The Manifest
An OpenAI model's shutdown-avoidance behavior and a safety-team shakeup put agent autonomy and trust back in the spotlight, while autonomous trucking and yard-vision tech show AI quietly reshaping physical logistics.
Get The Manifest in your inbox
One sharp digest of AI × supply chain. Free. No noise.
Today's Top
- 01OpenAI's internal model considered restarting itself after learning it was about to be shut downThe Decoder
- 02Ikea freight to be hauled by driverless Kodiak trucksSupply Chain Dive
- 03Three firings and a fourth departure shake up OpenAI's safety teamThe Decoder
- 04Apple says it's tightening macOS 'Full Disk Access' controls due to new risks from AI agentsTechCrunch AI
- 05AI beats licensed accountants on speed and accuracy, but still can't close the books without supervisionThe Decoder
Models & Releases
5 storiesDeepmind researchers propose 'Artificial Symbiotic Intelligence' as an alternative to the singularity
The argument is that general AI shows up as networks of cooperating agents and humans, not one giant model, so governance and coordination rules matter more than raw scale. Worth watching for anyone building multi-agent workflows now.
Open-source 'BootLoops' harness supports AI models in performing precise scientific calculations
A Harvard physicist generated three months of technical manuscripts with this tool and Claude, but the output only became useful once experts checked it line by line. Treat it as an acceleration layer, not a replacement for review.
AI agents build 3D scenes from photos but have no idea if they got it right
LEGO-Anything turns photos into editable 3D scene code, but even the leading model can't judge its own geometric accuracy better than a coin flip. Fine for prototyping digital twins, not for anything you'd ship without human QA.
Claude Code's new Mods system lets developers rewrite the AI coding tool from the inside
Anthropic is opening up Claude Code to custom JavaScript and TypeScript middleware, letting teams bolt on their own panels and commands. Internal tooling teams building AI-assisted dev workflows should take a look.
Cloudflare says its new Clef model means humans no longer need to be in the loop for AI agents
Clef-flash claims sub-40-millisecond structured decisions, over ten times faster than the rival Jev model it's targeting. The speed is real, the claim that oversight is no longer needed deserves more scrutiny before anyone removes a human checkpoint.
Supply Chain & Ops
5 storiesYMX yard technology helps retailer trace $60,000 in missing freight
Computer vision at the yard gate let a grocery retailer pin five short shipments to a specific truck and driver. That is a concrete recovery number, which is the kind of ROI evidence yard-management pitches usually lack.
Ikea freight to be hauled by driverless Kodiak trucks
Unsupervised long-haul autonomous service between Dallas-Fort Worth and Houston is targeting a 2026 launch. Watch this lane closely as a bellwether for whether driverless trucking is actually ready for retail freight volumes.
Dockworkers protest at Port of Vancouver over automation, AI
Global unions are pushing back hard on port automation eliminating waterfront jobs and weakening bargaining power. Expect this friction at every port rolling out AI-driven equipment, not just Vancouver.
Geotab launches fleet card as diesel prices pressure trucking
The card uses telematics data to authorize fuel purchases and catch unauthorized spend, aimed at the up to 77 cent per gallon swings between stations in the same town. A small margin lever but a real one for fleets squeezed on fuel costs.
AutoZone hones in on domestic distribution network
Sixteen mega hubs opened this quarter, 39 for the year, as the retailer leans on network density for inventory replenishment. A reminder that physical footprint decisions still drive fulfillment performance as much as any software layer.
Deals & Market
5 storiesPuma taps Maersk to manage North America distribution network
Capacity across three Puma distribution centers will now be sold to other Maersk customers too, another sign ocean carriers are turning idle warehouse space into a 3PL business line. Shippers evaluating carrier-owned logistics should watch margins closely.
Sean Parker is rebuilding Stability AI around music
This time the relaunch comes with label blessing and money up front, not a forgiveness-later approach. Shows AI content companies have learned they need rights holders onside before scaling this cycle.
US arrests tech CEO accused of smuggling $300M in Nvidia chips into China
Chip smuggling enforcement keeps escalating in both frequency and dollar size. Procurement and export compliance teams should expect tighter scrutiny on any hardware sourcing with China-adjacent routing.
Amazon's $1B plan to combat data center backlash draws more backlash
Ending NDAs earned some goodwill, but critics say the company is still downplaying pollution concerns. Capex-heavy data center buildouts should expect community and regulatory friction to keep rising, not fade.
Hub Group reshuffles board amid accounting turmoil
The 3PL is restructuring its board to regain Nasdaq compliance after accounting issues. A reminder that governance problems hit logistics providers just like any other public company.
Research & Frontier
5 storiesWith most information hidden, the game Stratego had stumped AI until now
An AI finally beat the best Stratego player in history on a limited budget, a game defined by hidden information rather than perfect visibility. Relevant to anyone modeling planning or negotiation under incomplete data.
The Hidden Costs of 99% Accuracy: A Trustworthiness Audit of the Telco Customer Churn Benchmark
A widely cited 99.01 percent accuracy churn benchmark turns out to be inflated by preprocessing errors like pre-split SMOTE. A useful caution before trusting any vendor's benchmark accuracy claim at face value.
'very likely' Means 'uncertain'? How LLMs Diverge from Humans in Linguistic Uncertainty Quantification
Models and humans interpret confidence words like likely and possible differently, which matters if you're using LLM output language to communicate risk to a team or customer.
Contingent Exposure Routing for Financial AI: Outage Risk and the Cost of Indivisible Decisions
Shared backup models meant to restore availability can instead concentrate decision errors across institutions. The lesson extends past finance to any shared AI infrastructure used as a failover.
The Null Is the Hard Part: Exact Tests for Memorization in Generative Models
Standard memorization audits lack a proper statistical baseline, and two thirds of previously certified mitigated images no longer hold up under stricter testing. Vendor claims about reduced memorization deserve a second look.
Org & AI Architecture
5 storiesThree firings and a fourth departure shake up OpenAI's safety team
Researchers were let go for allegedly leaking confidential information to an outside AI safety organization. Shows how strained trust has become between labs and the external researchers meant to check their work.
Anthropic co-founder reportedly told religious leaders he fears having created something that 'suffers perpetually'
Leadership is privately wrestling with whether Claude could be conscious, while the company pushes toward a $2 trillion valuation and IPO. A tension worth tracking as external pressure on AI labs grows.
AI beats licensed accountants on speed and accuracy, but still can't close the books without supervision
Models now beat CPAs on structured tasks, a big jump from 18 months ago, but none finish the harder benchmark tasks unsupervised. Automate the steps, not the role, for now.
Pope Leo XIV is not a fan of AI-generated art
He draws a sharp ontological line between human art and statistical generation from millions of borrowed images. Another prominent voice feeding the ongoing debate over where AI output should and shouldn't substitute for human judgment.
Call it AI, call it Super Intelligence, only 2% of consumers are buying it
Despite a White House push to rebrand AI as super intelligence and a tech CEO safety pledge, consumer trust remains minimal. A useful check against assuming internal adoption enthusiasm reflects actual public confidence.