May 1, 2026

TL;DR. Frontier labs are pouring tens of billions into RL data, but they're really buying research outcomes — not datasets — and almost nobody can actually deliver that at scale.

Takeaways

• Anthropic's $1B RL data budget will be "far exceeded" this year; OpenAI is in the same zone, and the window for "banal data projects" (Mercor-style) is rumored to close entirely in 2026.
• What labs pay for is hillclimbing: how directly your data moves SOTA benchmarks and saves the lab's own acquisition effort — not a reusable product.
• The market is still greenfield because few teams have the sophistication for on-demand, consistently QA'd research services; that's the moat, not the data itself.
• Mercor overtook Scale on the semblance of this product; upstarts are now using the same playbook against Mercor/Handshake/Surge.

TL;DR. An OpenAI staffer stitched daily Codex tasks into "Lord Bottleneck," an autonomous agent that now runs growth experiments end-to-end every morning.

Takeaways

• Don't try to automate the whole thing — use Codex to accelerate individual steps first, then connect them into a skill once they work.
• The unlock was telling Codex "do this every morning" — turning a manual loop into a standing agent that proposes experiments, writes code, runs them, and reports back.
• Named after the bottlenecks it kills (new user friction), it's reportedly generated "really serious" company value autonomously.

TL;DR. Cold DMs are a hell-yes-or-no game — kill the hedge, lead with negative curiosity, and write for the recipient's AI inbox.

Takeaways

• Subject lines should trigger negative curiosity and name-drop other investors if you want a GP to actually open it.
• Match the message to the decision-maker (GP, not the analyst), keep it short, and never bury the ask.
• Answer 3 personalization questions before sending; follow up even after a hard no.
• "Yo bro wanna collab," making it about yourself, and hedging your ask are the fastest ways to torch your reply rate.

TL;DR. Hype thread pitching a 1-hour lecture from a Jane Street junior who allegedly landed a $220K–$600K role using AI to mine trillions of data points.

Takeaways

• Classic "skip Netflix, watch this" engagement-bait framing — treat the salary claim as marketing, not verified fact.
• Promises a walkthrough of dataset research, AI pattern-finding, and turning raw data into decisions; worth a skim only if the actual video link delivers a real workflow.