The Manifest
China's Meituan releases a 1.6T-parameter open model running on domestic hardware, Amazon signals the end of crowdsourced data labeling, and IATA's new air freight liability rules force contract reviews across air forwarders.
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 Context and LongCat Sparse AttentionMarkTechPost
- 02Amazon will stop accepting new customers for Mechanical TurkTechCrunch AI
- 03News in Brief Podcast | Week 27 2026 | Air forwarder liability and Maersk optimismThe Loadstar
- 04Mark Zuckerberg tells staff that AI agents haven't progressed enoughtechcrunch.com
- 05Mistral CEO Mensch says proprietary AI models give labs a front-row seat to your business processesThe Decoder
Models & Releases
4 storiesMeituan Releases LongCat-2.0: A 1.6T-Parameter Open MoE Model with Native 1M Context and LongCat Sparse Attention
A 1.6T-parameter open model running entirely on domestic Chinese AI hardware raises the bar for open long-context inference outside the US cloud stack. Watch whether supply chain planning and document-processing workloads in the China market begin migrating to this infrastructure.
Baidu's 'Unlimited OCR' processes dozens of document pages in one pass by treating memory like human forgetting
Flat memory use regardless of document length removes a real constraint for teams processing multi-page bills of lading, contracts, and compliance filings in one pass. Worth benchmarking against your current document pipeline if throughput is the bottleneck.
Claude Code and Fable 5 ported the 2003 PC game Command & Conquer to native iOS in 'a few hours'
A full game port in hours is a headline, but the practical signal is that complex, multi-file legacy codebases are now tractable targets for AI-assisted modernization. Teams carrying aging ERP customizations should be paying attention.
LlamaIndex 'legal-kb': Agentic Retrieval over Index v2 with retrieve, find, read, and grep Tools
The filesystem-style tool API gives agents a more reliable way to navigate large document sets than pure vector search alone. A reference implementation worth reviewing if you are building agents over contract or compliance libraries.
Supply Chain & Ops
2 storiesNews in Brief Podcast | Week 27 2026 | Air forwarder liability and Maersk optimism
New IATA direct air waybill rules shifting liability onto forwarders will force contract reviews across air freight operations. Maersk's upward earnings revision separately signals that carriers see current demand as durable, not a temporary spike.
Borderlands Mexico: Cargo theft falls, but attacks on truckers turn deadlier
Fewer incidents but higher violence per attack signals a shift in threat profile on the Mexico-US corridor. Clarios opening a $147M battery distribution center in the same region shows manufacturers are still committing capital to nearshore supply chains despite the security environment.
Deals & Market
2 storiesAmazon will stop accepting new customers for Mechanical Turk
Amazon closing Mechanical Turk to new customers signals that crowdsourced data labeling at scale has been displaced by synthetic data and model-based annotation. Teams still budgeting for MTurk-style annotation pipelines should reassess before the next contract cycle.
Hollywood wants Seedance banned and reportedly also wants to keep using it
The MPA's first cease-and-desist against an AI video company, issued while studios quietly use the same product, illustrates the gap between official vendor policy and operational reality. Procurement and legal teams building AI governance frameworks will find this a useful case study in how that tension plays out.
Research & Frontier
3 storiesAI search agents don't fail at searching, they fail at asking the right questions when queries get ambiguous
Agents that retry searches instead of asking clarifying questions score worse on ambiguous queries, 51.9% versus strategies that pause to confirm intent. For anyone designing agent workflows over procurement catalogs or supply data, this is a concrete reason to build clarification loops into the architecture.
Structured PDF-to-JSON: A Guide to Open-Source Extraction Models in 2026
Most enterprise data still sits in PDFs and scans. This guide covers where open-source extraction models stand on the two distinct problems in document-to-JSON work, useful orientation before committing to a vendor or building extraction infrastructure in-house.
Qwen's Former Lead on What Hybrid Thinking Got Wrong and Why He Now Backs Agents
The gap between models that reason well and agents that reliably act in production is wider than benchmarks suggest. The RL infrastructure required to train agents that generalize across real workflows is still nascent, which is a grounding check for aggressive agent deployment timelines.
Org & AI Architecture
3 storiesMark Zuckerberg tells staff that AI agents haven't progressed enough
An unusually candid admission from the company betting heavily on agent-driven product development. If Meta's scale and resources are hitting this ceiling, internal roadmaps that assumed frontier agent capabilities were imminent deserve a realistic review.
Mistral CEO Mensch says proprietary AI models give labs a front-row seat to your business processes
The concern that AI labs accumulate customer business-process data and can eventually compete with their own clients is worth taking seriously in vendor reviews. The argument for open or self-hosted models deserves a seat at the table, separate from Mistral's own commercial interest in pushing the position.
AI private schools sell wealthy US families on personalized learning over traditional education
AI-fluent workers are being separated from those without access or training, and this dynamic is playing out in enterprise teams as clearly as in K-12 classrooms. The $75K tuition is the headline; the real signal is that early AI fluency is becoming a durable advantage that organizations need to build deliberately rather than hope for.