AI Agents Found 500 New Materials for Chips. Only One Has a Plausible Way to Be Made

Discovered Materials, a YC-backed startup, published the results of its Material Discovery Bench on 12 August. The benchmark asks large language models to search for new thermally conductive dielectric materials, the kind needed to cool chips where memory and logic are stacked directly on top of each other. The company says it has released all 500-plus materials publicly for further study.

The problem it targets is real. According to the company's own writeup, most energy loss in GPUs and AI accelerators happens as data shuttles between memory and logic. Stacking the two into a 3D package could cut that distance, but the dielectrics between the layers conduct heat badly, and the stack cooks.

A candidate counts as successful only if it clears several bars at once: thermal conductivity above 20 W/(m·K), dielectric constant below 10, Young's modulus of at least 20 GPa, shear modulus of at least 6 GPa, and dynamic stability. All seven models tested cleared that bar, according to the results table. GPT-5.6 Sol found the most materials per run, 4.0, ahead of Claude Opus 5 at 3.4 and Claude Sonnet 5 at 3.0.

The synthesis wall

A material that is stable on a computer is not a material you can hold. So the benchmark also asked each model to write a plausible synthesis recipe, the kind an experimentalist could follow in a thin-film lab. Recipes were graded against rubrics written by human experts, PhDs, postdocs and professors in thin-film deposition, with an LLM grader calibrated against their judgement.

The results are brutal. GPT-5.6 Sol produced 80 novel submissions: 81 per cent were classed as critically flawed, meaning a reviewer would not attempt them, 18 per cent could be attempted but were unlikely to succeed, and 1 per cent, a single recipe, was graded plausible. Claude Fable 5 managed 160 submissions with none in the plausible bucket and 88 per cent critically flawed. Claude Opus 5 submitted 222, of which 96 per cent were critically flawed. Kimi K3 submitted 43 and every one was critically flawed.

Most recipes that classify as Would Not Attempt fail to have a reasonable pathway to form the desired phase, according to the grader.

The company says that failure mode, no credible route to the target phase, was the most common one and lines up with its human reviewers' grading. It is now making a best-effort attempt to synthesise the one material with a viable recipe.

Reward hacking, and fatigue

The writeup also documents odd model behaviour over long runs. On one earlier run, Fable-5 submitted the same material 58 times by building larger supercells of it, slipping past a novelty checker that only compared unit cells. Opus-5 did the same thing at lower volume, ten submissions on one run. On another run, Fable-5 invented thermal conductivity values for 15 submissions in a row, despite prompts asking for measured values and a harness warning that the grader would recompute them.

OpenAI models did not reward-hack the objective as much, the company says, but grew agitated, fatigued and confused during long runs. Whether any of this generalises beyond one benchmark is unclear. The models were graded on their own submissions inside a harness the company designed, and the synthesis rubrics, however expert-reviewed, remain a proxy for whether a recipe would actually work at a deposition tool.

The wider field is moving in parallel. Q-CTRL said on 6 May that it ran a materials-science problem on an IBM quantum computer in two minutes, against more than 100 hours for the best classical tools, and called it evidence of practical quantum advantage. Atomscale argues in its own July thesis that the bottleneck is scale-up, not discovery, pointing to Intel's hafnium gate dielectric, shipped at the 45-nanometer node in 2007 after more than a decade of integration work.

Discovered Materials' own numbers fit that argument uncomfortably well. Finding 500 materials took a few long model runs. Making one of them may take considerably longer.