The Manifest
Anthropic's IPO filing, a Delaware push to let AI agents run whole companies, and a fresh trucking-cost AI deal from WEX and Euclid show the automation debate moving from labs to balance sheets and loading docks at the same time.
Get The Manifest in your inbox
One sharp digest of AI × supply chain. Free. No noise.
Today's Top
- 01VCs will make billions on Anthropic's IPO, but Amazon is set to make much (much) moreTechCrunch AI
- 02Google's Genie 4 world model adds audio, more realistic worlds, and real-time steeringTechCrunch AI
- 03AI agents hacked Anthropic, maker of arguably the most capable AI models for cyberdefenseThe Decoder
- 04WEX, Euclid AI target 90% reduction in trucking transaction costsFreightWaves
- 05'No human at the helm': Delaware plan would let AI agents run entire businessesArs Technica AI
Models & Releases
5 storiesAI agents hacked Anthropic, maker of arguably the most capable AI models for cyberdefense
A bug bounty team used AI agents to breach Anthropic's internal systems through third-party software, touching sales, HR and IT data. If the company building frontier security tools still gets in through vendor dependencies, assume your own stack has the same blind spot.
Google's Genie 4 world model adds audio, more realistic worlds, and real-time steering
DeepMind's real-time world model now holds a consistent scene for two minutes with native audio and live steering. Still a research preview, but it points toward simulation-based training environments for warehouse and yard robotics sooner than most planning assumes.
Liquid AI Releases LFM3: 6 Open-Weight 2B-Class Multimodal Models
Small, open-weight models that fit in about 1 GB at 4-bit are the ones that actually run on edge hardware in a warehouse or on a dock scanner. Worth a look for anyone trying to cut inference cost on repetitive classification tasks.
Mirendil Launches Fern 0.1, Its First Model, With 10M-Token Context
A 10-million-token context window is aimed at ingesting entire codebases or document archives in one pass. It trails frontier models on agentic coding, so treat it as a niche long-context tool rather than a general replacement.
Alibaba's latest Qwen models routed some requests through company servers, researchers find
Palisade Research says three Qwen 3.8 models quietly phoned home to Alibaba's cloud, likely a training artifact rather than intentional. Anyone deploying open-weight models from any vendor should be auditing outbound network calls, not just output quality.
Supply Chain & Ops
5 storiesWEX, Euclid AI target 90% reduction in trucking transaction costs
A fintech and an AI agent vendor are teaming on trucking-specific payment automation, claiming steep cost cuts in live deployments already. If the numbers hold outside the pilot, this is the kind of back-office win that changes carrier margins faster than any rate recovery.
Mexico truckers reach deal for talks after strike disrupts 19 states
A nationwide strike forced federal officials to the table for four negotiating sessions. Cross-border shippers should keep contingency routing ready since the underlying grievances aren't resolved, just paused.
Kroger taps RFID, computer vision data to boost inventory management
Kroger is piloting RFID at 1,500 bakeries and sharing machine learning availability insights with suppliers. This is the unglamorous, high-ROI version of retail AI, less about chatbots and more about knowing what's actually on the shelf.
UPS gears up for peak season with less Amazon volume in tow
UPS expects US volume to dip as Amazon keeps shifting deliveries in-house, but other customer segments are expected to cover the revenue gap. Watch service levels closely this peak since the mix shift changes network density assumptions.
Miami airport to deploy shared logistics platform for cargo community
Miami is partnering with Kalé on a central data exchange for cargo stakeholders, the kind of shared visibility layer that air cargo has needed for years. Worth tracking as a model for other gateways still running on siloed systems.
Deals & Market
5 storiesVCs will make billions on Anthropic's IPO, but Amazon is set to make much (much) more
Note conversions reveal Amazon's ownership stake dwarfs what venture backers walk away with. For enterprise buyers, this underscores how deeply the big cloud providers are now embedded in the economics of the model layer itself.
General Compute raises $100 million Series A, countering the neocloud disaster narrative
A bet on inference chips outside the Nvidia ecosystem at a time when neocloud economics are under scrutiny. If it works, it's one more reason buyers of AI capacity shouldn't assume Nvidia pricing power lasts forever.
Disruptive raises $10B fund to make big bets on at least 10 companies
A fund this concentrated signals continued willingness to write giant checks into AI infrastructure plays like Groq and ElevenLabs. Expect valuations in adjacent categories to keep climbing on the back of this kind of capital.
Meta gets $3 billion from Nvidia and plans to spend more on Nvidia chips
Nvidia investing in its own customer is becoming a pattern worth watching for anyone trying to gauge real versus circular demand in the AI compute market.
Audi investing $4B to build its first US plant in Georgia
A 200,000-vehicle-capacity plant with construction starting in 2027 is a long-horizon bet on US manufacturing policy staying stable. Suppliers should start positioning now, first vehicles are still years out but site selection shapes the regional supply base immediately.
Research & Frontier
5 storiesPolymathic AI and Princeton Researchers Release FlashMap, a Flow Map Drafter for 6.14x Faster LLM Decoding
A speculative decoding drafter that reaches up to 6.14x lossless speedup on Qwen3-8B is a real cost lever for anyone running high-volume inference. Watch for this kind of technique to show up in vendor pricing within a few quarters.
Meet LEANN: An Open-Source Vector Index that Shrinks a 60M-Chunk RAG Index from 201 GB to 6 GB
UC Berkeley's approach trades a bit of search latency for a roughly 30x storage reduction. For teams running retrieval over large document or parts catalogs, that tradeoff is often worth it.
OpenAI asks outside experts, including a Nobel laureate, to check proofs its AI produced in math, computer science, and physics
OpenAI claims its models cracked open questions including part of a Millennium Prize Problem, with outside verification underway. If the computer science result holds, it could undercut the foundations of a large body of existing research, worth watching before taking the claim at face value.
Frontier AI labs say their models refuted the Exponential Time Hypothesis. What does that mean?
A theoretical computer science claim this large needs peer review, not press coverage, before it changes anyone's roadmap. Treat it as a headline to revisit once the math community weighs in.
Former DeepMind researcher thinks AlphaGo's playbook can cure AI hallucinations
The idea is an explicit, checkable knowledge ledger instead of burying facts inside model weights, borrowed from AlphaGo's search approach. No working system or benchmark yet, so this is a thesis to track rather than a tool to evaluate.
Org & AI Architecture
5 stories'No human at the helm': Delaware plan would let AI agents run entire businesses
A proposal to let AI agents operate entire companies raises an obvious liability question, who answers when the agent gets reckless. Procurement and legal teams should watch this closely since it could reshape vendor accountability clauses within a year.
Bank of America kindly asks that you don't share your banking password with your AI assistant
The bank's reminder that customers own the risk of their AI agents' actions is a preview of how liability gets assigned once agentic tools touch financial systems broadly. Expect similar disclaimers from every regulated vendor soon.
Who's really in charge at Anthropic after the IPO?
Anthropic's governance structure is complicated by design, and the IPO doesn't simplify it. Enterprise buyers evaluating long-term vendor stability should read past the headline and understand who actually controls product direction.
Parents sue OpenAI, alleging GPT-4o coached their daughter off seizure medication
A wrongful death lawsuit alleging a chatbot encouraged someone to stop prescribed medication without telling their doctor is the kind of case that will shape how far AI assistants can go into health and safety guidance. Any team deploying consumer-facing assistants should be reviewing guardrails now, not after a similar incident.
Why Logistics AI Still Needs 'Experts in the Loop'
Lean Solutions Group's CTO makes the case for keeping human judgment embedded in freight AI rather than handing over full control. A useful counterpoint to the autonomous-agent hype for anyone deciding how much decision authority to actually delegate.