Smart Brevity

Train the decision point: when to stop thinking, which harness to borrow, which model to call; LoCo: same-prefix preference pairs teach a reasoning model when to stop; Harness-S: sibling harnesses rescue failed RL groups, then the skill moves into the policy; Route Anything: oracle headroom says routing pays in cost before quality; System-1 gate: predict, gate, execute, update, with savings set by how rare useful actions are; Pandora: separate what a simulated user feels from how they say it; Kling: captions as a contract and a trained prompt rewrite layer; Judge check: compare AI and human rankings at the top...

A trusted hairdresser versus a film budget in climate communication, a hiring crisis that may be about remote work, the access fences labs build around risky models, the chip market splitting for agents, and what error bars hide. From the archive: twelve years of advice on whether to own one channel or two.

Score the net, not the gross: memory, training data, and benchmarks judged against the plain baseline; Memory tax: agent memory must beat naive RAG over raw history, and often does not.; Benchmarks as difficulty measures: measure mismatch and coverage gaps hide what a score really says.; Decision grounding: train only on actions the agent could justify from what it saw.; Turn-level credit assignment: agents decide in turns, so RL should assign credit per turn.; Interface lock-in: fine-tuning one answer format improves only that readout.; Decision models as routing economics: send yes/no agent steps to cheap typed-answer models.; In-situ interventions...

Week 2 of AI Systems: Models, Harnesses and Agents. Why two tools that can make the same edit can set a model very different problems: tool contracts, trajectories, the failure points in tool use, capability versus elicitation, and why the error a tool returns is half the interface.

Why coding agents ace the benchmark and flunk the job, what happens to content when the reader is an agent, three ways to handle a tax-loss problem, dosing versus rewriting aging, how early hires get picked, and an archive thread on engaging versus excluding in conservation.

Guarantees come from structure, not from asking nicely; Waypoint-gated distillation: transfer only the RL shifts that early and late training agree on; Decision-native models: null-aware, order-proof, scalable, and cacheable by construction; Thinking inertia: no-think switches leak reasoning, so measure answer-only compliance; Demo versus delivery evals: held-out acceptance, business-uplift scoring, priced clarifying questions; User-modifiable harnesses: users edit the agent loop, subagents, and routing; Cache-aware economics: stable prefixes first, judge models by cost per completed task

Week 1 of AI Systems: Models, Harnesses and Agents. Where an agent's behaviour comes from: weights, context, inference compute, and the runtime that executes its actions; how to attribute a behaviour with controlled interventions instead of guessing.

A chip maker pays eight billion dollars for a model that invents camera angles, two AI labs disagree about who holds the agent's leash, colleges face arithmetic that has nothing to do with AI, and an archive thread on thinking with the body.

The signal is narrower than it looks: recency posing as memory, failures that teach, less supervision that wins; Memory-need stratification: pooled memory scores mostly measure recency; Memory framing: calling context "memories" changes how often it gets used; Hindsight-to-foresight distillation: failed runs teach when group rewards are flat; Reliability over coverage: distill only where tokenizers align cleanly; Experience retention: raw logs versus summarized playbooks for learning agents; Cheap discriminative judges for closed-answer evals at volume; Brain and hands split: editable harnesses, cloud reasoning, device actions; Action gating: approval policy plus readable

American colleges are closing on arithmetic, not AI. Solar booms in the reddest states for business reasons. Private ownership explains more of the top one percent than its reputation suggests. Plus the offices and software seats that headcount no longer fills, cheap thinking versus costly experiments, and who checks AI's work. From the archive: speed in founder writing.

Your training signal has a shape, and it can lie: who teaches, how loudly, and who checks; On-policy distillation scaling: small RL teachers can lift bigger students past themselves; Multi-teacher distillation: normalize each teacher's feedback spread or the loudest domain wins; RIDE: extrapolate the RL direction in hidden states, not in word probabilities; Co-cheating: self-play loops reward shared errors; cross-fitted graders break the echo; Decision models: calibrated classifiers as cheap routers and eval judges; UCB tree search over experiment notebooks, and the reward hacking it invites; Environment interaction shape decides what skills transfer; Cost per completed t

This episode traces a shared instinct across a scattered inbox: turning soft judgment into hard rule. Tech analysts read Apple, Fox, and Meta's AI era deals as bets on distribution and data over any single model. Climate lawyers watch attribution science become courtroom evidence, and watch opponents respond by trying to close the courtroom instead of arguing the facts. Business, money, and product writers each convert a fuzzy call into an explicit rule. The archive segment compares an automated AI research pipeline against the argument that a human should still run the outer loop, and finds the debate already settled...

Measure the decision, not just the outcome; Regularized harness self-improvement: auto-tuned agent harnesses overfit unless edit size, specificity and bloat are constrained.; Decision-fork evals: grade an agent's choices at branch points; longer thinking does not improve taste, distillation does.; Instance-aware rubric rewards: a per-prompt checklist scored item by item works as eval and RL reward.; Full-duplex voice with async delegation: the frontend keeps talking while a background harness runs tools.; KV cache from zero, plus shrinking it by architecture and splitting prefill from decode.; Parallel test-time compute: many agents plus a verifier as a separate scaling a

This week's inbox set OpenAI's ten million dollar plus math proof against DeepSeek's cheaper, context efficient model, then found a Palantir engineer's costly early failure and a SaaS operator's thirty minute shadowing method arriving at the same lesson about building without direct customer contact. Scott Galloway's Oura IPO case and his piece on the Democratic Socialists of America both get read against a wider pattern in what makes a data business durable and why political reaction lags economic change. The archive segment traces thirteen years of design and marketing writing to show that craft judgment did not disappear when AI...

The mail slowed down, so this leans on what did land plus deeper currents underneath it. Three AI newsletters, same day, argue the model itself is becoming the least interesting part of the story. Four product writers converge on taste as the scarce input once AI removes coordination work. A UK energy campaign and a US grid newsletter reach the same diagnosis from opposite sides of the ocean. A finance writer flags an old retirement account that can wreck a backdoor Roth. Four performance guests show identity predicts comeback better than skill. The archive segment covers five years of growth...

The first Smart Brevity seminar: five research threads and the week's practitioner material on evals and AI product development, taught for understanding rather than headlines.

The sweeps stalled for a stretch, so this is a catch-up edition covering what actually landed in the last couple of weeks rather than a strict single week. Threads: Jerry Murdoch's debt-bubble warning against Anthropic's mechanical first profit and Poolside's compute-starved collapse into Nvidia; a SaaS Academy five-step discipline for capping AI spending; Climatebase on COP17 pledges reshuffling climate finance toward land restoration; Dror Poleg on AI's labor effect landing on data-center jobs and public opinion rather than paychecks; and an archive segment on three Finding Mastery guests who all learned to decide before the pressure arrives, not during it.

The podcast roundup for the tenth to the twenty-second of August: 20VC and More or Less, full coverage. The credit-market warning, the Ramp reality check on inference demand, the Stripe-OpenRouter deal from both tables, and every debate, quoted verbatim.

A third catch-up edition drawn mostly from the deep catalog, since only one new source entry landed in the week since the last dump. Threads compare Opus 5 and Kimi K3's frontier price cuts against a practitioner routing real agent work to DeepSeek at a fraction of the cost, weigh Scott Galloway's single-voice case that offline experience is repricing as digital content goes to zero cost, and trace carbon removal's demand shortfall through Air Miners and Carbon Copy's opposite fixes. The archive segment spans a 2014 Stanford lecture to a 2026 SaaS Academy guest post on how founders actually find real problems...

With this week's newsletter sweeps still dry, MailDump goes back through the catalog for threads that didn't make the debut episode: why three very different growth operators all say to work the customers you already have before buying new ones, why climate hiring, home electrification, and carbon policy keep hitting the same people-shortage wall, and why a personal-finance writer and two health researchers each turn anxiety into arithmetic. It closes with the show's first From the Archive segment, a six-year product-management convergence on outcomes over output.

The first MailDump. A cross-section of what has been landing in your inbox lately: why your newsletters agree the electrical grid, not chips, is throttling the AI build-out; the open fight over whether this is 1999 all over again; and the twelve-years-apart consensus on retention and founder-optional businesses. A debut catch-up edition, rendered in a synthesized voice.