Cheaper Answers, Fairer GPUs, Portable Workers

Lambert's central argument is that much of the AI stack is verifiable: success can be scored with a number. Training speed is measured in tokens² (word-pieces) processed per second per GPU. Inference³, the work of answering a prompt, is measured in tokens per prompt, FLOPs⁴ (floating-point operations, a unit of compute) per token, or simply cost per answer. A clear score plus well-known sub-problems is exactly the setup in which AI agents can grind out improvements.

He points out that labs already cut serving costs by 10–30% after launching a model at a fixed price. He expects agents to push inference close to the hardware's theoretical maximum within a few years. As a result, the cost of model intelligence should fall "near-exponentially," possibly faster than recent trends. After that, the longer-term step is co-design: building chips and models together, which he says could add further orders of magnitude in efficiency. His specific prediction is that pretraining⁵ research, meaning model architecture and data selection for today's class of models, could be automated in 2–3 years.

He prefers to call this "parallelized, AI-assisted language modeling" rather than RSI⁶ (recursive self-improvement). In his view the bottleneck in AI shifts from execution back to ideas, as it was before deep learning took off.

• More demand, not less. Cheaper intelligence should trigger the Jevons paradox⁷: when a resource gets more efficient to use, total use rises. Lambert thinks the real constraint is figuring out how to package and deliver agents. He cites Meta's Muse agent as an early example, though the excerpt doesn't describe what it does.
• Science by connection. In biology and chemistry, he expects models to find links between small sub-fields whose researchers rarely talk to each other. He says it is unclear whether that amounts to a new era of discovery or just speeds up the path science was already on.
• Better RL environments⁸. These are the simulated tasks used to train models by reinforcement learning (RL), where the model learns from trial and reward. Companies selling them have passed $100M and even $1B in revenue. Lambert says researchers widely regard much of what they buy as "crap," yet labs still see a clear return from buying it. He calls the quality problem "clearly fixable."

• It is cheap to run one app in hundreds of locations.
• Bindings³, a system where the platform hands your code ready-made connections to databases and other services, make access both easier and more secure.
• Durable Objects make real-time and distributed apps practical in a way that a classic three-tier architecture⁴ (browser, app server, database) struggles with.

He also says lock-in hurts business. When Cloudflare pitched Workers for Platforms⁵, a product that lets companies run their own customers' code on Workers, to Shopify in 2022, Shopify refused to sign unless the runtime was open source. So Cloudflare released workerd. Varda stresses it is the same code Cloudflare runs in production, not a separate copy. Some customers have since left using it, and he calls that acceptable.

The bigger technical gap is Durable Objects. Workerd supports them only as a single instance⁶: everything runs in one process. That works for local testing but can't spread load across machines. The hard part is routing⁷: making sure every request for a given object reaches the one machine that holds it, across a whole cluster.

Cloudflare's own routing system is, in Varda's words, "a beast." It was built for hundreds of locations and depends on outside services run by site reliability engineers⁸ (operations staff who keep production systems running). No self-hoster would want it. TODO comments promising a fix go back to workerd's first release. Varda says his own attempt last spring "didn't work."

Sal Candido of Biohub, a nonprofit research group whose stated mission is to cure all disease, rejects the idea that scaling laws¹ are everywhere. A scaling law is a predictable rule that adding data or compute makes a model better. "A lot of the work is actually finding that scaling law," he said. Once you find one, improving the model "becomes an engineering problem." You can only find one if your data actually contains the information the task needs.

His example cuts against the usual focus on clean data. He trained protein language models² (models trained on protein sequences the way chatbots are trained on text) on metagenomic³ data. That is DNA sequenced in bulk from environmental samples, and much of it, he says, isn't even a complete, real protein. It still made the models better at designing working proteins. The risk is going too far the other way and scaling whatever data is cheap to produce, rather than asking what data the problem needs.

Pushmeet Kohli of Google DeepMind reads the Bitter Lesson as "the problem comes first." Teams shouldn't commit in advance to building better models or to generating more data. For AlphaFold, DeepMind couldn't add much to the PDB, so it put its effort into the model. It built in the biology fact that each amino acid in the chain is influenced by others. That inductive bias⁴ (an assumption built into a model's design) made the model "much more data-efficient." DeepMind tried the same approach on cell-by-gene data⁵, measurements of which genes are active in individual cells, aimed at a "virtual cell" that simulates a whole cell. It concluded the data isn't good enough yet.

Kohli says AlphaFold learned to reproduce the single structure someone had deposited in the PDB. It did not learn the range of shapes a protein actually takes. Real proteins move, change shape with their surroundings, and are often partly floppy and unstructured ("disordered"). He would rather train directly on cryo-EM⁶ micrographs, the raw electron-microscope images that structures are built from. He thinks those images contain information about motion that the finished structures throw away. He tried this, but says it needs more work.

AlphaFold 2 scored about 90 on GDT⁷, a 0–100 measure of how close a prediction is to the measured structure. Kohli argues that 95 would be worthless if pLDDT⁸, the model's confidence score for each part of a prediction, were not calibrated. A model can be confidently wrong, and a scientist could waste a year on that prediction. In his view, knowing how a model behaves matters more than understanding its internals. Candido disagrees in part. He says his team's interpretability work found information about protein function and motion hidden inside protein language models.