AI Builders Digest

Chai Discovery took antibody design from a 0.1% hit rate to about 15%

When the company started, state of the art antibody design bound about one molecule in a thousand, weak enough that most targets returned no hits at all and you could never see drug-like properties in the results. Chai-2 gets roughly 15%, so screening a thousand candidates returns about 150 and you can finally read real statistics off the output. The approach is imported straight from LLM work: refuse to add the twenty-fourth submodule, keep the architecture simple enough that scaling laws are findable, and treat antibodies, mini proteins, and everything else as different prompts to the same model. Their contrarian claim is that biology is more verifiable than code, since binding and manufacturability are objective readouts while code taste is not, and the discipline that demands is blunt: "You can fool yourself so easily in biology."

• #research
• #evals
• #startups

Prototype on the most expensive model, then move to open weights in 6 to 8 weeks

The playbook he sees emerging from founders is to build the first version on the best frontier model with cost and latency ignored entirely, then learn what users actually want before optimizing anything. Only after the workflow and UX are validated do you go hard at prompt engineering, model routing, harnesses, smaller models, and fine tuning. His timing claim is specific: open-weight models catch up roughly six to eight weeks later, and that is the moment to move production workloads. Starting with the cheapest model inverts the order, and most teams he talks to are stuck in step one.

• #open-source
• #products
• #pricing

Ask 10 IT leaders about their coding agent strategy, get 5 different answers

Enterprise AI has nothing like the two or three deployment patterns that defined early cloud, when a handful of infrastructure vendors constrained the options. Some companies standardize on ChatGPT or Claude, others offer a menu, and many have built their own orchestration layer so employees can reach any model. The divergence runs deeper on data access: some enterprises have agents assume the role of the user, others issue agents their own identities, and guardrails range from heavy to nonexistent. His conclusion is that anyone predicting ultimate market winners right now is probably wrong, because nothing has settled yet.

• #enterprise
• #agents
• #open-source

Google's Notebook drops the mode toggles for a single prompt bar

Notebook is now built around one unified prompt bar rather than a set of modes to switch between, positioned explicitly against products that keep adding more modes. The framing is that it should just do the thing you want without making you pick how. Ultra and Pro subscribers have it now, with a broader rollout next.

• #products

Technology adoption is emotional, so stop selling efficiency

People do not adopt a new technology because it makes them more efficient. They adopt it because someone similar to them adopted it and visibly got better off, or because they feel everyone around them already has and they will be left behind. That makes the right pitch social proof, not a 10x productivity claim, and it makes diffusion a social process rather than a rational one. Her practical version: the best AI training is not a course, it is pulling an agent into your team's group chat and letting people watch it work.

• #agents
• #enterprise

Knowledge graphs are back because intelligence got too cheap to meter

Ontologies and graph knowledge are trending now rather than three years ago for one reason: the hardest part of building a knowledge graph was always the intelligence, and good enough intelligence is finally too cheap to meter. As intelligence commoditizes, its complements rise in value, which is why knowledge graph talks are suddenly landing. He picked this up at a Midjourney meetup, which is a decent signal of where the practitioner conversation has moved.

• #research
• #agents

"Vercel is the Vercel for backends," with Factory AI at billions of requests

The positioning push is that Fluid compute makes Vercel a backend platform, not just a frontend one, with Factory AI running its API services on it at billions of requests per month. Separately he claims one line of code in the AI SDK cuts DeepSeek v4 Flash token spend through the AI Gateway by 90% or more.

• #infrastructure
• #pricing

SPC backs Rivo, agents that sweep your checking account into Treasury yield

Rivo connects agents to the checking account you already have, learns your cash flow, moves idle dollars into Treasury-backed yield, and brings them back before bills hit, so every dollar earns until the moment you spend it. The hard part is prediction under asymmetric cost: money returned a day early costs a little yield, money returned a day late bounces a bill and destroys trust. The bet is on the founder, Ambrish, who led the L4 autonomy work at Cruise through robotaxi launch, on the theory that self-driving money is the same problem with a friendlier failure mode. Agarwal also argued this week that AI deployed at scale does not need to be a black box, and that understanding how these models work is a prerequisite, not a nice-to-have.

• #funding
• #agents
• #research

The vibe-coded SaaS is really a funnel into a services business

His read on micro SaaS economics is that the product itself is probably not where the money is anymore. It works better as a self-serve funnel into a more expensive services offering, though he names the catch immediately: services drags you back to selling time for money. He also notes it is now easier to earn from X payouts than from a micro SaaS, and says he is hearing GPT 5.6 Luna High is much cheaper with basically unlimited usage, with his open question being whether it handles complex browser automation, roughly half of what he uses Codex for.

• #startups
• #products

AI use will become assumed and unimportant, and the humans stay heroes

His prediction is that the current strangeness about AI doing the work is temporary. Once what he calls the agency rupture heals and AI goes back to being invisible, we will think only about the humans and what they have done. Using AI becomes assumed and unremarkable rather than something worth crediting or discounting.

• #products

Altman: no essay about why it will not work has ever moved anything forward

He would rather be an optimist who works hard than a pessimist posting about why things will not work. He concedes the optimist path is much harder and that failure is the most likely outcome, but argues society fails if people do not try.

• #startups

Everyone says Airtable sold low, founders would take that deal tomorrow

The public reaction to Airtable's exit price was that it went far too cheap. His counterpoint is what SaaS founders say privately, which is that they would take it any day, because at least it was an exit.

• #funding
• #startups

Tan calls Prop 40 an asset seizure tax that will destroy California's tax base

He calls the California Democratic Party's endorsement of yes on Prop 40 insane, framing the measure as an asset seizure tax that wrecks the state's tax base. His housing argument runs alongside it: if you want prices to fall you build more housing and let markets work, while abolishing rent and pushing mass expropriation mainly serves whoever is fomenting it.

• #policy

Halvar Flake is joining OpenAI's Codex and ChatGPT team

Filed under "Better Cyber," Sottiaux says Halvar Flake will be joining soon. A security hire on the team behind OpenAI's coding agent is worth noting given how much code these systems now write and run.

• #security