The Manifest
Anthropic is having a rough weekend with a major music-industry lawsuit and a Claude Code usage cut disguised as a raise, while trucking data shows capacity exit rather than demand growth is driving the freight cycle and OpenAI pulls Cursor's access over its SpaceX acquisition.
Get The Manifest in your inbox
One sharp digest of AI × supply chain. Free. No noise.
Today's Top
- 01Sony and Warner sue Anthropic over 'one of the largest and most blatant ongoing thefts of intellectual property in history'The Decoder
- 02Supply driven trucking market cycle explained in the dataFreightWaves
- 03OpenAI cuts off Cursor after SpaceX acquisition, citing Musk's history of breaking contractsThe Decoder
- 04Anthropic Opens a Research Preview of the Model Hardware Standard (MHS): A Shared Specification for AI Agents to Safely Operate Physical DevicesMarkTechPost
- 05Google AI Releases Gemini Omni 1.1 Flash: 40-Second Scene Extension, First/Last Frame Control, and 4K UpscalingMarkTechPost
Models & Releases
4 storiesAnthropic's Claude Code limit change is a raise on paper but a cut in practice
A temporary usage boost expires and gets replaced by a smaller permanent one, a net 17 percent cut dressed up as a policy improvement. Teams running Claude Code in production workflows should budget for lower weekly caps after September 14.
Google AI Releases Gemini Omni 1.1 Flash: 40-Second Scene Extension, First/Last Frame Control, and 4K Upscaling
Longer scene extension and pinned first/last frames make this a more usable production tool than a demo toy. Worth a look for teams doing training or marketing video who want more control without a full studio pipeline.
Hugging Face Unveils Microduck: A $399 Open-Source 25 cm Biped You Train with Reinforcement Learning
A full sim-to-real robotics stack for under $400 lowers the bar for prototyping automation ideas in-house. Not a warehouse-ready robot, but a cheap way for ops teams to get hands-on with reinforcement learning before committing to bigger platforms.
Google's WikiSkill gives AI agents a persistent memory of past mistakes to sharpen future performance
Agents that document failures and successes instead of forgetting them after every run could make smaller, cheaper models good enough for planning and procurement tasks. Worth watching before betting on the biggest model for every agentic workflow.
Supply Chain & Ops
3 storiesSupply driven trucking market cycle explained in the data
Tender volumes are flat and rejections have tripled, which points to capacity leaving the market rather than freight demand actually picking up. Planners should track rejection rates as the real signal instead of assuming a demand recovery.
Broker Transparency Proposal Clears FMCSA and Heads to the White House
The rule has cleared its last internal checkpoint after missing two earlier deadlines, so publication is finally within sight. Brokers and shippers should start reviewing contract and data-sharing practices now rather than waiting for the final text.
Anthropic Opens a Research Preview of the Model Hardware Standard (MHS): A Shared Specification for AI Agents to Safely Operate Physical Devices
A shared driver spec that lets agents control lab and robotic hardware cut integration time from weeks to hours in early tests. Promising for manufacturing and warehouse automation, but agents still misread physical cause and effect, so keep humans in the loop for now.
Deals & Market
4 storiesSony and Warner sue Anthropic over 'one of the largest and most blatant ongoing thefts of intellectual property in history'
This lands just months after Anthropic's $1.5 billion book-author settlement, and it expands the legal exposure into music. Anyone evaluating AI vendors on training-data provenance should treat this as a live risk, not a settled one.
'We're not doing 30 bets a year': Vijay Pande on betting small after running $4 billion at a16z
A former mega-fund biotech investor is now making fewer, more concentrated AI-native bets and arguing open shared datasets, not walled-off ones, will actually move biology from discovery to engineering. Relevant for anyone in life sciences procurement watching where data access is headed.
OpenAI cuts off Cursor after SpaceX acquisition, citing Musk's history of breaking contracts
Access to AI coding tools can disappear overnight for reasons that have nothing to do with your own usage or payment. A reminder to keep a backup model provider in any workflow you depend on daily.
Nvidia's AI advantage is moving beyond the GPU
The next efficiency gains are coming from smarter data center traffic control, not just more chips. Buyers sizing up AI infrastructure spend should weigh networking and orchestration as heavily as raw GPU counts.
Research & Frontier
3 storiesMeet 'Code-as-World': An Agentic Loop That Rewrites Real Videos Into Executable MuJoCo Physics Programs
Turning ordinary video into editable physics simulations could cut the cost of building training environments for robotics and automation research. Early stage, but worth tracking if your team is exploring simulation-based training for physical operations.
Building Custom Batched Ensemble Weather Forecasting with NVIDIA Earth2Studio
Open tooling for ensemble weather forecasting matters for anyone whose planning depends on wind, storm, or temperature risk, from renewable energy scheduling to freight routing around severe weather.
LAION drops massive open video dataset with 10 million hours of footage for AI research
An 80-million-video open dataset lowers the barrier to training competitive vision models outside the big labs. Useful context for anyone tracking how much control a handful of companies actually have over the data pipeline behind AI video tools.
Org & AI Architecture
3 storiesThe skills that earn top grades are the ones AI can fake best
GPT-4o boosted grades on a marketing assignment by nearly a full point without anyone checking if students actually learned anything. A caution for any organization leaning on AI-assisted assessment or training programs to measure real skill, not just output quality.
AI-generated videos are already displacing actors and livestreamers across China's entertainment industry
95 percent of short dramas released in China this year were AI-generated, and labor disputes over voice and likeness rights are climbing fast. A preview of the contract and ownership fights headed toward other AI-driven content industries.
AI agents have no sense of time and are not aware of it
Coding agents overestimate task duration by up to 10x and rate their own work about 20 points too high. That gap between confidence and accuracy is exactly why long, autonomous agent runs still need a human checking the clock.