AI Builders Digest

The trillion-dollar cohort was punctuated equilibrium, not the new normal

Three companies went from roughly zero to a trillion in five years, and investors are now pricing dozens of follow-ons at that same velocity. The pushback: the market size may well be there, but the speed almost certainly isn't, and people are conflating the two. Inside the labs, compute scarcity has quietly turned into a human power law, with a few dozen researchers driving most results and hiring slowing because the cost isn't the researcher, it's the compute attached to them. The counterweight to lab RSI hype: smart, self-aware scientists have believed recursive self-improvement was eighteen months out, every eighteen months, for five years running. On when to sell: schedule the conversation on the calendar so it isn't emotional, because "your most productive years of your life are on the line right now."

• #funding
• #policy
• #agents

Nobody in AI has a moat except the venture firms

A layer-by-layer list of the whole stack, models, IDEs, harnesses, app builders, wrappers, inference, voice, data labeling, infra, neoclouds, generative media, each named with three companies and each declared moatless. The punchline is that the only durable franchise left is the venture firm underwriting all of it. Separately, a bet that brand marketing becomes the prized differentiating asset over the next decade, not launch videos but a sustained weekly answer to what a company stands for. If agents become the primary user of most products, standing out to the remaining human audience gets harder, and the marketers repeatedly passed over for CMO roles start getting cofounder seats.

• #funding
• #products

Gemini ships a ten-item fix list: Workspace tools, Projects, 49 connectors

A public status update against a prior list of user complaints, with six items marked done. Revamped Workspace tools land in one to two weeks, the new Projects design is finished and being implemented, and connector support is at 49 and climbing. 3.7 Flash showed improvements in tool calling with more coming, and the biggest over-triggering bugs are fixed. Notable mostly for the format: a roadmap graded item by item in public rather than announced.

• #products

Garry Tan open sources his 70-skill "Personal AGI" agent repo, MIT licensed

The setup from his Startup School talk is now free and public: a private GitHub repo seeded with 70 proven skills plus the start of a Karpathy-style knowledge wiki. It runs on an existing Claude Code or Codex subscription with no extra spend, works from the command line or desktop, and bootstraps by pasting a single image into the harness. Everything is MIT licensed.

• #open-source
• #agents

How to actually get good at evals: start from a workflow you already know cold

Pick a workflow you understand deeply and make its quality measurable, rather than starting from an eval framework. That means studying real traces, the prompt sequences typical users produce, what a good response looks like at each step and end to end, then deliberately manufacturing traces from your product's failure modes: messy tool call responses, missing context. Only after the eval is good does automation matter, and the harder ongoing problem is keeping the eval mirroring live traffic as user patterns drift.

• #evals
• #agents

Coding models are beating diffusion models at creative work

Recent procedural generation art, video editing, and 3D game demos are an update toward LLM coding models outperforming diffusion models across a lot of creative territory. The reason is structural rather than aesthetic: code is easier to edit, easier to nudge in a specific direction, and exports cleanly into existing tools, none of which is true of a generated pixel buffer. Backed with a live demo of the /design command in Claude Code.

• #products
• #agents

Data belongs on the balance sheet as an asset now, not in a footnote

"Data is the new oil" finally has a concrete referent: AI's thirst is deep enough that information is valuable in almost any form, and a recent data sale shows what that repricing looks like in practice. The implication is accounting as much as strategy, treating an organization's information as a balance sheet asset rather than exhaust. How well a company mines its own institutional intelligence becomes a deciding factor in competitiveness.

• #policy
• #products

Trajectory on continual learning: GRPO wasn't enough, so they went on-policy

A rundown of Trajectory's approach to the remaining data problems in continual learning, including why GRPO alone doesn't get there and what forced the move to on-policy training, plus the follow-on work of fixing everything that on-policy then breaks. Framed as one of the early leaders in the field being unusually specific about tradeoffs rather than results.

• #agents
• #evals

Cursor Origin hosts repos and deploys to Vercel, and is itself on Vercel

You can now keep repos in Cursor Origin and deploy from there to Vercel, closing a loop where Cursor Origin runs on Vercel itself. Delivered with a jab at GitHub's uptime.

• #products

Scanning code for vulnerabilities isn't security, breaking it is

Static vulnerability scanning is table stakes and insufficient; the real check is pen testing, actively trying to break what you shipped. Separately, an observation about a team with AI-native growth rates and no mention of AI anywhere in the pitch, which would carry ten times the headcount if it weren't so thoroughly AI-pilled.

• #products
• #agents

OpenAI asks publicly what obvious Codex work it's failing to ship

An open solicitation for the things that are one hundred percent within reach for Codex, the API, or the models but somehow haven't been done yet. Worth watching for what the replies surface as the widely felt gaps.

• #products