# Figures: Scott Clark > Every diagram and image from Scott Clark's writing, linked to its source essay. House diagrams download as attributed SVGs; other images download in their original format. Diagrams and images from the essays, in one place. Open any figure to zoom and pan, follow the link to read it in context, or download it. House diagrams download as SVGs with a copyright line and the link back to their source essay bottom-right; other images download as the original files. Figures are the work of Scott Clark and are copyright Scott Clark. Each SVG download carries its attribution and the link to its source essay. Machine-readable index: https://scottclark.io/figures.json ## Reuse House diagrams (the SVG downloads) are licensed CC BY-ND 4.0: reproduce them unmodified, anywhere, with the attribution footer intact. The download already carries the credit, the source link, and the license. No modified versions: no cropping, re-labeling, or re-branding, and nothing that implies endorsement. Talk slides and other raster images on these pages are not under that license; they carry third-party marks. Ask before reusing them. License: https://creativecommons.org/licenses/by-nd/4.0/ ## Accelerating Science with AI: Surfing the Frontier Between the Bitter Lesson and No Free Lunch (talk recap) https://scottclark.io/blog/accelerating-science-with-ai - [Title slide: Accelerating Science with AI: Surfing the Frontier Between the Bitter Lesson and No Free Lunch. Scott Clark, Founder and CEO, with the Distributional and Talaria Scientific logos.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-01) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-01.png (© 2026 Scott Clark) - [Slide: three takeaways. 1. AI x Science can lead to abundance. 2. Researchers will always have a role. 3. Better tools can accelerate this.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-02) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-02.png (© 2026 Scott Clark) - [Slide: what abundance with AI actually means, a four-step build: new scientific discoveries, new materials and more efficient energy, advances in medicine, aerospace and more, living in a sci-fi future.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-06) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-06.png (© 2026 Scott Clark) - [Slide: improving quality of life, not just replacing repetitive work.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-07) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-07.png (© 2026 Scott Clark) - [Slide: we can now meaningfully accelerate science with AI.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-08) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-08.png (© 2026 Scott Clark) - [Slide: scale beats cleverness, but specificity beats generality. Rich Sutton, The Bitter Lesson, 2019; Wolpert and Macready, No Free Lunch Theorems for Optimization, 1997.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-09) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-09.png (© 2026 Scott Clark) - [Figure: the bitter lesson drawn as a rising wave. Problems arranged left to right by how much evidence exists for ones like them, performance on the vertical axis. Three waterlines, t1 to t3, each generation raising the water and washing further right; specialized methods drawn as narrow peaks, the left ones under water since t2, one held out until t3; past the dashed out-of-distribution line a specialized solver still stands above the waterline.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-12) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-12.png (© 2026 Scott Clark) - [Slide: there will always be a frontier. Surf the wave. Don't climb a static mountain that is about to be washed over.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-13) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-13.png (© 2026 Scott Clark) - [Figure: the same wave with two annotations, a gray peak labeled "a static mountain" that the water has already covered, and an amber arrow labeled "surf here" pointing at the edge where the newest waterline meets the out-of-distribution region.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-14) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-14.png (© 2026 Scott Clark) - [Slide: we are still a long way from /goal solve stable fusion containment.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-15) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-15.png (© 2026 Scott Clark) - [Slide: things that were previously impractical are now achievable.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-16) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-16.png (© 2026 Scott Clark) - [Figure: the cost, quality, speed triangle, with the feasible region at its center grown outward past the old edges.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-21) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-21.png (© 2026 Scott Clark) - [Figure: the rising wave again, with a callout reading "Use the best of both worlds to push right" and arrows from the top waterline and from the specialized solver toward the right edge.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-22) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-22.png (© 2026 Scott Clark) - [Slide: I have been trying to solve this problem for the last 20 years of my career, with varying degrees of success.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-24) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-24.png (© 2026 Scott Clark) - [Slide: whatever the field, the final step was always parameter tuning. A timeline strip: Oregon State, finite elements, 2004 to 2008; UC Davis, protein folding, 2006; Max Planck, quantum, 2007; Cornell PhD, applied math, 2008 to 2012; Los Alamos, metagenomics, 2009; Berkeley Lab and JGI, genome assembly, 2010. Methods for efficient optimization given an objective.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-25) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-25.png (© 2026 Scott Clark) - [Figure: a one-dimensional knob on the horizontal axis, the quantity you care about on the vertical, and a handful of expensive evaluations drawn as dots.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-26) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-26.png (© 2026 Scott Clark) - [Figure: the same dots with a Gaussian process fit through them: a mean curve and an uncertainty band that pinches to nothing at the measured points and opens up between them.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-27) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-27.png (© 2026 Scott Clark) - [Figure: below the belief, an acquisition curve (expected improvement) scores every point on the knob; its peak marks the next evaluation.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-28) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-28.png (© 2026 Scott Clark) - [Figure: the new point added, the band collapsed where it was measured, the mean moved, and the acquisition function pointing to the next evaluation.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-29) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-29.png (© 2026 Scott Clark) - [Slide: SigOpt, Bayesian optimization as a service. A timeline strip: founding, 2014, with the Y Combinator and Andreessen Horowitz logos; building, 2015 to 2020, with the American Express, Netflix, and OpenAI logos; scaling, 2020 to 2023, with the Intel logo. Optimization at scale for problems that mattered.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-30) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-30.png (© 2026 Scott Clark) - [Slide: the optimizer worked, except when it didn't. It overfit, or optimized the "wrong" thing.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-33) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-33.png (© 2026 Scott Clark) - [Slide: the best thing about a black box optimizer is that it will optimize anything you tell it to. The worst part about a black box optimizer is that it will optimize exactly what you tell it to.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-35) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-35.png (© 2026 Scott Clark) - [Slide: you need to be able to trust the numbers.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-36) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-36.png (© 2026 Scott Clark) - [Slide: pre-production AI testing, 2023 to 2025. High dimensional Bayesian statistical tests for chaotic and non-stationary AI systems and agents. Stamped: It failed. 1. People didn't know what to test (or have the data before deployment). 2. They couldn't understand tests well enough to act. 3. They shipped to prod anyway.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-40) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-40.png (© 2026 Scott Clark) - [Slide: a second column, post-production agent analytics, 2025 to 2026. Post-production analytics on agent traces to find the evals that you aren't monitoring for. Stamped: It failed. A feature, not a product (let alone a business or a startup).](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-44) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-44.png (© 2026 Scott Clark) - [Slide: knowing what to measure matters. Trusting what you measure matters.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-45) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-45.png (© 2026 Scott Clark) - [Slide: a third column, computational science harness, 2026 and on. A multi-agent, multi-FM harness for computational science research acceleration, using domain specific guardrails rooted in physics and simulation to ensure trustworthiness.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-48) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-48.png (© 2026 Scott Clark) - [Slide: ideas move science forward, but testing them is expensive in researcher time.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-49) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-49.png (© 2026 Scott Clark) - [Figure: from idea to trusted result, where the week goes. Panel 1, today: the scientist walks the whole span by hand, with the science (ideas, taste, judgment) getting the scraps above a gray band labeled the validation schlep: data wrangling, environments, baselines, sweeps, checking. Caption on the slide: AI is now good at the mechanical middle of research.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-51) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-51.png (© 2026 Scott Clark) - [Figure: the second panel of the same figure, in the suit: the scientist holds ideas, taste, and judgment end to end, with no schlep between idea and trusted result; below a dashed boundary labeled "judgment is proposed, never assumed", the harness carries the same five mechanical steps plus the full search recorded, adversarial review, and provenance on every run.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-52) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-52.png (© 2026 Scott Clark) - [Slide: a computational science what-if machine. Start new research from scratch. Respond to reviewers. Convert old academic software. Extend current papers.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-53) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-53.png (© 2026 Scott Clark) - [Figure: the Talaria architecture. A research loop of five actions, idea, plan, implement, run, synthesize, with supporting actions and gates between them, all depositing into a system of record (references, concepts, hypotheses, results, findings, an append-only event journal) kept as markdown and JSONL in a git repository; record inspection, extra rigor, maintenance, and meta-usage actions hang off the record.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-54) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-54.png (© 2026 Scott Clark) - [Slide: receipts. Hypotheses, what you asked, frozen; results, what ran, immutable; findings, what you now believe. Every run pinned to exact code, configuration, and environment, so a rerun is a reproduction. A result with no hypothesis is not allowed to exist, because orphan numbers are how spurious findings start. Kept whether or not it worked. The lab notebook, inverted: a record the work keeps of itself.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-55) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-55.png (© 2026 Scott Clark) - [Slide: a mech suit instead of a replacement. The scientist pushes out of distribution, and the model fills in behind. So what can we do with it?](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-57) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-57.png (© 2026 Scott Clark) - [Slide: every field that simulates the physical world keeps a portfolio of models, from the full-fidelity solvers it trusts to the cheap surrogates it can afford. Surrogate: any cheap model standing in for an expensive one. Fidelity: how faithfully it reproduces the metric you care about, in the regime you care about.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-58) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-58.png (© 2026 Scott Clark) - [Figure: a plane with cost per answer on a log scale across and fidelity up. A handful of points: empirical fits and crude heuristics at the cheap end, a few reduced-order models and response surfaces in the middle, full-order solvers at the expensive end; open circles mark hand-built historical one-offs.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-59) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-59.png (© 2026 Scott Clark) - [Figure: the same plane with a new row of blue points, neural operators and learned surrogates, filling the cheap-but-decent region the classical stack left empty; cheap per answer once trained, valid where validated.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-60) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-60.png (© 2026 Scott Clark) - [Figure: the multi-fidelity loop. Explore with cheap surrogates, exploit what they learn to call the next, costlier rung, verify the answers that matter with the reference solver or the lab; every run lands in the record, cheap or dear, with its provenance, in a model ledger and a run ledger, and the record feeds the next pass.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-62) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-62.png (© 2026 Scott Clark) - [Figure: the same cost-fidelity plane with the frontier drawn as a curve through today's models and a dashed amber curve above it marking where a harness pushes it, with amber circles for the points a harness would add.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-63) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-63.png (© 2026 Scott Clark) - [Slide: the models, validated surrogates on a dense frontier, and the optimizer, the right model per job via multi-fidelity Bayesian optimization. A denser frontier makes a sharper search; the search buys the runs that train and validate the next models; the policy itself can be learned, over surrogates, acquisition functions, and fidelities.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-64) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-64.png (© 2026 Scott Clark) - [Figure: the strategy in four parts. 1, Talaria manufactures validated physics surrogate models; a dense Pareto frontier of accuracy versus cost per job; 2, TauOpt picks the right model per job and optimizes and searches over them; 3, the harness keeps the research reproducible and self-improves the process; 4, the outputs: fluids, materials, and energy search and design, and just-in-time physics models for robotics and world models.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-65) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-65.png (© 2026 Scott Clark) - [Slide: search over surrogates toward what you build next. Materials discovery, energy, physical AI: the classic multi-fidelity design search, with a dense set of validated models to search over parameter and model space.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-66) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-66.png (© 2026 Scott Clark) - [Slide: built in the open, free for open science. Private beta today. Public, open source, and always free for open scientific work. Public launch at NeurIPS, December 2026.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-67) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-67.png (© 2026 Scott Clark) - [Slide: thank you, with QR codes for talariasci.com and scottclark.io. Researching computational physics or applied math? Running physics simulations? Private beta signups, open source launch at NeurIPS 2026, always free for open science.](https://scottclark.io/blog/accelerating-science-with-ai#fig-slide-68) - image: https://scottclark.io/images/blog/accelerating-science-with-ai/slide-68.png (© 2026 Scott Clark) ## How Better Evals Can Bring Abundance through Accelerated Scientific Discovery (talk recap) https://scottclark.io/blog/better-evals-abundance - [Title slide: How Better Evals Can Bring Abundance through Accelerated Scientific Discovery. Scott Clark, Founder and CEO, with the Distributional and Talaria Scientific logos.](https://scottclark.io/blog/better-evals-abundance#fig-slide-01) - image: https://scottclark.io/images/blog/better-evals-abundance/slide-01.png (© 2026 Scott Clark) - [Slide: what abundance with AI means, a four-step build from new scientific discoveries to living in a scifi future.](https://scottclark.io/blog/better-evals-abundance#fig-slide-02) - image: https://scottclark.io/images/blog/better-evals-abundance/slide-02.png (© 2026 Scott Clark) - [Slide: improving quality of life, not just replacing repetitive work.](https://scottclark.io/blog/better-evals-abundance#fig-slide-03) - image: https://scottclark.io/images/blog/better-evals-abundance/slide-03.png (© 2026 Scott Clark) - [Slide: we can now meaningfully accelerate science with AI.](https://scottclark.io/blog/better-evals-abundance#fig-slide-04) - image: https://scottclark.io/images/blog/better-evals-abundance/slide-04.png (© 2026 Scott Clark) - [Slide: I have been trying to solve this problem for the last 20 years of my career, with varying degrees of success.](https://scottclark.io/blog/better-evals-abundance#fig-slide-05) - image: https://scottclark.io/images/blog/better-evals-abundance/slide-05.png (© 2026 Scott Clark) - [Slide: grad school timeline across Oregon State, UC Davis, Max Planck, Cornell, Los Alamos, and Berkeley Lab. Whatever the field, the final step was always parameter tuning.](https://scottclark.io/blog/better-evals-abundance#fig-slide-06) - image: https://scottclark.io/images/blog/better-evals-abundance/slide-06.png (© 2026 Scott Clark) - [Slide: SigOpt, Bayesian optimization as a service, with the founding, building, and scaling timeline.](https://scottclark.io/blog/better-evals-abundance#fig-slide-07) - image: https://scottclark.io/images/blog/better-evals-abundance/slide-07.png (© 2026 Scott Clark) - [Slide: the optimizer worked, except when it didn't. It overfit, or optimized the wrong thing.](https://scottclark.io/blog/better-evals-abundance#fig-slide-08) - image: https://scottclark.io/images/blog/better-evals-abundance/slide-08.png (© 2026 Scott Clark) - [Slide: the best thing about a black box optimizer is that it will optimize any eval you tell it to. The worst part is that it will optimize exactly the eval you tell it to.](https://scottclark.io/blog/better-evals-abundance#fig-slide-09) - image: https://scottclark.io/images/blog/better-evals-abundance/slide-09.png (© 2026 Scott Clark) - [Slide: we are still a long way from typing goal solve stable fusion containment.](https://scottclark.io/blog/better-evals-abundance#fig-slide-10) - image: https://scottclark.io/images/blog/better-evals-abundance/slide-10.png (© 2026 Scott Clark) - [Slide: Distributional attempt one, pre-production AI testing, 2023 to 2025: high dimensional Bayesian statistical tests for chaotic and non-stationary AI systems and agents.](https://scottclark.io/blog/better-evals-abundance#fig-slide-11) - image: https://scottclark.io/images/blog/better-evals-abundance/slide-11.png (© 2026 Scott Clark) - [Slide: the same testing slide grayed out with It failed stamped over it, and the three reasons: people didn't know what to test or have the data before deployment, they couldn't understand tests well enough to act, and they shipped to prod anyway.](https://scottclark.io/blog/better-evals-abundance#fig-slide-12) - image: https://scottclark.io/images/blog/better-evals-abundance/slide-12.png (© 2026 Scott Clark) - [Slide: Distributional attempt two on the timeline strip, post-production agent analytics, 2025 to 2026: analytics on agent traces to find the evals you aren't monitoring for.](https://scottclark.io/blog/better-evals-abundance#fig-slide-13) - image: https://scottclark.io/images/blog/better-evals-abundance/slide-13.png (© 2026 Scott Clark) - [Slide: the pivot slide grayed out with It failed stamped over it. A feature, not a product, let alone a business or a startup.](https://scottclark.io/blog/better-evals-abundance#fig-slide-14) - image: https://scottclark.io/images/blog/better-evals-abundance/slide-14.png (© 2026 Scott Clark) - [Slide: the timeline strip gains its third column, Computational Science Harness, 2026 and beyond: a multi-agent, multi-FM harness for computational science research acceleration, using domain-specific evals and guardrails rooted in physics and math.](https://scottclark.io/blog/better-evals-abundance#fig-slide-15) - image: https://scottclark.io/images/blog/better-evals-abundance/slide-15.png (© 2026 Scott Clark) - [Slide: from idea to trusted result, where the week goes. Two rows compare a computational scientist's week today, dominated by the validation schlep done by hand, with the same week in the harness, where data wrangling, environments, baselines, sweeps, and checking run with provenance on every run.](https://scottclark.io/blog/better-evals-abundance#fig-slide-16) - image: https://scottclark.io/images/blog/better-evals-abundance/slide-16.png (© 2026 Scott Clark) - [Slide: thank you, with QR codes for talariasci.com and scottclark.io. Researching computational physics or applied math? Building world models? Private beta signups, open source launch at NeurIPS 2026, always free for open science.](https://scottclark.io/blog/better-evals-abundance#fig-slide-17) - image: https://scottclark.io/images/blog/better-evals-abundance/slide-17.png (© 2026 Scott Clark) ## Making New Mistakes Faster as a Second-Time Founder https://scottclark.io/blog/make-all-new-mistakes-faster - [The map I was actually on, drawn with axes I could only name later. Vision-market fit is the top-left quadrant: resonance keeps climbing, the pull line never gets crossed, and the quadrant keeps you alive exactly long enough to matter. Rooms lean in, rounds come together, pilots start, and none of it compounds on its own.](https://scottclark.io/blog/make-all-new-mistakes-faster#fig-make-all-new-mistakes-faster) - SVG: https://scottclark.io/figures/make-all-new-mistakes-faster/make-all-new-mistakes-faster.svg (© 2026 Scott Clark; CC BY-ND 4.0) ## Building this blog: Part II https://scottclark.io/blog/building-this-blog-part-ii - [A site written for two readers. The dossier repo holds verified public facts, the resume repo holds bio truth, and the site composes both into two surfaces: HTML for the person, and a machine-readable stack (markdown twins, an entity graph, llms.txt, a full-content feed) for the agent they sent. The connective tissue between repos is markdown contracts. Facts flow one direction.](https://scottclark.io/blog/building-this-blog-part-ii#fig-two-readers) - SVG: https://scottclark.io/figures/building-this-blog-part-ii/fig1-two-readers.svg (© 2026 Scott Clark; CC BY-ND 4.0) --- Source: https://scottclark.io/figures (Scott Clark)