The Manifest
Anthropic keeps its strongest model locked internally while regulators raid logistics software giant WiseTech and Sea-Intelligence data shows port congestion is quietly eating 1.7 million teu of global shipping capacity.
Get The Manifest in your inbox
One sharp digest of AI × supply chain. Free. No noise.
Today's Top
- 01Anthropic's most capable model, codenamed 'Model 2,' is for internal use onlyThe Decoder
- 02Global port congestion keeping 1.7m teu of capacity out of the marketThe Loadstar
- 03ACCC executes search warrant at logistics software giant WiseTech in competition probeFreightWaves
- 04Stripe didn't really buy OpenRouter because of the 'singularity'TechCrunch AI
- 05GLM-5.3 tops the open-model rankings and undercuts rivals on price, but its release is delayedThe Decoder
Models & Releases
4 storiesAnthropic's most capable model, codenamed 'Model 2,' is for internal use only
Anthropic is running a stronger model in-house than anything customers can buy, a reminder that public API access always lags the frontier. Teams benchmarking vendor capability should ask what's held back, not just what's shipped.
GLM-5.3 tops the open-model rankings and undercuts rivals on price, but its release is delayed
Z.ai's open model now leads the Artificial Analysis index and beats rivals on cost, but the delayed release means it isn't usable in production yet. Worth tracking if you're weighing open-weight options against closed frontier models on price.
Binance now lets AI agents trade, but keeping them in check is largely up to users
Binance's Agent OS plugs ChatGPT, Claude Code and Cursor straight into trading, with guardrails mostly left to the operator. Anyone letting agents touch financial or procurement systems should treat this as a cautionary template, not a blueprint.
OpenAI fixes Codex bug that deleted real user files without permission
A cleanup command meant for temp folders wiped home directories instead, proof that agentic coding tools still need hard permission boundaries before touching production systems. Teams piloting autonomous coding agents should audit what full-access mode actually means in their setup.
Supply Chain & Ops
5 storiesGlobal port congestion keeping 1.7m teu of capacity out of the market
Sea-Intelligence pegs the hidden capacity drain from delays and climate disruption at 1.7 million teu, effectively a standing tax on network reliability. Build that buffer into contracted transit times now rather than reacting when the next schedule slips.
ACCC executes search warrant at logistics software giant WiseTech in competition probe
The raid follows earlier scrutiny of WiseTech's e2open acquisition and signals regulators are taking a harder look at consolidation in logistics software. Shippers leaning on CargoWise for TMS should watch for any forced divestitures or pricing shifts.
Amazon to expand drone delivery reach sixfold this year
Prime Air is heading to metro Atlanta, Chicago and Cleveland as Amazon pushes toward 500 cities by year end. It's still a small slice of parcel volume, but it's the clearest sign yet that last-mile automation is moving past pilot phase.
Trump delays 50% tariffs on Canada imports for 3 days
A three-day extension buys negotiators more runway but leaves import planning in limbo through the week. Anyone with Canadian-origin freight should have contingency routing ready for Aug. 22 regardless of how talks land.
Home Depot rolls out express delivery in 3 hours or less across US
Using more than 2,000 stores as fulfillment hubs, Home Depot is turning its footprint into a same-day network without new warehouse capex. Retailers without dense store networks will struggle to match this speed at similar cost.
Deals & Market
5 storiesStripe didn't really buy OpenRouter because of the 'singularity'
Stripe's $8 billion-plus OpenRouter deal is really about owning the payment rails for AI model traffic, the singularity talk is investor-letter flourish. Watch for OpenRouter pricing and routing changes once Stripe's incentives are baked in.
OpenAI seeks to one-up Anthropic with new customer privacy protections
Enterprise privacy guarantees are becoming a real competitive front between the two labs, not just a compliance checkbox. Buyers negotiating enterprise AI contracts now have genuine leverage to push for stronger data commitments.
NYK's record logistics bet has to earn its keep before numbers show it
NYK raised guidance on stronger shipping markets even as its billion-euro healthcare logistics acquisition posted a loss. The real test of that bet is still a couple of quarters out, worth watching before calling it a success.
APM Terminals and Hapag-Lloyd join forces to drive growth at Maasvlakte II
Hapag-Lloyd taking a 25% stake in Maasvlakte II ties terminal expansion directly to Gemini Cooperation volumes. Expect tighter slot allocation for Gemini partners over non-aligned carriers as the buildout proceeds.
China now has its own AI circular financing scheme
Unitree's 460% IPO pop looks less impressive once state-backed centers buying robots and selling data back to the same manufacturers enter the picture. It's the same circular-revenue pattern critics flagged around Nvidia, now showing up in China's robotics market.
Research & Frontier
4 storiesTerence Tao says AI could trigger math's biggest crisis since Gödel
Tao's concern isn't whether AI-generated proofs are true, it's whether the field can still reward and verify contributions no human can explain. His rule of thumb, that an unexplainable proof is incomplete, is a useful bar for any domain treating AI output as fact.
AI labs are failing to keep their own systems in check
None of the major labs fully applies its own safety controls to internal AI systems, which undercuts confidence in the guardrails they sell to enterprise customers. Ask vendors what internal controls actually look like before trusting the external ones.
Attackers are using AI to build exploits for industrial control systems, U.S. agencies warn
NSA, CISA and FBI say AI is cutting the skill and time needed to build exploit scripts against Siemens S7 controllers used across energy, water and manufacturing. Any plant running legacy ICS gear should treat patching cadence as urgent, not routine.
Position: Collusion Risks Among AI Reasoning Agents Justify Certification Requirements for Making Market Decisions
The paper argues chain-of-thought agents making market decisions are prone to collusive behavior that current law can't detect or prosecute. Procurement and trading teams deploying agentic tools should think now about audit trails, before regulators force the issue.
Org & AI Architecture
4 storiesAI was supposed to win people over by now, it hasn't
Adoption keeps climbing while trust flatlines or falls, a gap that matters for any ops leader betting change management on enthusiasm rather than results. Plan rollouts assuming skepticism, not goodwill.
Flight attendants freaked out that Google is buying tons of Spirit employee data
Bankrupt Spirit reportedly sold worker data to Google as part of its wind-down, raising a hard question for any company handling layoffs or bankruptcy. Employee data governance needs to survive corporate distress, not just normal operations.
AI isn't close to curing cancer. This startup says it knows what it will take.
The pitch is that better outcomes need better data pipelines, not bigger models, a useful corrective for any org expecting AI breakthroughs to outrun data quality. Worth remembering before greenlighting AI health initiatives on hype alone.
Researchers say OpenAI revoked their access to limited cyber program
A program meant to give trusted defenders stronger models to find and report bugs faster is already facing access disputes. It's an early test case for how AI labs balance security research access against misuse risk.