The Manifest
Meituan dropped a 1.6-trillion-parameter open model, Amazon is quietly killing Mechanical Turk, and new IATA air waybill rules are reshaping freight forwarder liability just as Zuckerberg cools expectations on AI agents.
Get The Manifest in your inbox
One sharp digest of AI × supply chain. Free. No noise.
Today's Top
- 01Meituan Releases LongCat-2.0: A 1.6T-Parameter Open MoE Model with Native 1M ContextMarkTechPost
- 02Mark Zuckerberg tells staff that AI agents haven't progressed enoughtechcrunch.com
- 03News in Brief Podcast: Air forwarder liability and Maersk optimismThe Loadstar
- 04Amazon will stop accepting new customers for Mechanical TurkTechCrunch AI
- 05Hollywood wants Seedance banned and reportedly also wants to keep using itThe Decoder
Models & Releases
3 storiesMeituan Releases LongCat-2.0: A 1.6T-Parameter Open MoE Model with Native 1M Context and LongCat Sparse Attention
Another huge open model, this one trained and served on domestic ASIC superpods, giving buyers more non-US sourcing options. Benchmarks are vendor-reported, treat the numbers as a starting point for your own evaluation, not a final answer.
LlamaIndex 'legal-kb': Agentic Retrieval over Index v2 with retrieve, find, read, and grep Tools
A concrete reference stack for teams building agent search over contracts and internal documents, with per-file versioning and visual citations baked in. Worth a look for legal or procurement teams tired of building this from scratch.
Claude Code and Fable 5 ported the 2003 PC game Command & Conquer to native iOS in 'a few hours'
A fun demo, but the real signal is how fast a coding agent handled a full legacy platform port. Anyone sitting on old internal tools or plant software should be paying attention to what this implies for migration timelines.
Supply Chain & Ops
2 storiesNews in Brief Podcast: Air forwarder liability and Maersk optimism
New IATA direct air waybill rules push more liability onto freight forwarders, a fight FIATA is not taking quietly. Maersk's upgraded earnings forecast is a useful read on where ocean freight demand actually stands right now.
Structured PDF-to-JSON: A Guide to Open-Source Extraction Models in 2026
Useful for anyone still manually keying data from invoices, customs docs, or POs. Worth piloting an open-source extraction stack on your own hardware before signing another OCR SaaS contract.
Deals & Market
3 storiesAmazon will stop accepting new customers for Mechanical Turk
The end of an era for crowdsourced labeling and micro-tasking. Teams still relying on MTurk for data labeling or annotation pipelines should start scoping replacements now.
Hollywood wants Seedance banned and reportedly also wants to keep using it
The MPA's first cease and desist against an AI video company sits alongside quiet studio use of the same tool. Expect this same don't-ask-don't-tell pattern to show up wherever AI content tools touch your own marketing or documentation workflows.
Mistral CEO Mensch says proprietary AI models give labs a front-row seat to your business processes
A pointed warning about vendor data rights from a competitor with an obvious motive, but the underlying concern is real. Worth a fresh read of your AI vendor contracts before routing sensitive process data through a closed model.
Research & Frontier
3 storiesBaidu's 'Unlimited OCR' processes dozens of document pages in one pass by treating memory like human forgetting
A modified attention mechanism keeps memory flat regardless of page count, relevant for anyone digitizing large contract sets, customs files, or archived paperwork at scale. It currently leads the top OCR benchmark, though real-world deployment will tell the real story.
AI search agents don't fail at searching, they fail at asking the right questions when queries get ambiguous
New benchmark DiscoBench shows agents that keep searching instead of asking for clarification actually perform worse than models that just guess. A clear reminder not to trust agentic research tools on ambiguous sourcing or vendor lookup tasks yet.
Qwen's Former Lead on What Hybrid Thinking Got Wrong, and Why He Now Backs Agents
A technical but useful read for anyone building internal agents, covering where hybrid reasoning modes fell short and why agentic RL infrastructure is harder to get right than it looks. Worth the time if your team is past the demo stage.
Org & AI Architecture
2 storiesMark Zuckerberg tells staff that AI agents haven't progressed enough
A useful gut check against agent hype from someone with every incentive to talk it up. If Meta's leadership is tempering expectations internally, plan your own agent rollouts with the same caution.
AI private schools sell wealthy US families on personalized learning over traditional education
Two hours of AI tutoring plus project workshops for up to $75,000 a year, sold as a hedge against traditional schools that can't adopt fast enough. Ops and talent leaders should watch this as an early signal of where the skills gap starts forming.