Executive Summary
Computational biology and software engineering reported the same shift today. Once generating output gets cheap, the constraint moves to the loop that validates it and acts on it. Fabian Theis argues that the value in virtual-cell work is moving from the model to closed-loop experiment design, and that wet-lab cycle time limits it, not compute. Vijay Raji describes OpenAI's coding agents overwhelming Git first, then review and CI. In both accounts the model is the least scarce part. A third, thin signal (Berman) suggests that model weights reveal what they were trained to do, which weakens the idea that weights are the asset worth protecting. For enterprise buyers the conclusion is to fund the verification pipeline and the proprietary data, not more generation capacity.
What Changed
- OpenAI replaced Git-based source control for its monorepo. Agent traffic went more than 10x past what the system was designed for. The replacement assumes agents that are always online, and OpenAI claims about 100x the throughput. The hard problems were checkout throughput, sparse checkout and caching, not merge collisions on main.
- Perturbation data reached the scale world models need. Perturb-seq now runs across 100+ cell lines, and single experiments produce 10M transcriptomes. Lab Compass, an inverse model that proposes the experiment needed to reach a target cell state, has completed 3–4 real cycles of model-proposed experiments.
- Baselines exposed inflated claims. Linear and mean-prediction baselines matched foundation models on perturbation tasks until the Arc challenge cleanup fixed the metrics and holdouts. After that, nonlinear models won only on harder tasks.
- Early demo: weights leak task identity. A frontier model named the game Breakout from 17k raw weights alone. This is one demo on a tiny network.
Cross-Expert Synthesis
Theis and Raji agree on three things. First, the bottleneck sits downstream of generation: assay cycle time for Theis, review and CI/CD for Raji. Neither names compute. Second, defensibility comes from what model progress can't commoditize. Theis puts the moat in proprietary perturbation data plus tight experimental integration. Raji's test is that a model release should improve your product for free, not threaten it. Both are saying the same thing: own the data and the loop, rent the model. Third, the hardest problem is organizational. Experimentalists resist spending weeks on experiments a model suggested. Typical software shops grew token usage about 2x in a year, against 17x at frontier labs.
They diverge on evidence. Theis's central lesson is that impressive claims collapsed against simple baselines once holdouts were audited. Raji's numbers (100x repo throughput, a $1B ads run rate in 7–8 months with about 50 people, 17x token growth) are self-reported and unaudited, and he concedes tokens are a crude proxy. Apply Theis's discipline to Raji's numbers: the direction is plausible, the magnitudes are not established.
Berman's result connects to Theis's moat argument. If weights encode what a model learned, a fine-tuned model is a lossy copy of the proprietary data that is supposed to be the moat.
Enterprise Implications
Coding-agent rollouts sized by seat count will hit a wall that has nothing to do with licensing. Enterprises running parallel agents against a large Git-hosted monorepo should expect checkout and CI saturation before they see quality problems. Almost none can build a replacement version-control system as OpenAI did. They will need caching layers, sparse checkout, ephemeral build environments and elastic CI capacity. That is infrastructure work, squarely in AI-infrastructure advisory territory, and most agent vendors don't scope it.
For life sciences, and by extension other R&D-heavy sectors, the model is not the purchase. The purchase is the data-generation program and the integration into experiment scheduling. The moat Theis describes is a pharma consortium buying early access to an organoid perturbation atlas, not a licence for model weights.
For private AI deployments, weights shipped to partners, customers or edge devices should be treated as derived data, not as inert code.
What To Do About It
- Model the load before scaling agents. Multiply PRs per day by CI minutes and checkout size at 10x current volume. Put review automation and CI capacity in the same budget line as agent seats.
- Measure at merge-to-production. Count what reaches production, not tokens or PRs generated. Generated volume rewards the wrong behavior once prototypes are cheap and mostly discarded.
- Make baseline comparisons a procurement requirement. For any "foundation model for X" pitch, require results against linear and mean baselines on audited holdouts. A vendor who hasn't run them is a red flag.
- Classify weights with their training data. Put fine-tuned weights under the same handling policy as the data they were trained on, until weight-leakage research shows what can and can't be recovered at scale.
- Customer question: "If your agents produced 10x the PRs tomorrow, which fails first — repo, CI, or human review — and who owns that system today?"
Risks and Blind Spots
OpenAI's experience may not transfer. Its monorepo, agent density and engineering talent are outliers. Most enterprises will hit review-quality failures in security and correctness before they hit repo throughput limits, and today's sources say nothing about the defect rates of agent-written code or the real cost of discarded prototypes. Theis's lab-in-the-loop evidence comes from a handful of groups and one validated system, in hematopoiesis, so "the loop is the moat" is a strong hypothesis rather than a proven pattern. The weight-leakage result comes from a fragment covering one demo on a 17k-parameter network; keep it on the watch list, not the roadmap. Finally, every source today speaks from the vendor or research side. None gives an enterprise buyer's account of agents or virtual-cell models in production, and that is the gap that matters most for advising clients.