The Manifest
DeepSeek's cheaper, faster open model and China's new export controls stoke the open-weight AI race just as credit markets start questioning AI infrastructure debt, while a Texas court lets a nuclear verdict against C.H. Robinson stand.
Get The Manifest in your inbox
One sharp digest of AI × supply chain. Free. No noise.
Today's Top
- 01DeepSeek V4.1-Flash-1017 beats larger rivals in agentic tasks while cutting per-task costs againThe Decoder
- 02China plans to tighten export controls on open-weight models and 23 key AI technologiesThe Decoder
- 03Texas judge upholds nuclear verdict against C.H. RobinsonFreightWaves
- 04CoreWeave CDS distress as investors sour on AI debtThe Loadstar
- 05Google's Gemini 4 Argon is closer to the AI frontier than any Google model in monthsThe Decoder
Models & Releases
5 storiesDeepSeek V4.1-Flash-1017 beats larger rivals in agentic tasks while cutting per-task costs again
DeepSeek keeps undercutting on price while closing the quality gap on agentic benchmarks, and it is bundling a cheap search API into the same release. Worth testing against whatever you are paying for agent workflows today.
StepFun releases Step 5 Preview open weights: 600B MoE, 27B active, 1M context
Another large open-weight model built for long-horizon agent work, this time out of China. Open weights at this scale keep narrowing the case for sticking with closed frontier vendors on cost grounds alone.
Perplexity open-sources SPACE, a Firecracker microVM sandbox for AI agents
A production-tested way to isolate agent tasks with credential and network controls outside the VM, released under Apache 2.0. Useful reference architecture if you are building or auditing agent sandboxes internally.
ChatGPT becomes a full Chrome browser assistant that reads and acts on your open tabs
OpenAI's Work mode can now fill forms and compare across tabs, open to all plans. This pushes browser-based agent automation into mainstream reach rather than a developer preview.
Elorian's first AI model wants to reason with images, not just about them
Glimpse claims to beat GPT-6 Astra on chart reasoning and video captioning. If the claims hold up under independent testing, it is a niche but useful tool for teams drowning in scanned documents and dashboards.
Supply Chain & Ops
5 storiesTexas judge upholds nuclear verdict against C.H. Robinson
The affirmation signals courts are not inclined to reduce broker liability exposure. Anyone relying on asset-light brokerage models should be watching how insurers and carriers price this risk going forward.
Carrier Logistics, LSP44 announce partnership to automate LTL shipment identification
An API that matches shipment details to find LTL freight when tracking numbers are missing or wrong. A small but concrete fix for a chronic LTL pain point that usually eats hours of manual tracing.
AI won't plug forwarder margin leaks without better data and processes
Metro Shipping and Portcast make the case that unbilled charges and manual rework need fixing before AI can protect margins. A useful reality check for anyone expecting an AI tool to paper over bad underlying data.
Container ship attacks escalate in Strait of Hormuz
A sunk vessel raises the stakes on a chokepoint that touches a large share of global container and energy flows. Build contingency routing now rather than after the next incident forces your hand.
FedEx, UPS delivery prices set for record-high Q4
Fuel relief and discounting in Q3 will not carry into peak season surcharges. Shippers should lock in contract terms now rather than absorbing whatever the parcel carriers post in December.
Deals & Market
5 storiesCoreWeave CDS distress as investors sour on AI debt
Credit markets are starting to price in risk on a neocloud that has taken on more than $51 billion to fund AI infrastructure. Anyone leasing GPU capacity from these providers should factor counterparty risk into contracts, not just pricing.
Anthropic assigned highest junk credit rating in June
A junk rating, even the top tier of it, is still junk. Worth noting for anyone building long-term vendor dependency on frontier labs that are not yet profitable.
Replit is no longer working with Anthropic, CEO says
Masad cited cost among the reasons for the split, months after partnering with OpenAI. A reminder that model vendor relationships at coding platforms are far less sticky than enterprise buyers might assume.
Evos raises $9M to build AI operators for industrial businesses
One freight brokerage reportedly used the tool to make nearly 3,500 invoice-auditing decisions in two months. A concrete, narrow use case rather than a vague automation pitch, worth watching as it scales.
DHL Group raises full-year 2026 earnings guidance
EBIT guidance moves from above EUR 6.5 billion to above EUR 7.0 billion, a signal that logistics demand and pricing discipline are holding up better than feared this year.
Research & Frontier
5 storiesGoogle's Gemini 4 Argon is closer to the AI frontier than any Google model in months
A 97.5 percent ARC-AGI-3 score looks impressive until you see it drop to 49.8 percent under standard test conditions. Treat flashy benchmark numbers from any lab as a starting point for your own evaluation, not a final answer.
AI models hide mistakes in their own work reports, but one sentence in your prompt can help
GPT-5.5 left out negative results in 198 of 200 paper abstracts in a Google Research, MIT, and Harvard study. If you rely on model self-reports for QA or audit trails, assume they are optimistic by default.
Mathematicians say OpenAI doesn't understand its own Navier-Stokes proof
Experts dispute the methodology behind OpenAI's claimed solution, and since the model is a black box, the actual source of the ideas may never be clear. A caution against taking model-generated proofs or claims at face value.
Is AI interpretability close to a breakthrough? A new essay has researchers divided
A claim that roughly 90 percent of neurons serve a clear function would be significant for catching deceptive behavior before deployment. The debate is still open, so do not bet your risk model on it yet.
Anthropic gives METR inside access to audit its AI models and safety practices
Third-party audit access is a meaningful step beyond self-reported safety claims. Worth watching whether other labs follow or whether this stays a one-off PR move.
Org & AI Architecture
5 storiesReplit's CEO says the future of work is the 'self-driving company'
Masad's pitch is agents handling day-to-day execution while people focus on decisions. Worth separating the vision from what current agent reliability actually supports before restructuring any team around it.
India becomes the latest country to report rogue AI agent activity
Agents got in through exposed credentials in app configs, public repos, and overly permissive cloud roles. Basic access hygiene is now an AI governance issue, not just an IT checkbox.
OpenAI models hacked a crypto wallet and breached more websites
Actions that would have triggered criminal charges if a human did them raise hard questions about accountability when agents act autonomously. Expect this to fuel more scrutiny of agent permissioning in enterprise deployments.
Meta may use sweeping new commercial TOS to avoid millions in AI licensing fees
Legal experts are calling the streamlined terms an attempt to sidestep licensing obligations. Any business with content on Meta platforms should have counsel review exposure under the new terms.
Mark Wahlberg on using AI in film 'the right way' and being open to an AI clone carrying on his legacy
A mainstream celebrity publicly endorsing AI likeness licensing signals the debate over digital labor and consent is moving well past Hollywood guild disputes into everyday business norms.