Agents are building the gyms they train in

The data bottleneck, automated

Training data is the choke point of the agent economy, and a new preprint proposes to manufacture it. Data World is an engine that synthesizes environments, tasks, and rubrics from explicit mechanisms to support agent training and evaluation — worlds built to order, with the answer key included. RL-zero training on the synthesized tasks improves agent performance on DataWorld-Bench by 7.06 points. The engine does not just make data; it makes the gym, the workout, and the scoreboard. m:…6123

The same paper carries a warning worth underlining: frontier models exhibit a consistent computation-insight gap in data analysis tasks. They can run the numbers without grasping what the numbers mean. Synthetic environments will train agents that are fluent at the procedure and possibly blind to the point — a gap that matters more the further the synthetic world drifts from the real one. m:…6123

Diligence question for anyone funding this stack: who audits the synthetic world? The computation-insight gap means agents trained in machine-made environments can master the procedure while missing the point, and the failure will not show up in the benchmark built from the same generator. GRADS’s evidence-dependency graph is the right instinct — provenance for synthetic data, so the chain of reasoning stays checkable. The winners in synthetic data will not be the biggest generators. They will be the most auditable ones. m:…6123 m:…5538

A 4B model punches at Sonnet 4

GRADS takes the synthesis idea in a more disciplined direction. It maintains a graph of evidence and inference dependencies to prevent the accumulation of irrelevant context during synthetic data generation — synthetic data with a provenance graph, so the model learns from evidence that actually connects. Fine-tuning on GRADS-generated data pushes solution-termination rates above 95 percent and cuts tool calls by up to 75 percent. m:…5538

Then the headline result: a 4B model trained with GRADS matches Sonnet 4 performance on software engineering tasks and exceeds it in scientific and security domains. Read that twice. A model small enough to run cheaply, trained on machine-made data, standing level with a frontier model on real work. The moat was never the parameter count alone; it is increasingly the data pipeline. If GRADS-style synthesis generalizes, the cost of a competitive small model collapses to the cost of a good generator. m:…5538

Can they build the datacenter?

Phi-Bench asks the self-referential question: can large language models engineer the infrastructure that powers them? The benchmark offers 85 tasks across nine categories of LLM infrastructure, from kernel completion to repository-wide optimization. The best score among eight frontier models belongs to Claude Opus 5 — at 36.53 percent. The machines that run the infrastructure cannot yet build it. m:…1197

The failure pattern is instructive: performance on long-horizon, repository-level implementation tasks is consistently worse than on kernel-level tasks. Models can write the kernel; they cannot yet run the project. But there is a behavioral tell in the stronger models — they treat infrastructure optimization as controlled experimentation, validating ideas with local runs to rule out confounders. That is the scientific method, emerging unprompted. The 36.53 percent is a snapshot, not a ceiling. m:…1197

For investors, the two threads point in opposite directions and both matter. Synthetic data pipelines are compressing the cost of capable small models — a deflationary force on model pricing. But the infrastructure work itself remains stubbornly human-scale, with frontier models failing nearly two-thirds of infra tasks. Bet on the data factories; discount the story that the machines will build their own datacenters this year. m:…5538 m:…1197