Welcome back! Jury selection in Elon Musk's lawsuit against Sam Altman opened yesterday with a witness list that puts more than a decade of OpenAI's internal decisions into sworn testimony. While the trial began, OpenAI tore up its Microsoft contract and gained the right to ship on AWS, China ordered Meta to stop a $2B AI deal that was already complete and the researcher behind AlphaGo raised $1.1B to build AI that learns without human data.
In today’s Generative AI Newsletter:
Musk v Altman: What does the witness list reveal about what is actually on trial?
OpenAI x Microsoft: Why is AWS the real winner of the renegotiated contract?
China v Meta: What happens when Beijing treats AI startups like semiconductors?
Ineffable Intelligence: Can AI learn forever without ever touching human data?
Latest Developments
Musk v Altman trial forces a decade of OpenAI into the public record
Jury selection in Elon Musk's lawsuit against Sam Altman opened yesterday in San Francisco, with Musk asking for up to $134B in damages and the unwinding of OpenAI's for-profit conversion. The witness list is the part to watch.
The details:
Damages claim: $134B redirected back to the OpenAI nonprofit, which Musk co-founded with Altman in 2015 before departing the board in 2018.
Witness list: Microsoft CEO Satya Nadella and a roster of former OpenAI executives, with Altman and Musk both expected to take the stand.
Scope of testimony: Years of internal decisions on safety reviews, the 2023 board ouster, the for-profit restructuring and the company's commercial pivot all enter sworn record.
IPO context: OpenAI and xAI are both preparing public offerings, with the trial timeline running alongside both filings.
What enters the record is the harder cost. People who were inside every controversial OpenAI decision will be under oath while the company prepares to go public. This means a pre-IPO market reads the same transcripts as the jury.
Special highlight from our network
Plurai just introduced a new method called vibe training. It focuses on one goal: making your agents reliable in production.
Here’s how it works.
A small model is trained on your exact use cases. Agents simulate debates around those scenarios, improving how the model evaluates outputs before deployment.
The results are hard to ignore:
8× more cost-effective than GPT-5 mini
Real-time inference (<100 ms)
43% fewer errors vs GPT-5 mini
Why this matters:
General models struggle with consistent evaluation. Specialized models trained on your workflows perform better where it counts.
OpenAI's new Microsoft deal hands AWS the $50B Bedrock runway
OpenAI and Microsoft rewrote their partnership terms, ending Microsoft's exclusivity over OpenAI's IP and removing the AGI clause that would have ended Microsoft's revenue share. OpenAI gains the right to ship products on any cloud.
The details:
Microsoft: Loss of IP exclusivity and the AGI trigger, in exchange for revenue share through 2030 and Azure-first launch access through 2032.
AWS: OpenAI is now free to deploy on Amazon Bedrock, the platform that just signed a $50B compute deal with the company earlier this month.
Jassy: AWS CEO Andy Jassy called the announcement "very interesting" hours after it landed.
AGI clause: Originally written to end Microsoft's claims when OpenAI declared it had reached general intelligence. Replaced with calendar dates because nobody could agree on what would set it off.
Microsoft kept the cash and the Azure-first launch slot. AWS came away with the right to host the models enterprise customers actually want. Teams already standardized on Bedrock can stay on Bedrock for OpenAI workloads, and the strongest practical reason to migrate to Azure for those workloads just disappeared.
China blocks Meta's $2B Manus deal
China's National Development and Reform Commission reclassifies AI startups as export-controlled assets. They ordered Meta and Manus to stop the $2B acquisition Meta announced in December. The order does not stand alone.
The details:
Pattern: Chinese regulators have also instructed AI labs Moonshot and Stepfun to refuse US capital, making three such directives in three months.
Singapore: Manus relocated there specifically to escape Beijing's reach. The veto shows that domicile does not protect a startup whose talent and IP are Chinese in origin.
Travel: Manus executives are reportedly barred from leaving China during the probe.
Unwind: Meta said the transaction "complied fully with applicable law" and the two teams are already integrated, with Manus's site reading "now part of Meta."
China is now applying to AI labs the export-control logic the US applies to chips. The talent and the model weights are national assets that stay inside the country regardless of corporate structure.
Ineffable Intelligence raises $1.1B to build AI without human data
David Silver, the DeepMind researcher behind AlphaGo and AlphaZero, launched Ineffable Intelligence with a $1.1B seed at a $5.1B valuation. The London-based lab is betting that AI learns better from experience in simulation than from scraped human data.
The details:
Track record: Silver led DeepMind's reinforcement learning team for a decade and was the lead researcher on AlphaGo, AlphaZero, AlphaStar and AlphaProof.
Method: Ineffable's models skip pre-training on human data. Agents learn through repeated experience in simulated environments.
Framing: Silver describes human data as "a kind of fossil fuel." His approach is "a renewable fuel, a model that can just learn and learn and learn forever."
Round size: $1.1B at a $5.1B valuation, the largest seed round in European history.
AlphaZero already proved a version of the thesis by learning chess and Go from scratch through self-play. The open question is whether reality is game-like enough for the same approach to scale into messier domains where human data has so far been mandatory. The next two to three years will tell whether the renewable fuel argument holds outside a board game.
Tool of the Day: Lovable mobile
Lovable is an AI app builder that turns plain-language prompts into working web apps. The new mobile app brings that flow to your phone, letting anyone spin up a prototype and share a link without a laptop.
Try it yourself: Install Lovable from the iOS or Android store and sign in or create a free account. Tap "New project" and describe what you want to build in a sentence or two, the same prompt style you would use on the desktop site.
Lovable generates a live preview you can hand off via URL, and follow-up prompts let you tweak layout, copy and behavior on the fly. This turns a coffee break into a working prototype, even for non-technical people who want to capture an idea the moment it lands.
Light Bytes
OpenAI hardware: OpenAI is reportedly working with MediaTek, Qualcomm and Luxshare on a phone targeting 2028 mass production.
Adobe Firefly: Adobe opened its multi-app Firefly AI Assistant to public beta, letting Creative Cloud subscribers prompt across their tools while keeping outputs editable.
GPT-5.5 in Arena: GPT-5.5 went live on the Arena AI leaderboard, where users can compare its outputs head-to-head against other frontier models.
Trademark filings: Taylor Swift filed three federal trademarks for her likeness and voice, joining Matthew McConaughey in legal moves to block AI-generated impersonations.
Meta's solar bet: Meta agreed to buy up to a gigawatt of solar power from Overview Energy, a startup developing satellites that beam solar electricity to Earth, with an in-space demo targeted for 2028.
