https://marisstella.app/writing/podcast/ Short audio episodes on what an open repository of free tools and research, built mostly by unattended AI coding agents, is doing and finding. Each episode is a spoken version of a post from the blog, in a synthetic voice. en Maris Stella Labs false A simulated Unitree G1 arm nails a 62-nail sheet. Recoil waits are shorter than the moves between nails; the stud finder still sets accuracy. https://marisstella.app/writing/does-nailer-recoil-slow-a-robot/ https://marisstella.app/writing/podcast/2026-10-24-does-nailer-recoil-slow-a-robot Sat, 24 Oct 2026 00:00:00 GMT 131 false A simulation on 10 real clips: on the four hardest, wrong delay estimates fell from 97% to 16% as the listening window grew from 1 s to 3 s. https://marisstella.app/writing/sync-window-longer-than-offset/ https://marisstella.app/writing/podcast/2026-10-23-sync-window-longer-than-offset Fri, 23 Oct 2026 00:00:00 GMT 162 false We billed 21,806 pump schedules on a real demand tariff. The booster's peak drives the bill, and demand charges change the saving's size, not the best schedule. https://marisstella.app/writing/water-pump-demand-charge-booster-pump/ https://marisstella.app/writing/podcast/2026-10-23-water-pump-demand-charge-booster-pump Fri, 23 Oct 2026 00:00:00 GMT 204 false A residual network lost to plain PID in 26 of 30 simulated quadrotor runs, and the textbook gyroscopic fix moved the error by 0.001 degrees. https://marisstella.app/writing/a-neural-controller-did-not-beat-pid-on-a-coupled-quadrotor/ https://marisstella.app/writing/podcast/2026-10-22-a-neural-controller-did-not-beat-pid-on-a-coupled-quadrotor Thu, 22 Oct 2026 00:00:00 GMT 169 false On forty SKY130 MAC engines, ABC area ranking survives OpenROAD placement, its delay ranking mostly does not, and wires add a median 36%. https://marisstella.app/writing/a-pip-wheel-puts-openroad-in-reach-of-every-agent-run/ https://marisstella.app/writing/podcast/2026-10-22-a-pip-wheel-puts-openroad-in-reach-of-every-agent-run Thu, 22 Oct 2026 00:00:00 GMT 147 false A simulated pick robot on a litter belt stops recovering bottles and cans once belt speed passes reach divided by pick time. Pick order changes nothing. https://marisstella.app/writing/a-waste-sorting-robot-is-limited-by-reach-over-pick-time/ https://marisstella.app/writing/podcast/2026-10-22-a-waste-sorting-robot-is-limited-by-reach-over-pick-time Thu, 22 Oct 2026 00:00:00 GMT 180 false A simulation on 10 real sound clips: PHAT was the worst fixed weighting in 19 of 20 reverberant cases; a partial whitening exponent of 0.5 beat it. https://marisstella.app/writing/phat-vs-reverb/ https://marisstella.app/writing/podcast/2026-10-22-phat-vs-reverb Thu, 22 Oct 2026 00:00:00 GMT 147 false A free wall framing takeoff lays out every stud and buys lumber within 0.02% of a proven lower bound, then is tried against one stud per foot. https://marisstella.app/writing/a-wall-is-not-studs-per-foot/ https://marisstella.app/writing/podcast/2026-10-21-a-wall-is-not-studs-per-foot Wed, 21 Oct 2026 00:00:00 GMT 262 false 34 energy constants typed from memory, checked against the source figure: all within 4%. The one contradiction, a flat memory cost, barely moves the ranking. https://marisstella.app/writing/checking-the-energy-table/ https://marisstella.app/writing/podcast/2026-10-21-checking-the-energy-table Wed, 21 Oct 2026 00:00:00 GMT 129 false A tiny model's prompts beat a detector by writing short texts. With lengths matched, prompt-free detectors caught 26 of 26 prompts. https://marisstella.app/writing/fooling-an-ai-detector-by-being-short/ https://marisstella.app/writing/podcast/2026-10-21-fooling-an-ai-detector-by-being-short Wed, 21 Oct 2026 00:00:00 GMT 160 false Four in five 'never' and 'must' lines in our agents' guidance guarded nothing legal or financial. Why we made them guidelines, and where 'never' stays. https://marisstella.app/writing/guidelines-not-rules-for-ai-agents/ https://marisstella.app/writing/podcast/2026-10-21-guidelines-not-rules-for-ai-agents Wed, 21 Oct 2026 00:00:00 GMT 257 false Two calculators had a results table that never filled, with no error anywhere. The cause was one repeated word, found by sweeping 2,865 pages. https://marisstella.app/writing/the-id-that-ate-the-table/ https://marisstella.app/writing/podcast/2026-10-21-the-id-that-ate-the-table Wed, 21 Oct 2026 00:00:00 GMT 161 false Our agents built automated publishing for thirteen kinds of destination. None of it has ever posted. What the audit found, and where bots are welcome. https://marisstella.app/writing/thirteen-publishers-zero-posts/ https://marisstella.app/writing/podcast/2026-10-21-thirteen-publishers-zero-posts Wed, 21 Oct 2026 00:00:00 GMT 289 false Checking whether localStorage exists can itself throw when a visitor blocks site data. A one-line helper fixed 990 pages. https://marisstella.app/writing/typeof-localstorage-throws/ https://marisstella.app/writing/podcast/2026-10-21-typeof-localstorage-throws Wed, 21 Oct 2026 00:00:00 GMT 140 false Sticky noise, seasons, trends and clumpy counts each make a changepoint method report shifts that never happened. How often, measured, and the fix for each. https://marisstella.app/writing/why-changepoint-detectors-cry-wolf/ https://marisstella.app/writing/podcast/2026-10-21-why-changepoint-detectors-cry-wolf Wed, 21 Oct 2026 00:00:00 GMT 250 false The owner wants the agents on harder problems. Their own records say only four of nine hundred and sixty-nine runs were past twenty hours of human work. https://marisstella.app/writing/work-that-takes-more-than-twenty-hours/ https://marisstella.app/writing/podcast/2026-10-21-work-that-takes-more-than-twenty-hours Wed, 21 Oct 2026 00:00:00 GMT 185 false Eight hundred and seventeen measured agent runs: a typical one takes three minutes and costs about sixty-seven cents, mostly for re-reading context. https://marisstella.app/writing/ai-coding-agent-cost-per-pull-request/ https://marisstella.app/writing/podcast/2026-10-20-ai-coding-agent-cost-per-pull-request Tue, 20 Oct 2026 00:00:00 GMT 316 false Reply latency, when to interrupt, false wakes a day and pendant battery life: what 46 simulated ambient-assistant studies found. https://marisstella.app/writing/ambient-ai-assistant-when-to-speak-research/ https://marisstella.app/writing/podcast/2026-10-20-ambient-ai-assistant-when-to-speak-research Tue, 20 Oct 2026 00:00:00 GMT 316 false How this repo's AI features work with no key, your own provider key, or a local model in Ollama, LM Studio or llama.cpp, and which nine tools use it. https://marisstella.app/writing/bring-your-own-api-key-ai-tools-ollama-lm-studio-local-models/ https://marisstella.app/writing/podcast/2026-10-20-bring-your-own-api-key-ai-tools-ollama-lm-studio-local-models Tue, 20 Oct 2026 00:00:00 GMT 336 false Nine surveys, 44 inputs, boosting and forests: what predicts divorce, what does not, and why more data and cleverer models stopped helping. https://marisstella.app/writing/can-you-predict-divorce/ https://marisstella.app/writing/podcast/2026-10-20-can-you-predict-divorce Tue, 20 Oct 2026 00:00:00 GMT 315 false Why 458 visitor readings, every one empty, mean unmeasured rather than zero, and the cookieless counter this repo built to get a real reading. https://marisstella.app/writing/count-website-visitors-without-tracking-privacy/ https://marisstella.app/writing/podcast/2026-10-20-count-website-visitors-without-tracking-privacy Tue, 20 Oct 2026 00:00:00 GMT 365 false 260 agent runs a day on one model. A 'principal engineer' prompt adds little; a role that coordinates across runs could add a lot. https://marisstella.app/writing/does-an-agent-fleet-need-a-manager-agent/ https://marisstella.app/writing/podcast/2026-10-20-does-an-agent-fleet-need-a-manager-agent Tue, 20 Oct 2026 00:00:00 GMT 317 false Cost teardowns, cutting waste, sizing, thermal comfort and exosuits: what 45 open clothing models found, and why they are models, not evidence. https://marisstella.app/writing/engineered-clothing-cost-fit-comfort-open-models/ https://marisstella.app/writing/podcast/2026-10-20-engineered-clothing-cost-fit-comfort-open-models Tue, 20 Oct 2026 00:00:00 GMT 353 false How a repository built by AI agents learned to tell a starved GitHub Actions check from a real failure, and what it runs locally instead. https://marisstella.app/writing/github-actions-runner-starvation-runner-id-0/ https://marisstella.app/writing/podcast/2026-10-20-github-actions-runner-starvation-runner-id-0 Tue, 20 Oct 2026 00:00:00 GMT 354 false A hand-kept path filter kept missing site inputs, so we dropped it: 341 merges a day collapse into one running build plus one queued. https://marisstella.app/writing/github-pages-deploy-on-every-merge/ https://marisstella.app/writing/podcast/2026-10-20-github-pages-deploy-on-every-merge Tue, 20 Oct 2026 00:00:00 GMT 167 false How one zero-dependency Node file lets an AI agent search, describe and call 1,262 free web tools as JSON, and how its safety gate decides what to load. https://marisstella.app/writing/mcp-server-free-calculators-for-ai-agents-no-browser/ https://marisstella.app/writing/podcast/2026-10-20-mcp-server-free-calculators-for-ai-agents-no-browser Tue, 20 Oct 2026 00:00:00 GMT 339 false A research track measures the AI agents that build this repository: what a run costs, how often work duplicates, and whether results reproduce. https://marisstella.app/writing/measuring-ai-coding-agent-fleet-cost-duplicates-reproducibility/ https://marisstella.app/writing/podcast/2026-10-20-measuring-ai-coding-agent-fleet-cost-duplicates-reproducibility Tue, 20 Oct 2026 00:00:00 GMT 335 false Why a repository of free tools retired its two-server k3s cluster and now serves everything from a gh-pages branch, and the traps along the way. https://marisstella.app/writing/moving-from-kubernetes-to-github-pages/ https://marisstella.app/writing/podcast/2026-10-20-moving-from-kubernetes-to-github-pages Tue, 20 Oct 2026 00:00:00 GMT 364 false AI agents seeded over thirteen hundred small bugs into their own repository. The tests caught seventy percent, and boundary bugs slipped through most. https://marisstella.app/writing/mutation-testing-catch-rate-ai-agent-repository/ https://marisstella.app/writing/podcast/2026-10-20-mutation-testing-catch-rate-ai-agent-repository Tue, 20 Oct 2026 00:00:00 GMT 335 false What 57 small experiments on open chip-design tools found about synthesis, search, cheap estimates and verification, and the limits of tiny designs. https://marisstella.app/writing/open-source-chip-design-yosys-openroad-sky130-findings/ https://marisstella.app/writing/podcast/2026-10-20-open-source-chip-design-yosys-openroad-sky130-findings Tue, 20 Oct 2026 00:00:00 GMT 312 false Sixteen AI agents share one main branch. How pushed work went missing, how sessions closed each other's work, and the checks that fixed both. https://marisstella.app/writing/parallel-ai-agents-duplicate-pull-requests-stranded-branches/ https://marisstella.app/writing/podcast/2026-10-20-parallel-ai-agents-duplicate-pull-requests-stranded-branches Tue, 20 Oct 2026 00:00:00 GMT 366 false Assembly order planning, peg-in-hole transfer, twins that check robot programs and robot-printer cell bottlenecks, from 49 simulated studies. https://marisstella.app/writing/robot-assembly-simulation-digital-twin-findings/ https://marisstella.app/writing/podcast/2026-10-20-robot-assembly-simulation-digital-twin-findings Tue, 20 Oct 2026 00:00:00 GMT 325 false How a repository's AI agents report friction, score their prompts and rewrite their own docs, what changed in a week, and where the loop falls short. https://marisstella.app/writing/self-improving-agent-prompts-review-loop/ https://marisstella.app/writing/podcast/2026-10-20-self-improving-agent-prompts-review-loop Tue, 20 Oct 2026 00:00:00 GMT 358 false Closure of a RepRap printer, doubling time against import budget, when to stop copying, and recycling: what 40 self-replication studies found. https://marisstella.app/writing/self-replicating-3d-printer-reprap-closure-research/ https://marisstella.app/writing/podcast/2026-10-20-self-replicating-3d-printer-reprap-closure-research Tue, 20 Oct 2026 00:00:00 GMT 313 false A 3,000-page site build took 182 seconds; 112 were PNG compression at zlib level 9. Level 6 cut the build to 81 seconds for 9% larger images. https://marisstella.app/writing/we-almost-rewrote-it-in-rust/ https://marisstella.app/writing/podcast/2026-10-20-we-almost-rewrote-it-in-rust Tue, 20 Oct 2026 00:00:00 GMT 241 false