Things I read

“And what I'm interested in is investing in people.” — Arthur Rock VC investing isn’t what it was. By that I mean, the majority of investment isn’t done by a loose partnership of generalists investing

Y Combinator is looking for startups working on these ideas. If you're working on something we're interested in, we'd love to hear from you.

Discover how untrusted open-weight AI models can potentially leak data, even when self-hosted. Learn the security risks and how to protect your enterprise.

Gonzo first-person report from a Chelsea "peptide party" sponsored by needle-free peptide startup afterglow, mingling with tech-adjacent biohackers, event planner "Birb," and Aella's sex-network app Glosso; the author questions a hedonism promising sin without consequence.

Iconiq, Menlo, & Lightspeed are set to be the biggest venture winners in the upcoming IPO. None have more than 2% of the company.

Co-authored with Antaripa Saha. Decision models have exploded onto the scene recently, starting with TypeSafe’s Jev, followed by Perplexity’s Decider, Liquid’s d1, Cloudflare’s Clef, and OpenAI’s

An Anthropic researcher just told a room full of PhD students that their career plan no longer exists His 3 pieces of advice for students: 1. Use AI every day. If you're on the exponential, you'll never fall behind it. 2. Choose projects that get better when compute scales. The rest will be solved before you finish them. 3. Stop planning your career as if it were stable. The safe option no…

Google Research repo/paper for Harness², a framework that optimizes an agent's scaffolding (prompts, tools, memory, skills) rather than model weights, via a Propose–Probe–Compose loop plus persistent cross-task lessons; reports gains up to ~11.5 pts on WorkBuddy-Bench, JobBench and LAB.

In his book The Hard Thing About Hard Things, Ben Horowitz divides the life of a company into two conditions. Peacetime is when a company has a big lead over its competitors in a market that is

Nick Cammarata argues mechanistic interpretability is in better shape than feared: polysemanticity/superposition in InceptionV1 and LLMs was largely an artifact of bad architecture and weak tools, and apparent mixing is "recursive generalization" — so networks can be read directly.