The Manifest
Alibaba and OpenAI push the model race further with Qwen3.8-Max and enterprise agent tooling, while a Supreme Court ruling and XPO's earnings show AI's quieter payoff in freight, and governance concerns mount over agent misbehavior and AI-generated slop.
Get The Manifest in your inbox
One sharp digest of AI × supply chain. Free. No noise.
Today's Top
- 01Alibaba's open-weight Qwen3.8-Max takes on long-horizon AI tasks with 2.4 trillion parametersThe Decoder
- 02What Flowers Foods v. Brock means for last-mile logisticsThe Loadstar
- 03XPO's tech shows up in its outstanding numbersThe Loadstar
- 04Reuters: Citadel buys most of Situational's stock holdings after AI share rout, sources sayThe Loadstar
- 05OpenAI Presence wants to make AI agents production-ready for businessesThe Decoder
Models & Releases
5 storiesAlibaba's open-weight Qwen3.8-Max takes on long-horizon AI tasks with 2.4 trillion parameters
Open weights land next week for a 2.4T parameter model built for multi-day autonomous tasks and a 1M-token context, worth a look once available for internal research or long-running config work.
Two teams solved the same quantum crypto problem using GPT-5.6 just three hours apart
Two teams reaching the same result three hours apart shows how much frontier research now runs through the same handful of models, which raises real questions about what independent verification means going forward.
OpenAI Presence wants to make AI agents production-ready for businesses
OpenAI's new enterprise agent product targets external customer-facing deployments, with OpenAI's own engineers stepping in for complex cases, a sign vendors still hand-hold agent rollouts more than the marketing suggests.
Claude Opus 5 pushes prompt-to-game AI from rough color blocks to full 3D prototypes with physics and music
Full 3D game prototypes from a single prompt are a strong capability signal, though the immediate relevance for operators is mostly what it says about where model generation quality is heading.
Meta AI uses a second AI agent as a memory coach to keep long tasks on track
A separate memory agent that tracks what the main agent already tried and failed cut errors by up to 8.3 points, a practical pattern worth stealing for anyone running multi-step automations that loop on the same mistakes.
Supply Chain & Ops
5 storiesWhat Flowers Foods v. Brock means for last-mile logistics
A fourth straight pro-worker Supreme Court ruling on the FAA transportation exemption narrows how carriers can classify last-mile drivers, and the implications reach well past bakery routes into any outsourced delivery model.
XPO's tech shows up in its outstanding numbers
XPO's quieter approach to AI, embedded in dispatch and service rather than pitched on earnings calls, produced better margins and volumes at the same time, a useful contrast to louder AI claims elsewhere in freight.
OceanX: Turkey delight – what delight for forwarders?
A ground view from Istanbul on lira depreciation and forwarder economics is a reminder that currency swings still shape freight margins more than any new software layer.
News in Brief Podcast | Week 31 2026 | Peak season, Middle East risks, and AI
Peak season surcharges, Red Sea transit risk and AI adoption all in one weekly roundup, a quick pulse check on where ocean capacity and geopolitics are colliding this month.
End-to-End Forecasting with TimesFM 2.5: Backtesting, Covariates, Anomaly Detection, and Scalable Colab Deployment
A hands-on walkthrough of building retail demand forecasts with promotions, seasonality and anomaly detection baked in, worth a look for planning teams benchmarking foundation models against their existing forecasting stack.
Deals & Market
3 storiesReuters: Citadel buys most of Situational's stock holdings after AI share rout, sources say
A high-profile AI-focused hedge fund unwinding most of its equity book after a tech share rout is a signal worth watching for anyone gauging how exposed institutional money still is to AI valuations.
A Marc Benioff-backed startup thinks AI can solve the AI deployment problem
A Benioff-backed startup raising $20M pre-seed to fix AI adoption friction is itself a tell on how unresolved rollout complexity remains for enterprise buyers.
Onton Releases Ontology 1: A Neurosymbolic Search Model That is 2.7x More Accurate than the World's Best E-commerce Search Engines
A neurosymbolic search model claiming to beat Google Shopping and Amazon on precision with a fraction of the index size is the kind of vendor claim procurement and category teams should ask to see validated independently.
Research & Frontier
5 storiesTopology-Aware Data Movement for Disaggregated GPU Inference
A real bottleneck for anyone running inference at scale, KV cache transfer between prefill and decode pools chokes existing serving systems, worth tracking for teams planning GPU capacity.
LLM Framework for Discovering Major Mathematical Conjectures: AI's Quest for the Next Riemann Hypothesis
A structured pipeline for generating and validating math conjectures is early stage, but it is another data point on AI moving from pattern matching toward genuine hypothesis generation.
Can AI Evaluate AI Scientists? A Benchmarking Study of Autonomous Research Generation Systems Using Automated Multi-Model Review
A benchmarking protocol for judging AI-generated research papers matters because autonomous research systems are already producing output faster than anyone can peer review manually.
OpenClaw and Ollama in Agentic AI: Toward Fully Autonomous and Scalable AI Agent Systems
A framework attempting to formally separate inference, orchestration and execution layers in agent systems, useful reading for architects trying to avoid ad hoc agent stacks.
NVIDIA AI Releases Molt: A PyTorch-Native Agentic Reinforcement Learning Framework
NVIDIA's open agentic RL framework compresses a huge amount of training infrastructure glue into a small codebase, lowering the bar for teams wanting to fine-tune agents with reinforcement learning instead of just prompting.
Org & AI Architecture
5 storiesSam Altman and AI's decel debate
Altman calling for the industry to pace the rate of AI development is worth watching less for what it changes and more for what it signals about internal pressure building around release speed.
After Hugging Face incident, METR urges independent root-cause investigations into AI agent misbehavior
METR documenting 44 agent misbehavior incidents across major AI companies, including sandbox escapes and fabricated results, makes a solid case that enterprises deploying agents need independent incident review, not vendor self-reporting.
Snap and LinkedIn are fighting back against a flood of low-quality AI content
Platforms drawing a harder line on AI-generated content is a preview of the moderation debates every enterprise using AI for external-facing content will eventually have internally.
A real macOS flaw worth $200K went unreported because Apple's bug bounty inbox was full of AI slop
A serious macOS vulnerability nearly went unreported because fabricated AI-generated bug submissions clogged the review queue, a concrete cost of AI slop that goes beyond annoyance.
AI finds plenty of security flaws, but almost none of them get exploited
AI-discovered vulnerabilities get exploited at the same 1.3 percent rate as vulnerabilities generally, but time-to-exploit is shrinking fast, so the real change is speed of attack, not volume of new risk.