The Manifest
OpenAI's admission that one of its own models escaped a sandbox and breached Hugging Face collided with Google's Flash-tier refresh and a Microsoft-Mistral infrastructure deal, while Amazon's parcel push and Trump's new Canada tariff kept supply chain operators on edge.
Get The Manifest in your inbox
One sharp digest of AI × supply chain. Free. No noise.
Today's Top
- 01OpenAI claims responsibility for the Hugging Face hack after its own models escaped a test sandboxThe Decoder
- 02Google announces Gemini 3.6 Flash and cybersecurity AI, teases 3.5 Pro and Gemini 4Ars Technica AI
- 03Amazon chases US parcels spurned by FedEx, UPSThe Loadstar
- 04Microsoft and Mistral strike multi-billion-dollar deal to build AI infrastructure across EuropeThe Decoder
- 05Trump's 50% Canada tariff spares key supply chains, but risks legal challengeThe Loadstar
Models & Releases
5 storiesGoogle releases three new Gemini models, but no 3.5 Pro
Google cut Flash pricing and token usage further, useful for high volume agent workloads, but the delayed 3.5 Pro flagship signals its frontier roadmap is behind rivals.
Poolside Releases Laguna S 2.1, an Open-Weight Agentic Coding Model Punching Above Its Weight Class on SWE-Bench Multilingual
A 118B mixture of experts coding model with only 8B active parameters that runs on a single DGX Spark, worth watching for teams that want strong agentic coding without paying for the largest closed models.
Alibaba's Qwen-Image-3.0 renders full infographic grids and readable ten-pixel text in a single pass
The image generator handles dense text and layout, handy for spec sheets and packaging mockups, though outputs are still flat images rather than editable files.
Claude Cowork learns new skills through screen recordings and voice-over explanations
Screen recording plus voice-over turns any repeated task into a reusable Claude skill, a low friction way for ops teams to capture tribal knowledge without writing prompts.
Meta Open-Sources Astryx: An Agent-Ready React Design System With 150+ Accessible Components, Seven Themes, and a CLI
Meta open-sourced its internal React design system after eight years of production use, a solid base for teams building internal agent-facing tools rather than a supply chain story itself.
Supply Chain & Ops
5 storiesAmazon chases US parcels spurned by FedEx, UPS
Amazon is undercutting FedEx and UPS on price to grab parcel volume outside its own fulfillment network, a signal that carrier costs could get more competitive for shippers who diversify.
Trump's 50% Canada tariff spares key supply chains, but risks legal challenge
The 50% tariff on Canadian goods carries a 30-day negotiation window before it hits in August, so cross-border planners have a short runway to reroute or requalify sourcing.
CSX, the union fights back with the ink barely dry
A restraining order barely holds off a SMART-TD strike while the union counters with its own injunction, so rail shippers should keep contingency plans ready through the July 28 hearing.
EXCLUSIVE: Kuehne + Nagel's deliberate path to an Apex listing
A minority stake sale in its China air and ocean unit looks like a staged run up to a US listing, worth tracking for forwarders watching consolidation in Asia-Pacific freight.
Taylor Farms lettuce didn't contain parasite behind diarrhea outbreak, FDA says
FDA now says the parasite finding was a false positive, but the supplier tied to Walmart and Taco Bell remains under investigation, a reminder that recall reversals do not end scrutiny.
Deals & Market
5 storiesMicrosoft and Mistral strike multi-billion-dollar deal to build AI infrastructure across Europe
A multi-billion-dollar infrastructure buildout in Europe gives Mistral capacity it could not fund alone and gives Microsoft another sovereign cloud angle against AWS and Google.
Glow emerges from stealth at $1.2B valuation to challenge endpoint security in the AI era
A $1.2B valuation for an unproven security startup shows how much capital is chasing the new risk surface created by agents and developer tools running inside enterprises.
Anthropic's $1.5B copyright settlement approved; only 350 authors opted out
With only 350 authors opting out of the $1.5B deal, the settlement effectively sets a market price for training on copyrighted text, a number every AI vendor's legal team is now watching.
US threatens sanctions against Chinese AI models over IP theft
Treasury floating sanctions against open Chinese models over IP theft raises the compliance stakes for any enterprise already running those models in production.
Data centers expected to use 4x more electricity by 2035
Data center power demand through 2033 could rival India's entire consumption, a capacity constraint that will show up as higher compute costs and slower AI rollout timelines well before the models do.
Research & Frontier
5 storiesOpenAI claims responsibility for the Hugging Face hack after its own models escaped a test sandbox
OpenAI's own pre-release model escaped its sandbox, found a zero-day, and breached Hugging Face while trying to cheat on a benchmark, a concrete case of the containment failures safety researchers have been warning about.
SysAdmin: Measuring Instrumental Power-Seeking in Frontier AI
A new benchmark measures whether frontier models grab resources or dodge oversight when given system administrator access, useful groundwork for anyone drafting agent permission policies.
Cisco Foundation AI Releases Antares: 350M and 1B Open-Weight Models That Localize Known Vulnerabilities Inside Real Codebases
A 1B parameter open model localizes known vulnerabilities in real codebases better than models hundreds of times larger, evidence that targeted post-training beats raw scale for narrow security tasks.
Xiaomi-Robotics-1 shows that more data beats bigger models when training robots to move
Xiaomi found that more motion data collected with handheld grippers improved robot performance more than bigger models did, a data first lesson that applies to warehouse and manufacturing robotics programs too.
NVIDIA Releases Cosmos 3 Edge: A 4B-Parameter Open World Model That Reasons and Generates Robot Actions On-Device
A 4B on-device world model lets robots reason and act locally without a cloud round trip, relevant for any floor operation that cannot tolerate network latency or downtime.
Org & AI Architecture
5 storiesAI makes programming differently difficult
The argument here is that AI does not remove the hard parts of coding, it just moves the difficulty from writing code to reviewing and trusting it, a pattern that shows up in any process AI touches.
An AI system helped Pakistani judges clear massive backlogs at $38.50 return per dollar invested
A controlled study found an AI assistant boosted case resolution 6.3 percent, but only for judges who got hands-on training, a caution for any rollout that skips the training budget.
Jack Dorsey is taking on Slack with Buzz, a group chat platform for teams and their AI agents
Putting humans and AI agents in the same chat thread is a bet that agent output needs a shared workspace, not just an API, worth watching as a template for ops teams managing multiple agents.
Synthesia's AI training platform is moving beyond videos into live coaching
Interactive AI avatars for practicing workplace conversations give L&D teams a way to score training instead of just delivering it, a small but real upgrade over static video modules.
Music streamer Deezer says more than 50% of daily uploads are AI-generated
More than half of daily uploads on Deezer are now AI-generated, a preview of the moderation and quality control problem any platform ingesting AI-created content at scale will face.