Hi, I'm Sean Moran.

A weekend spent on two open problems with parallel agents, exact-arithmetic verifiers and a proof assistant. Hadamard 668 resisted every route and stayed open. Maxwell's equilibrium problem yielded a proof candidate: Lean checks its algebraic core, no specialist has read it, and its novelty is unestablished. The experiments were cheap. The knowledge was not.

Moonshot's Kimi K3 is a 2.8-trillion-parameter open-weight model that ranks near the top of third-party leaderboards — yet the report shows most of the work isn't the architecture. A read of what actually separates a frontier model: linear attention and sparse mixture-of-experts routing, the environments that grade tasks, reward models built to resist gaming, multi-teacher distillation, and a serving stack that stays cheap at scale.

A 28M-parameter transformer you never train beats a tuned XGBoost on tabular data. An introduction to tabular foundation models — how in-context learning on rows and columns works, and an independent reproduction of TabICLv2 on TabArena.

How time-series foundation models forecast, using patches and quantiles. Where t0-alpha lands on GIFT-Eval, and what its reproducible result says about the field.

Fit a baseline, then model only the part it leaves unexplained — don't tell me what the baseline already knows. Residualisation isolates a feature's unique contribution, giving clean attribution when effects are confounded and modular designs where each model adds genuinely new signal.

Token consumption is an input, not an outcome — like CPU hours or a cloud bill. Borrowing the FinOps playbook, TokenOps connects AI spend to delivery, so the question shifts from how many tokens you burned to what those tokens produced.

Deep Deterministic Policy Gradient in full: the maths, a code sketch, the implementation details that decide whether it trains at all, and the off-policy actor-critic algorithms it grew into.

AI has collapsed the cost of research execution — literature review, coding, analysis — but not the cost of scientific judgement. As execution commoditises, the scarce skill becomes choosing the right questions, recognising flawed results, and verifying claims.

Representing uncertainty in structured systems, through directed Bayesian networks and undirected Markov networks. Runs from foundational concepts through practical inference to the weighted logical rules of Markov logic.

Not all agents are jobs with a clear endpoint. Durable workflow agents carry long-lived processes through time, handling waits, retries, failures and real-world side effects with persistent state and event-driven resumption.

Stainless built an SDK generator — something any AI can replicate. What it couldn't replicate was the dependency graph, the production edge cases, and the canonical position. When the moat isn't the app, it's who relies on you.

Many enterprises treat AI as a procurement problem — why the real challenge is building internal capability to evaluate and adapt AI systems, not just buying them.

Speeding up individual tasks rarely makes an end-to-end process faster — it shifts the constraint elsewhere. A reading of JPMorgan's AI-transformation experience through the lens of the Toyota Production System, and what it implies for workflow design in knowledge work.

When software becomes cheap to create, it no longer has to be generic. What changes once applications can be shaped around an individual instead of an average user.

Walkthrough of a decentralised dead-man's-switch built on a modern smart-contract stack: the components involved, the trust assumptions, and the design choices that hold it together end to end.

What attention maps, residual streams, and hidden states can tell you about a transformer's internal computation, and the limits of current mechanistic-interpretability techniques.

Practical lessons from time-series forecasting: split design, leakage, calibrated baselines, drift detection, and the recurring failure modes of naive models.

How to get the most out of Claude Code today, why the harness around the model matters more than the model alone, and what to drop from the 2025 playbook.

The released OpenAI emails, read as a record of how AGI strategy looked in the year before the transformer architecture reshaped the field.

The 1952 statistical model behind chess ratings, A/B test ranking and the reward model in RLHF, derived from first principles with a working implementation.

Two years of Y Combinator batches read as a signal about the AI market: which categories are growing, which are commoditising, and where founders are positioned against consensus.

Current LLM-agent frameworks sit at an early, exploratory stage analogous to phone phreaking. A sketch of the more durable engineering patterns underneath them.

An agent shipped a prototype end-to-end while I was at the coffee shop. What the IDE becomes when the inner loop runs without you, and what the developer's job becomes instead.

The computational "circuits" that emerge inside transformer attention layers, with a hand-traced example of a model reasoning across several steps.

Accuracy and recall rank RAG and agent systems poorly. Bits-over-random compares them better, which changes how you choose between embedders.

Some production RAG queries have answers that dense similarity cannot reach at all. Why that happens, and the hybrid retrieval patterns that address it.

Self-evaluation, uncertainty calibration, and abstention as first-class design choices in multi-agent systems, and how each one contributes to operational reliability.

Errors compound 17× in "bag-of-agents" architectures. The structural choices that limit that compounding.

Close reading of the NeurIPS 2025 Best Paper on attention gating: what the ablations show, how attention sinks behave, and what it means for inference cost.

Nonparametric clustering of deep embeddings with the Distance-Dependent Chinese Restaurant Process, for settings where you do not know the number of clusters in advance.

How to keep your engineering judgement intact when an AI assistant is doing more of the writing — a personal "internal KL divergence" check for when you're drifting from your own thinking.

Victor Lavrenko's relevance models, traced through cross-encoder reranking to modern attention, drawing out the continuities between classical IR and the transformer.

Kahneman's "Noise" applied to AI-assisted decision making: where calibrated models reduce variance, where they amplify it, and how to tell which is happening to you.

Pairing an LLM with a symbolic proof assistant to do mathematics neither side manages alone. What the architecture looks like, and where it still fails.

The JPMorgan engineering blog on Senatus, an internal code-to-code recommendation engine built on De-Skew LSH that scales sub-linearly across very large codebases.