A recurrent-depth language model ships six ways to stop early. Measuring one of them at its own default: it selects better stopping points than chance, and on one task it still loses accuracy on the answers. Both are true, and separating them needed two different experiments.

A regime where the optimizer is not the bottleneck and cannot become one — how to recognise it, why every optimizer-side fix I tried did nothing, and why the one data-side intervention that appeared to work had to be withdrawn.

A three-act investigation into Safe Superintelligence Inc. that kept overturning itself: the ‘straight shot’ thesis is elegant but rests on an unverified premise; scaling didn’t die, it fragmented into four axes SSI can’t access; and the data flywheel’s reward signal is measurably broken — which turns SSI’s ‘no product’ from fatal weakness into audit qualification. Each act’s evidence almost closed the case; each layer of depth reopened it.

We put interpretable, locally-computable factors from four mechanistically-different domains — software-defect prediction, paper acceptance, AI-text detection, and a world-model physics proxy — on one shared per-factor scorecard. Predictive power has a domain-dependent ceiling, the factors don’t transfer across domains, and a single-factor reading nearly had us call an industrial benchmark a ‘collapse’ when the signal was just multivariate.

We replaced the single-LLM mutation step of an AlphaEvolve-style evolve loop (OpenEvolve) with a proposer→critic→aggregator panel and tested it rigorously across three problems. It never won — equal on easy optimization, significantly worse on hard optimization (p=0.01), and on correctness-gated ARC it cracked nothing a single call couldn’t. A clean negative result, with the baseline/ablation harness that makes it trustworthy.

A field note on 2025–2026 agentic RTL generation: every result that moved the needle came from the verification loop wrapped around the LLM, not the LLM itself. For someone who builds compilers and ML infra, that loop is the product.

As RL-with-verifiers scales, the Lean verifier is both the reward source and the throughput bottleneck — a CPU/scheduling problem almost nobody benchmarks on commodity, on-prem hardware. We profiled it: on an 8-core box, ~4.5% of proofs run away and eat most of the single-thread wall, and a tuned per-proof deadline beats adding cores. The honest-measurer story (including the claim we had to retract) is the point.

A tiny adversarial benchmark shows small local LLMs confuse token-distance for elapsed wall-clock time. A zero-dependency continuous-time memory fixes it for sub-millisecond cost — but whether a model will even accept that fix turns out to be a model-dependent deployability property that capability leaderboards don’t measure.

A research note on the GPU scale-up interconnect battle — why NVSwitch’s non-blocking all-to-all is NVIDIA’s real moat, where the open UALink camp and AMD’s Infinity Fabric actually stand, and when the fabric becomes the binding constraint on LLM training.

On June 12, 2026, a US government directive forced Anthropic to abruptly disable Claude Fable 5 and Mythos 5 for all customers. A careful, source-verified account of what actually happened — separating the confirmed facts from Anthropic’s characterization, and from the jailbreak-research claims that are not yet independently established.