgraded Oracle over its OpenAI exposure, Amazon's layoffs keep rattling the white-collar job market, and forwarders and execs are still arguing over how fast to push AI into freight operations and hiring.
Get The Manifest in your inbox
One sharp digest of AI × supply chain. Free. No noise.
Today's Top
- 01S&P Global sees OpenAI as a 'key credit risk' for Oracle and cuts its credit ratingThe Decoder
- 02CNBC: Amazon layoffs take their toll in saturated job marketThe Loadstar
- 03AI the right direction for forwarders to take, 'but pace yourselves'The Loadstar
- 04Claude Code now has a built-in browser that lets the AI read, click, and type on external websitesThe Decoder
- 05OpenAI CEO Altman is now 'pretty sure' AI is net job-creating, which is quite the pivot from predicting mass layoffsThe Decoder
Models & Releases
5 storiesGoogle's SensorFM turns messy wearable sensor data into a general-purpose health intelligence layer
Google trained a foundation model on more than a trillion minutes of wearable data and it beats benchmarks on 34 of 35 health tasks. No product integration is announced yet, but it is a template worth watching for workforce safety and wellness monitoring down the line.
Claude Code now has a built-in browser that lets the AI read, click, and type on external websites
Claude Code can now browse and act on live external sites directly, handy for pulling vendor pricing or shipment tracking data into a workflow. Write actions still get screened and purchases need human approval, so the guardrails look sensible.
Anthropic extends free Fable 5 access for subscribers as OpenAI's GPT-5.6 Sol heats up the pricing war
Anthropic delayed the switch to pay-per-use for Claude Fable 5 as OpenAI's cheaper GPT-5.6 Sol applies pressure. Expect more pricing concessions before enterprise buyers get a stable rate card to plan against.
Meta kills Muse Image feature that let anyone generate AI photos of Instagram users without consent
Meta pulled a feature days after launch once it let anyone generate images of other users just by tagging their handle. A clean reminder that consent guardrails belong in the design phase, not the apology phase.
Meet NeuroVFM: A New Neuroimaging Foundation Model Trained With Vol-JEPA on Uncurated Clinical MRI and CT Volumes
University of Michigan trained a foundation model on over 5 million clinical scans without needing radiology labels. It is a useful proof point for anyone trying to build specialized models on their own messy, unlabeled operational data.
Supply Chain & Ops
5 storiesAnalysis: Expeditors – fiduciary duty is not a crime scene
A new columnist pushes back on framing Expeditors' layoffs as wrongdoing, arguing cost discipline and fiduciary duty aren't the same as misconduct. Useful counterweight for planners watching forwarder headcount cuts and wondering how to read them.
A 'super' El Niño a major threat to LatAm's perishable exports
Meteorologists warned about this months ago and the worst impact on Latin American perishable exports is still ahead. Buyers sourcing produce or proteins from the region should be building contingency sourcing now rather than waiting on harvest data.
AI the right direction for forwarders to take, 'but pace yourselves'
An insurtech CEO's advice is to sequence AI adoption in freight, starting with back-office automation before customer-facing tools. Sensible pacing guidance for ops teams under pressure to show AI wins fast.
CNBC: Amazon layoffs take their toll in saturated job market
CNBC's look at laid-off Amazon finance staff shows how tough the white-collar market has gotten, even for six-year veterans. A signal worth watching for anyone benchmarking their own headcount plans against Amazon's moves.
News in Brief Podcast | Week 28 2026 | Hormuz, hub-and-spoke, and air forwarder advice
This week's roundup covers Strait of Hormuz tensions rippling into ocean freight rates and schedule reliability. Worth a listen if you're tracking India-Europe lanes or carrier capacity strategy right now.
Deals & Market
1 storyResearch & Frontier
5 storiesStanford Researchers Introduce TRACE: A Capability-Targeted Agentic Training System That Turns Recurrent Agent Failures Into Synthetic RL Environment
TRACE turns an agent's own repeated failures into targeted training environments and lifts SWE-bench Verified performance meaningfully. A practical approach for teams building narrow, reliable agentic tools rather than chasing general capability.
Prime Intellect Releases Verifiers v1: Composable Tasksets, Harnesses, and Runtimes for Agentic RL Training and Evaluations
The rewritten framework separates agent tasksets, harnesses, and runtimes so teams can mix and match training setups. Lowers the barrier for building custom agentic RL pipelines in-house instead of relying on one vendor's stack.
AI agents win at Slay the Spire 2 after researchers replace growing chat logs with structured memory
Swapping ballooning chat logs for structured memory layers kept the prompt near 5,000 tokens instead of ballooning past 500,000, and the agent won consistently where log-based competitors won zero games. A concrete data point for anyone worried about context costs in long-running agent tasks.
AI boosts research careers but narrow the span of ideas explored: study
A study finds AI tools help individual researchers' output but narrow the range of ideas a field explores overall. Worth a caution for any R&D or innovation team leaning too hard on AI-assisted brainstorming.
GATS: Graph-Augmented Tree Search with Layered World Models for Efficient Agent Planning
A lighter world model cuts the number of costly LLM inference calls needed for multi-step agent planning. Aimed at making agent planning cheaper and less stochastic, which matters once you try to run these in production rather than demos.
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
Altman's reversal from warning about vanishing professions to calling AI net job-creating is notable, but the actual data still doesn't clearly back either the doom or the optimism camp. Treat both positions as opinion, not evidence.
Claude Cowork's biggest use case is the mundane office work nobody wants to own, Anthropic says
Anthropic's look at 1.2 million Cowork sessions found most usage goes to status reports, onboarding checklists, and slide decks, the admin work around the work. That's a useful signal for where AI adoption is actually landing versus where the hype points.
Grades dropped from 96 to 48 percent when a Brown professor made students take the exam without AI
An in-person retest sent average scores from 96 to 48 percent once AI access was removed, and larger studies from China and Berkeley back up the pattern. A hard data point for any org rethinking how it certifies skills people learned with AI assistance.
LinkedIn is the undisputed king of long-form AI slop, according to a study spanning five platforms
A five-platform scan found 41 percent of long-form LinkedIn posts are AI-generated, far more than any other platform measured. Worth discounting thought-leadership posts there a bit more than you already do.