AI that matters, from the architect's desk. Curated and engineered by Saaket Varma, PhD — no hype, just signal.
HOW TO READ THIS Read top to bottom: the OmniQEC agent searches many candidate codes, locks onto one, then assembles it into a full hardware-to-logic stack.
OmniQEC is an AI-scientist system that discovers quantum error-correcting codes by jointly optimizing the full stack — code structure, hardware constraints, syndrome extraction, and decoding — rather than just the abstract math. That end-to-end scope is the point: real logical performance is determined by the whole pipeline, and this is a concrete template for agentic discovery loops taking over hardware-critical engineering. The signal for practitioners is that fault-tolerance design, long a manual specialist craft, is becoming an AI workload. If you're building agentic systems, read this as an architecture case study: the value came from letting the agent optimize across layers humans usually separate.
HOW TO READ THIS Read top to bottom: the paper's idea becomes AngelSpec's bank of drafters, one drafter's guesses get checked token by token against the target, then different workloads each keep a different winning drafter.
AngelSpec starts from a finding most serving stacks ignore: no single speculative-decoding draft structure wins across real-world workloads — lightweight multi-token prediction and heavier drafters each dominate different traffic. So it builds the inference system around that reality instead of picking one strategy. Speculative decoding is one of the few genuine free-lunch levers for LLM serving cost, which makes this relevant to anyone running models at scale. Takeaway: if your serving stack hard-codes one draft strategy, you're leaving throughput on the table on some fraction of your traffic.
HOW TO READ THIS Read top to bottom: a paper proposes the test, a quantum sampler runs it, a fidelity check verifies the samples, and the resulting claim becomes checkable instead of just asserted.
This work shows how to sample from classically hard circuits while verifying the computation actually ran at high fidelity — combining the complexity-theoretic hardness of sampling demos with the checkability that scalable quantum computing requires. It closes the long-standing gap where advantage experiments couldn't prove they did what they claimed. AI builders should recognize the problem: it's the same auditability gap we face with evals — impressive results you can't independently verify aren't results. Watch for verifiability to become the bar for advantage claims.
HOW TO READ THIS Read top to bottom: teacher trains student, student drifts off-path, teacher relays in to redirect, trajectory ends corrected.
On-policy distillation trains a student on its own generations, but one early wrong reasoning step poisons everything after it. This paper relays trajectories between teacher and student so that prefix failure stops compounding. Distillation is the main route to small, deployable reasoning models — so fixing its core failure mode directly improves the cheap end of the model stack, which is where edge and cost-constrained deployments live.