The Manifest
Supply chain legal and structural fights took center stage with a RICO suit against CH Robinson and South Africa's port reform breakthrough, while Washington launched a new AI oversight unit and Google quietly reshaped Gemini's free tier and open source bug bounty program.
Get The Manifest in your inbox
One sharp digest of AI × supply chain. Free. No noise.
Today's Top
- 01CH Robinson losing the headline war, the plaintiffs' bar has noticedThe Loadstar
- 02South Africa's port reform takes an important step forwardThe Loadstar
- 03Trump launches 'Super Intelligence Force' that has nothing to do with actual superintelligenceThe Decoder
- 04Google's new Gemini tiers cut free users to its weakest model and lock $5/month subscribers out of ProThe Decoder
- 05GPT-6 Astra vs GPT-6.1 Sol vs Gemini 4 Argon vs Claude Fable 5.1: Which Frontier Model Fits Which JobMarkTechPost
Models & Releases
4 storiesYandex Introduces Sona: A Single Generative Recommender That Replaces Entire Recommendation Cascade
For teams running recommendation-driven demand sensing or merchandising, collapsing a multi-stage recommender stack into one transformer means less infrastructure to tune and maintain, worth watching if it generalizes past music streaming.
The Story of Qwen: Alibaba's AI Models From 7B to 2.4T
A useful reference if you're evaluating open-weight models for in-house deployment, shows how fast scale and license terms have shifted across one vendor's lineup in three years.
Aleph Alpha Releases Kolibri: A 78.1B Open-Weight English-German MoE Model With Only 3.46B Active Parameters
A sub-80B model that runs on a single GPU is a practical option for teams needing English-German coverage without hyperscaler dependency, especially if EU data residency is a requirement.
Can an Open Model Do Security Research? Cantina's apex-flash-1 Solves 40 of 60 Held-Out Bug Tasks
An open model tuned for vulnerability research passing 40 of 60 held-out bug tasks is relevant for ops teams wanting to run their own security scanning instead of paying for proprietary tools, though BF16 deployment still needs serious GPU memory.
Supply Chain & Ops
5 storiesWhen a truck derails a train, who pays?
The federal insurance floor for trucks at grade crossings hasn't moved since 1985, a liability gap that will eventually land on shippers and carriers when the next collision isn't a minor one.
CH Robinson losing the headline war, the plaintiffs' bar has noticed
Six family-owned carriers accusing CH Robinson and TQL of routing freight through chameleon carriers under a RICO complaint is a serious escalation, brokers should expect more scrutiny on carrier vetting practices going forward.
South Africa's port reform takes an important step forward
Separating the port authority from Transnet's rail arm is overdue, shippers routing through South African ports should watch for capacity and reliability improvements over the next year.
Borderlands Mexico: Identity thieves target customs payments in cross-border freight
Identity thieves targeting customs payments is a reminder to tighten verification on cross-border brokers, while LEGO's $400M Mexico manufacturing expansion signals nearshoring demand isn't slowing down.
Where freight forwarders actually lose money
Uncollected demurrage charges and pre-dispatched haulage that never gets reimbursed are quietly eating margins, a solid audit checklist for any forwarder's finance team.
Deals & Market
2 storiesGoogle's new Gemini tiers cut free users to its weakest model and lock $5/month subscribers out of Pro
Free users get bumped to the weakest model while Pro gets locked behind a paywall, expect this pattern to repeat industry-wide as frontier model compute costs keep rising.
Trump launches 'Super Intelligence Force' that has nothing to do with actual superintelligence
A new White House unit coordinating with AI companies, critical infrastructure operators and interest groups, the practical impact on procurement and compliance is still unclear but worth tracking.
Research & Frontier
5 storiesGoogle researchers find a way to keep self-improving AI agents from memorizing their tests
RRSI reins in benchmark memorization and lifts scores on unseen tasks by up to 4.7 points using 30 percent fewer tokens, relevant for anyone building self-improving automation pipelines rather than one-off agents.
NASA and IBM's open source lunar model turns 17 years of orbiter data into a foundation for lunar science
Not supply chain directly, but it's a template worth stealing: a foundation model trained on 17 years of sensor history sharpens long-horizon prediction, the same pattern applies to asset or demand data operators sit on.
Google Research Moves Federated Learning Into TEEs: Gboard Now Trains With Externally Verifiable Differential Privacy
Externally verifiable differential privacy on attested server-side hardware is a meaningful step for anyone who needs to audit a vendor's privacy claims rather than just take them on faith.
MACTS-EM: Multi-Agent Collaborative Time Series Forecasting with Emergent Memory
Multi-agent forecasting with shared memory targets exactly the regime-shift problem planners deal with when demand patterns break, still early research but worth tracking for next-gen planning tools.
Choosing Before Acting: Comparative Value Estimation for Long-Horizon Tool-Use Agents
Addresses weak credit assignment in long tool-call chains, groundwork relevant to anyone building agentic procurement or planning workflows that string together many tool invocations.
Org & AI Architecture
3 storiesCan 'super intelligence' and a non-binding safety pact solve AI's image problem?
This reads more like a rebrand exercise than a policy shift, don't expect it to change anyone's actual compliance obligations.
Google froze its open source bug bounty program due to a 'significant rise' in AI submissions
Low-quality AI-generated vulnerability reports overwhelming a human review process is a preview of what happens to any ops channel that opens itself to AI-generated submissions without gatekeeping.
Chinese AI models parrot state doctrine or refuse to answer on sensitive topics
Before trusting a Chinese open-weight model for anything customer-facing or compliance-adjacent, test it on your own sensitive topics first rather than assuming it behaves like the benchmarks suggest.