Microsoft AutoTuring has GPT-6 Sol and other models design GPU accelerators
Microsoft Research's AutoTuring had Claude Opus 5.5, GPT-6 Sol/Luna and Grok 4.7 design FP16 GEMM kernels in a 15-dimensional GPU space, beating a modeled H200 by 5.4% while cutting simulator calls by 70%.
this is pure f*cking wonderful Claude Opus 5.5 and Grok 4.7 proved AI understands computer architecture Autonomous agents frequently design specialized chips, yet conventional benchmarks fail to prove whether models reason about physical machines or hill-climb over anonymous parameters. Stripping domain semantics out of the problem while holding the evaluation harness fixed reveals the truth Researchers from Microsoft Research introduced AutoTuring to evaluate frontier agents across a 15-dimensional GPU accelerator space. Orchestrating Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and Grok 4.7 across 9 FP16 GEMM kernels outperforms a modeled NVIDIA H200 by 5.4% while cutting simulator calls by 70.1% Workload basket → Causal parameter synthesis → Parallel kernel simulation → Falsification audit → Physical silicon tape-out Four operational principles separating physical deduction from stochastic coordinate search: 1. Semantic Grounding vs Coordinate Bookkeeping Exposing real architectural knobs allows Claude Opus 5.5 to partition kernels by bandwidth and compute bottlenecks. Handed the identical space as anonymous variables on [0,1], models degrade into coordinate bookkeeping and stagnate on noise plateaus 2. Causal Hypothesis Falsification Execution transcripts confirm physical deduction over trial and error. When an aggressive sweep gutted L1 and L2 cache allocations to buy extra compute cores, memory-bound kernels suffered a 39 µs latency regression. The agent immediately retracted the thesis on the next turn, classifying cache as load-bearing infrastructure 3. Source Code Introspection When paired with an external Critic recommending an increase in off-chip memory bandwidth, the architect inspected the Python evaluator codebase. Discovering that HBM bandwidth was hardcoded as a fixed constant and ignored overrides, the model dismissed the advice rather than corrupting the design 4. The Multi-Agent Critic Illusion Structured critique and domain grounding behave as substitutes rather than complements. While an adversarial Critic improved the blind optimizer by 4.7%, forcing the Critic onto a domain-grounded architect degraded mean-best latency by 3.4%. Physical knowledge renders multi-agent debate redundant Benchmarking end-to-end latency cannot determine whether an agent comprehends hardware physics. Autonomous chip design requires deterministic DAG execution tracing to isolate causal deduction from accidental coordinate convergence Read the complete breakdown in the article below ↓
This post quotes another post: x.com/marfinxx/status/2094016175617241109