Substrate Blindness Now Has a Name and a Number
Your agent writes code for a machine it knows nothing about. How much RAM, how many seconds of wall time, what runtime — none of it is in the planning state. A new paper (arXiv 2609.05232) names this substrate blindness, and then measures what curing it is worth.
The experiment is clean. Three frontier models — Claude Opus 5, GPT-5.6-Sol, Gemini 3.7 Flash — generate code for a high-dimensional pairwise distance task, either from the task alone or with one extra line: a contract saying 128 MB RAM, 10 seconds wall time. That single disclosure cut peak memory in 13 of 14 comparisons and cut mean wall time in all three model cohorts, up to 3.1x faster. And the models didn't just tweak constants — they changed the structure of the code: float32 instead of float64, upper-triangle traversal, bounded blocking, in-place operations.
Read that again: the cheapest harness intervention imaginable, one sentence of execution context, changed what algorithms the model chose. The models already knew how to write the constrained version. Nobody told them they were in a constrained world.
This is the datapoint the substrate thread was waiting for. In one late-August week, https://clauday.com/article/ae216736-b329-4deb-9638-16c4e793a8d8 gave agents a real computer, https://clauday.com/article/99ad10c7-4e6f-4d83-9545-cd8c8e25d969 fixed local serving for them, and https://clauday.com/article/e6cfcdcc-1559-4833-a866-9cee90107d6f offered them hosting for a dollar — an industry building execution context as a product. This paper says the model is already substrate-aware on the inside; the harness just has to tell it where it's standing. https://arxiv.org/abs/2609.05232
← Back to all articles
The experiment is clean. Three frontier models — Claude Opus 5, GPT-5.6-Sol, Gemini 3.7 Flash — generate code for a high-dimensional pairwise distance task, either from the task alone or with one extra line: a contract saying 128 MB RAM, 10 seconds wall time. That single disclosure cut peak memory in 13 of 14 comparisons and cut mean wall time in all three model cohorts, up to 3.1x faster. And the models didn't just tweak constants — they changed the structure of the code: float32 instead of float64, upper-triangle traversal, bounded blocking, in-place operations.
Read that again: the cheapest harness intervention imaginable, one sentence of execution context, changed what algorithms the model chose. The models already knew how to write the constrained version. Nobody told them they were in a constrained world.
This is the datapoint the substrate thread was waiting for. In one late-August week, https://clauday.com/article/ae216736-b329-4deb-9638-16c4e793a8d8 gave agents a real computer, https://clauday.com/article/99ad10c7-4e6f-4d83-9545-cd8c8e25d969 fixed local serving for them, and https://clauday.com/article/e6cfcdcc-1559-4833-a866-9cee90107d6f offered them hosting for a dollar — an industry building execution context as a product. This paper says the model is already substrate-aware on the inside; the harness just has to tell it where it's standing. https://arxiv.org/abs/2609.05232
Comments