Presented by
Good morning, AI enthusiasts. The three-week legal battle many expected to shake OpenAI to its core ended in 90 minutes, and not the way Elon Musk intended.
A jury in Oakland found that Musk waited too long to file his case, and Judge Yvonne Gonzalez Rogers agreed on the spot. Musk has already promised an appeal to the Ninth Circuit. Has the most disruptive lawsuit in AI history finally found its floor?
In today's recap:
Musk's OpenAI lawsuit dismissed after 90 minutes
Dell and OpenAI bring Codex on-premises
Navigate large codebases fast with Claude Code
Cursor ships Composer 2.5 with new RL training
4 new AI tools, prompts, and more
OPENAI
Jury rejects Musk's lawsuit against OpenAI
Getty Images
Recaply: A federal jury just dismissed Elon Musk's lawsuit against OpenAI after just 90 minutes of deliberation. The jury found he waited too long to sue Sam Altman and Greg Brockman, with the case barred by the statute of limitations.
Key details:
Musk filed the suit in February 2024, but the jury ruled he knew about the disputed behavior as early as 2021, placing his claim well outside the legal window.
Musk co-founded OpenAI and gave $38M in early funding before departing the board in 2018. The lawsuit alleged Altman and Brockman "stole a charity" through OpenAI's for-profit transition.
Judge Yvonne Gonzalez Rogers accepted the jury's findings as her own, saying there was "substantial evidence" behind their decision and that she had been prepared to dismiss the case on the spot.
Musk told the court earlier this month, "I was a fool. I gave them free funding to create a startup." He is planning an appeal to the Ninth Circuit Court of Appeals.
Why it matters: Musk's lawyers argued this case was about protecting charities from being looted for private gain. OpenAI's attorneys called it a competitive attack filed years past its deadline. The jury agreed. For OpenAI, the verdict removes the biggest legal threat to its planned IPO and recent corporate restructuring. Musk says he'll appeal to the Ninth Circuit, which means the final chapter of this story hasn't been written.
PRESENTED BY HYPERAGENT
Run AI agents that ship real work
Recaply: Most AI tools give you a chat window. Hyperagent gives AI a real computer with browser access, shell, code execution, and hundreds of integrations per task, so agents can finish work instead of just describing it. Built by the Airtable team.
With Hyperagent, you get:
Agents that run in a real computer, not a chat window
Memory that builds over time so agents get sharper on your work
hundreds integrations, no glue code needed
Already in production. A DTC brand runs 14 agents via Slack. An insurance brokerage runs 11 for compliance work, 50+ times a day.
Apply to Founding500 and get $10M in AI credits for agent-first companies. Applications close May 31.
DELL
Dell brings OpenAI Codex on-premises for enterprises
OAI
Recaply: Dell Technologies just announced a partnership with OpenAI to bring Codex to hybrid and on-premises enterprise environments, connecting the AI coding agent to the Dell AI Data Platform and Dell AI Factory where most enterprise data already lives.
Key details:
Codex connects to Dell's AI Data Platform, letting it access internal codebases, documentation, business systems, and operational workflows without requiring enterprise data to leave on-premises infrastructure.
Codex is one of OpenAI's fastest-growing enterprise products, with over 4M developers using it weekly. Dell AI Factory 2.0 now anchors partnerships with both OpenAI and Google to power enterprise AI workloads.
According to Ihab Tarazi, SVP and CTO of Dell's Infrastructure Solutions Group, the partnership lets enterprises "deploy AI where enterprise data already lives," giving customers "a practical, secure path to deploying AI agents at scale."
Codex access via Dell's AI Data Platform is available now. Integration with the Dell AI Factory, including ChatGPT Enterprise and other OpenAI API solutions, is still being explored.
Why it matters: Dell is tying together its AI Factory 2.0 with both OpenAI and Google at the same time, which signals it's playing for the whole enterprise AI stack. Most AI adoption inside large organizations stalls because sensitive data can't leave on-premises environments. By meeting companies where their data already lives, this partnership opens Codex to sectors, such as finance, healthcare, and government, that have largely been locked out of agentic AI until now.
GUIDES
Navigate any large codebase fast with Claude Code
Recaply: In this tutorial, you will learn how to configure Claude Code for a large or unfamiliar codebase using CLAUDE. md files, hooks, and skills, so it can find what it needs without burning through your context window.
Step-by-step:
Start in the subdirectory, not the repo root. Go to the folder where you're actually working and run
claude /initto let Claude generate a starter CLAUDE. md file scoped to that part of the codebase.Keep the root CLAUDE. md lean: one line per top-level folder describing what lives there. This gives Claude a directory map to scan before opening files, instead of searching blindly across millions of lines.
Specify test and build commands per subdirectory in each CLAUDE. md file. Running the full suite across a monorepo when Claude changed one service wastes context and causes timeouts. Scope the commands to what applies locally.
Install a code intelligence plugin for your primary language in Claude Code Settings, then enable the Language Server Protocol. Claude can then search by symbol instead of by string, filtering thousands of grep matches down to only the references pointing to the same function.
Test navigation by asking Claude to trace a call path across the codebase. If it opens irrelevant files or hits context limits, add a
.claude/settings.jsonfile withpermissions.denyrules to exclude build artifacts and generated code.
Pro tip: Add a stop hook that prompts Claude to suggest CLAUDE. md updates at the end of each session while context is fresh. This turns every coding session into a setup improvement, and the configuration gets sharper the more you use it.
TOGETHER WITH 1440 MEDIA
Daily news for curious minds.
Be the smartest person in the room. 1440 navigates 100+ sources to deliver a comprehensive, unbiased news roundup — politics, business, culture, and more — in a quick, 5-minute read. Completely free, completely factual.
CURSOR
Cursor ships Composer 2.5, its most capable model
Cursor
Recaply: Cursor just shipped Composer 2.5, its most capable coding model yet, calling it up to 10x more efficient than comparable models and built on Moonshot's Kimi K2.5 open-source checkpoint with a new RL training technique called targeted textual feedback.
Key details:
Composer 2.5 uses targeted textual feedback during RL training, inserting correction hints at specific points in the model's reasoning rather than applying one reward over the full rollout, letting it fix behaviors like bad tool calls and communication style.
Cursor trained Composer 2.5 on 25x more synthetic tasks than Composer 2. It's partnering with SpaceXAI to train a significantly larger model using 10x more compute on Colossus 2's million H100-equivalents.
The training team found Composer 2.5 attempting creative workarounds during training, including decompiling Java bytecode to reconstruct a deleted function signature, illustrating what Cursor calls "the increasing care necessary for large-scale RL."
Composer 2.5 is live in Cursor now across all plans. Cursor is doubling included usage for the first week of availability.
Why it matters: Cursor is building its own models instead of relying on frontier providers, which is a meaningful shift for a coding tool that started as a model wrapper. Targeted textual feedback lets it fix behaviors that benchmarks don't capture, like how hard the model works and how it explains its thinking. The SpaceXAI partnership on a next, larger model suggests Cursor is treating model capability as a real competitive moat, not a product feature.
TOOLS
Trending AI Tools
⚙️ Composer 2.5 - Cursor's most capable coding agent, built on Kimi K2.5 with new targeted RL training and up to 10x more efficient than comparable models
🎥 Creatify Agent - Multi-agent orchestrator that researches competitors, writes direct-response copy, and generates video ads from a single prompt
🤖 Agora-1 - Odyssey's multi-agent world model letting up to four human or AI participants interact inside the same real-time simulation
🧪 Starchild-1 - Odyssey's first real-time multimodal world model, generating interactive world simulations with audio for the first time
NEWS
What Matters in AI Right Now?
Anthropic acquired Stainless, the SDK and MCP server platform that has powered every official Anthropic SDK since the earliest days of its API. Hundreds of companies use Stainless to generate type-native SDKs, CLIs, and MCP servers across TypeScript, Python, Go, Java, and Kotlin.
Anthropic launched self-hosted sandboxes in public beta and MCP tunnels in research preview for Claude Managed Agents, letting enterprises run agents inside their own infrastructure. Supported providers include Cloudflare, Daytona, Modal, and Vercel.
OpenAI partnered with Dell to deploy Codex in hybrid and on-premises enterprise environments via the Dell AI Data Platform. Codex is already one of OpenAI's fastest-growing enterprise products, with over 4M weekly developers.
Odyssey released Agora-1, the first multi-agent world model letting up to four human or AI participants interact inside the same real-time simulation, alongside Starchild-1, the first real-time multimodal world model to generate interactive simulations with audio.
Sapient Intelligence introduced HRM-Text, a 1B-parameter reasoning model trained on $1K of compute and 40B structured tokens using a brain-inspired hierarchical latent architecture, achieving competitive performance with roughly 1/1,000 of the training data of comparable models.
🧡 Enjoyed this issue?
🤝 Recommend our newsletter or leave a feedback.
How'd you like today's newsletter?
Cheers, Jason
