AI Builders Digest

Personal agents will be a bigger shakeup than the App Store

The form factor of something like Muse means handing off a task and trusting it gets finished end to end, which requires the agent to operate your tools or bring its own. That flips who you compete for: not the user's attention but the agent's. Whoever builds the best rails for agents to order food, run ecommerce transactions, book flights, and touch local commerce gets used most. Levie calls it the biggest opportunity and disruption in consumer tech since the App Store itself.

• #agents
• #products

Richard Socher: anything you can simulate, AI will solve

The premise behind Recursive's $650M raise is that scientific progress has stalled because knowledge fractured into thousands of subfields nobody can cross, and AI is the tool that weaves them back together. Socher's sharpest heuristic: superhuman capability arrives wherever you have a simulator or a verifier, which is why math and programming fall first and biology waits on data. He is explicitly not a hard-takeoff believer, since clinical trials still take years no matter how good the molecule. Two contrarian threads: hallucination is a feature when you want novel proteins, and his own prompt-engineering paper (DecaNLP) was rejected by essentially every reviewer as useless. "Boy, are we far away from the true upper bounds of any of the spaces of intelligence."

• #agents
• #research
• #funding

A VC spends two hours a day coding and gets asked why daily

Kothari's answer to the DM he gets every day: things move too fast now to learn AI secondhand. He breaks investing in this era into three things you have to feel firsthand, not hear about — model capabilities so you can see where the next generation improves, the infra tools models need for long-horizon tasks, and the harness capabilities that make autonomous companies possible. Talking to researchers and founders still works, he just never learned well without getting his hands dirty. His pitch: tinker daily for ten days and you'll be hooked.

• #agents
• #funding

Agent announcements imply undisclosed nation-state breaches

Reading the wave of agent releases from Google, OpenAI, and Anthropic, Agarwal takes it as a certainty that multiple major companies have already been penetrated by nation-state actors without knowing it. The reasoning is implicit in the timing: the capability that makes agents useful is the same capability that makes intrusion cheap and quiet.

• #security
• #agents

The best part of this AI wave is that everyone wants to build

Rauch's read past the short-term drama: the defining feature of the moment is that people are excited to create and ship rather than just consume, and that is an immeasurably good development. He paired it with Reagan's 1989 shining city passage about walls with open doors.

• #products

A markdown file plus a skill turned into a weekly taste engine

Yang keeps his movie, TV, and book ratings in a plain markdown file and runs a /tastemaker skill over it, which scans new releases weekly and surfaces what he hasn't seen. He's asking whether to open source it. He also teased an upcoming episode with the product lead for ChatGPT Finances at OpenAI covering six personal finance prompts for proactive tax savings, credit card rewards, and interactive forecasts toward financial independence.

• #agents
• #products

Muse and Instinct both refuse to place calls

Yu tried getting both agents to make phone calls and each said it couldn't, asking publicly whether the capability is still rolling out. A small data point on the gap between announced agent capabilities and what's actually live.

• #agents
• #products