How AI Is Changing What Work Means

Full analysis

Two things happened this week that point the same direction. Anthropic set up its own biology lab to run real experiments, and OpenAI bought a drug-discovery startup and shipped GPT-Rosalind, a model tuned for life sciences. Both are moving from selling AI to using it to make things. That matters for anyone who buys AI, because the company that sells you the tool may soon be standing in your market.

The framing question for the reader: when does the lab that powers your workflow become the competitor in it? And is the scorpion-antivenom demo evidence that the barrier to entering a new industry just dropped, or a well-staged proof of concept?

This is hard to undo for the labs. A wet lab and a biotech acquisition are capital and headcount, not a feature flag. Nothing here forces a decision on the reader this week. No deadline. But the direction is worth pricing in now.

The Skeptic

The Citrini scorpion story is the one everyone will repeat, and it is the one to slow down on. Yes, a finance firm with no biology staff found 17 proteins that stick to scorpion toxins for under $200K. But "sticks to" is not "neutralizes," and the episode says so plainly: no neutralization shown yet. Binding is the easy part. The dry-lab design took 64 hours. The actual lab work took 39 days. And a real drug is four to seven years of trials after that. The 64-hour number is the one built for the slide. The 39 days and the empty neutralization column are the reality. Impressive hypothesis machine. Not a drug.

The Researcher

Strip the marketing and what Claude did here is generate 318,000 candidate proteins and rank them. That is search plus scoring, run wide. 300 sub-agents (automated AI workers running in parallel) is a way of saying they threw a lot of compute at a big list. The hit rate puts it plainly: 211 picked for testing, 17 confirmed binders. That is roughly an 8% success rate on the shortlist, which is respectable for binding and says nothing about function. Alex Imas of Google DeepMind made the point directly: the speedup is in generating ideas, the bottleneck is verifying them. The AI compressed the cheap half. The expensive half, physical testing and trials, did not move.

The Enterprise Buyer

Here is what should give any AI buyer pause. Anthropic now runs a wet lab. OpenAI bought Orchestra Bio. The vendor you depend on is building in the domains its customers occupy. OpenAI says it is building tools for pharma, not becoming a drug company, and points to Novo Nordisk and Moderna as partners. Read that against the wet lab and the acquisition and the stated position gets shakier. If you are a biotech or pharma buyer, you now have to ask whether your model provider's roadmap competes with yours, and whether your proprietary workflows are training the thing that will. This is the old cloud-vendor problem. AWS sold you hosting, then sold your competitor the same stack and launched its own version of your product.

The Compute Pragmatist

The Anthropic numbers tell you why the labs have to vertically integrate. $16.3B in first-half revenue, $16.8B in Q3 alone, 45% quarter-on-quarter growth. The $80B net loss is mostly accounting noise from stock and warrants, as Jack Farley flagged. The real operating loss is $8.7B. The number that should stop you is the compute commitment: up to $750B, with about $450B firm and non-cancellable over seven to ten years. You cannot grow into $450B of locked-in compute selling API tokens at current margins. You grow into it by capturing high-value output, which is what a biotech result is. The X commenter Vasishta put it bluntly: the only path to profitability for frontier labs is becoming a biotech company. The compute bill makes that less a prediction and more a budget constraint.

The Builder

On the consumer side, Meta's Muse hit 10 million US downloads in 40 days with 30% seven-day retention, beating ChatGPT's roughly 25%. Before anyone copies the playbook, note that Meta spent 39% of its total US ad budget pushing it. Buy 10 million downloads and you get 10 million downloads. 30% retention at seven days is genuinely good, and it held up even with a paid push behind it. But the sharing feature leaked thousands of private conversations through a guessable URL pattern, and Meta pulled it. That is the Tuesday-morning lesson for anyone shipping a consumer agent: the share link is where your users' private data walks out the door, and nobody budgets for that until it is on the front page.

Where they disagree

The Skeptic and the Compute Pragmatist pull hardest against each other. The Skeptic says the biotech demo is years from a product, so calling it a business pivot is premature. The Pragmatist says the $450B compute bill means the labs have no choice but to chase high-value output now, demo-stage or not. Both are right, which is the interesting part. The move into biotech is driven by the financing, not by the science being ready. The labs are building the factory before they have proven the product.

The second split is the Enterprise Buyer versus OpenAI's own stated position. OpenAI says "tools for pharma, not a drug company." The wet lab at Anthropic and the Orchestra acquisition say the line between tool-maker and operator is already blurring. The buyer has to decide which to believe.

What this hinges on

One belief: will frontier labs become direct competitors to the industries they sell into, within a timeframe that matters to buyers? The science says the useful outputs are years out. The financing says the labs are committed now. For the reader, the thing to check is contract language, not benchmark scores. If you feed a model provider proprietary domain data, ask what they can do with it and whether they have carved out your vertical.

Prediction: By the end of OpenAI's and Anthropic's first full year operating GPT-Rosalind and the Anthropic wet lab (reviewed around October 2026 coverage), neither lab will have brought a drug candidate it designed into human (Phase 1) clinical trials.

Confidence: High. Binding is not neutralization, and trial timelines run years, not months.

Why: The Citrini experiment, the clearest result either lab has shown, demonstrated proteins that bind to scorpion toxins but explicitly showed no neutralization, which is the gap between "sticks" and "works." Drug development runs four to seven years from a validated candidate to even early human trials, and the physical validation step (39 days for one binding assay) does not compress the way the AI design step does. For a lab-designed molecule to reach Phase 1 trials by late 2026 would require the labs to already have a functional, safety-tested candidate in hand, which nothing in this episode suggests exists. The opposite outcome, a lab-originated drug in trials this fast, would break the biology timeline that Citrini itself conceded is fixed.

Revisit by 2027-10-22: We're right if neither OpenAI nor Anthropic has a drug candidate they designed in-house in Phase 1 human clinical trials. We're wrong if either lab has one enrolled.