4,715 karma · joined May 4, 2013
https://westurner.org/
These comments are publicly archivable and logged: https://westurner.github.io/hnlog/
[ my public key: https://keybase.io/westurner; my proof: https://keybase.io/westurner/sigs/3t_WnYf465hEbUSdwjMXGIeLieAd_81SDYQUdDUMsI4 ]
Even with a perfect sandbox, an agent can slip in vulnerable code intentionally or accidentally. How different of a problem is that from hiring and HR is that? Idk. Perhaps easier to terminate a model.
"On the Impossible Safety of Large AI Models" (2022) https://arxiv.org/abs/2209.15259
"LLMs + Security = Trouble" (2026) https://arxiv.org/abs/2602.08422
I agree. What does coverage-guided fuzzing fuzz if there is 100% test coverage?
So, then, 100% branch test coverage is not a sufficient metric (because it doesn't indicate whether the code is fuzzed or formally verified for example).
Would Branch coverage even be a sufficient software quality metric if we were to instead measure how many times each branch of code is covered by tests? How to verify that one test which executes 100% of the code and runs only one assertion on, say, a CLI utility exit code integer is actually sufficiently covering?
> I think a codebase generated by AI is actually more understandable than one generated by humans at this point,
From doing a larger port (of sphinx, docutils, myst-md-parser, pygments, to rust in westurner/dsport) with a lot of human in the loop and currently ~80% branch coverage, this seems to be at least initially true but just like real life there's drift from even a good plan that you pay a more expensive model to prepare.
I suppose it's the same challenge as architectural drift in open source non-LLM-assisted products and the solutions are pretty much the same: give better instructions (AGENTS.md,) and use better sufficiency criteria as an engineering manager (branch test coverage, fuzzing, formal methods, TLA+), and train and pay humans to do secure code review.
Sometimes the agent doesn't notice that the code already solves for that and implements its own implementation with tests and it's wastefully redundant when the code should be refactored and the tests should be refactored so that we can delete code in order to minimize bloat.
Unfortunately often, just like IRL software development, the response from the agent is not sufficient to close the issue.
One proposed solution for this that is in retrospect obvious and also essential to success in "normal"/"traditional"/"legacy" (non-AI) engineering projects, is to always verify whether the candidate solution satisfies the criteria;
From "Groundtruth – checks your AI coding agent's claims against the Git diff" https://news.ycombinator.com/item?id=48838209 :
> "Follow up to verify that the work was actually satisfactorily completed"
> Are there other sound management practices that aren't yet effectively implemented in current gen agents?
Oh, and always write tests, docs, commit messages, and changelog entries; but don't waste tokens on documenting something that doesn't verifiably pass sufficient tests.
An eval of this is likely worthwhile;
Re: "Grounded in logic" and "Grounded in theory"
Ground and justify all of the responses with logic and theory and real observations from qualified experiments with citations.
Present a coherent argument borne of logical premises with extant sufficient proven evidence of support. Assess and critique the response given such criteria that all responses should be valid logical arguments, and revise before responding
> The initial phase of Linked Discoveries™ is an experimental pilot resource developed by the National Library of Medicine (NLM) that helps users explore how articles with PubMed® abstracts connect to the surrounding literature. Linked Discoveries creates neighborhoods of related articles, highlights shared conditions, genes, and chemicals across the neighborhood, and surfaces context like NIH funding, citation relationships, retractions, and reviews.
> Linked Discoveries uses data from PubMed, along with related condition, gene, and chemical data linked from other NLM databases like Gene, MedGen, and PubChem®.
So it does clique / principal components / flow analysis of the citation graph? Or GraphRAG and clustering?
OTOH, probably also useful: ChEMBL, UniProt, UpToDate, actual EHR charts, Folding@home, AlphaFold's db's, Co-Scientist that writes hypotheses, the LODcloud Linked Open Data Cloud
Re: cheminf: https://news.ycombinator.com/item?id=32894745
healthcare-ai GitHub topic: https://github.com/topics/healthcare-ai
Which dictionaries use IPA to spell out how to pronounce words?
Pronunciation respelling for English > International Phonetic Alphabet: https://en.wikipedia.org/wiki/Pronunciation_respelling_for_E... alphabet
open-dict-data/ipa-dict: Monolingual wordlists with pronunciation information in IPA; data/en_US.txt : https://github.com/open-dict-data/ipa-dict
Re: language learning: https://news.ycombinator.com/item?id=38317345 :
> [ Phonemic awareness, Phonological reasoning, ]
> What are some of the more evidence-based early literacy / reading curricula? OTOH: LETRS, Heggerty, PAL
Phoneme posters with IPA letters? Speech Pathology?
Re: the new WebMCP spec: https://news.ycombinator.com/item?id=45623194 :
> Could this help with accessibility reviews?
> "Lighthouse accessibility score" https://developer.chrome.com/docs/lighthouse/accessibility/s...
> awesome-a11y > Tools: https://github.com/brunopulis/awesome-a11y/blob/main/topics/...
I've been working with Playwright e2e tests (End to End) in a number of projects lately.
Playwright is a useful and justified abstraction in the testing workflow because: Playwright tests can be run interactively in vscode (with just bwrap sandboxing but unfortunately no VM isolation), Playwright tests can be run in a dedicated local browser or in a headless browser for example in a container on GitHub Actions (a CI Continuous Integration system that runs then tests and saves the output artifacts automatically when any change is pushed to the source code repo), Playwright test can be recorded from an interactive browser session, and Playwright tests can be written by AI prompted to implement 100% branch test coverage with tests and e2e tests with {Playwright or Puppeteer} in {language}.
And then there's the Computer Use model (and the need for process or full VM sandboxing e.g. with ai-sandbox)
Find existing open source agent instructions and agent skills for UI/UX agents that could or should be adapted to a11y testing?
But can there be holograms within the outer hologram?
How much code is needed to implement Classical+PQ (Hybrid PQ) or PQ-only (Only PQ) cipher selection restrictions just?
FWIU, with golang:
# This allows X25519MLKEM768 (Hybrid PQ)
GODEBUG=fips140=on
# This prevents any PQ ciphers from being used:
GODEBUG=fips140=only
But it looks like [MLIR and all other implementations of] SIMD only accept vectors; so there can't be Zero-Copy there because the tensor must (?) be copied to a vector to pass to a SIMD e.g. matmul routine, and then the resultant vector must be copied back into a tensor only if there are subsequent references to the complete tensor instead of just a slice?
FWIU, AFAICS, GPUs are designed for 3x3 tensors (and affine transformation to 2D) but for greater degrees like for 4x4 tensors (e.g. for SQG) you must implement shaders?
From https://news.ycombinator.com/item?id=49855122 re: gravity from QED without any GR spacetime curvature, using the ~amplituhedron to solve n-body gravity with scattering amplitudes, and SQG/DDF:
> normed tensor gaussian splatters work well as a simulation primitive for this too because they do conservation.
N-body gravity derived from QED with scattering amplitudes (without spacetime curvature, in a Flat Minkowski space)
[...]
From this chat: https://share.google/aimode/O0pKsl6nYdv8Nh2fD :
> The "shear-jamming" of Fedi’s fluid space is the physical, hydrodynamical translation of hitting a geometric boundary facet of the amplituhedron
And, normed tensor gaussian splatters work well as a simulation primitive for this too because they do conservation.
Matrices are Tensors but with the matrix product operator instead of the tensor product operator.
Pandas supports MultiIndex DataFrames but the pandas docs recommend xarray for 3D and N-Dimensional data.
xarray supports N-Dimensional data as for example NetCDF but not tensor arithmetic.
xarray_jax: https://github.com/google-deepmind/xarray_jax :
> This library solves that problem. It registers xarray data structures as custom JAX PyTrees. This allows JAX to seamlessly flatten xarray objects into their raw arrays for accelerated computation and then unflatten the results back into fully labeled xarray objects, preserving critical metadata like dimension names and coordinates.
flatten and unflatten with datatypes is necessary for unrolling loops for performance.
Which is the correct logic for probabilistic logic, for expressions with frequentist or symbolic distributions as values? Are quantum logic and quantum statistical mechanics the appropriate or useful tools for all probabilistic logic?
dist_a1 <operator> dist_a2
uncertainties does mean±dev in Python with numpy types.
From https://news.ycombinator.com/item?id=41411280 :
> W3C CSVW supports per-column schema. ( with URIs for datatypes )
> Serialize a dict containing a value with uncertainties and/or Pint (or astropy.units) and complex values to JSON, then read it from JSON back to the same types. Handle datetimes, complex values, and categoricals
IEEE-754 specifies NaN (null), ±0, three infinities (positive, negative, and unsigned), but IEEE-754 does not specify a representation for categoricals, datetimes (like ISO8601), or complex numbers.
XSD (XML Schema Datatypes), which RDFS vocabularies often use to specify the rdfs:range of an rdfs:Property, does not specify how to specify abstract complex numbers; but OpenMath RDF, and QUDT (Quantities, Units, Dimensions, and Types) and OM Ontology all have a way to save complex numbers to disk, too.
Notes re: how IOMMU GPU passthrough with device selection would be a helpful feature to add to QEMU cli, virt-manager,: https://news.ycombinator.com/item?id=46750715 :
> rutabaga_gfx does GPU paravirtualization: https://github.com/magma-gpu/rutabaga_gfx
Here's a picture of that: "Fact-checking Trump’s plaques for past presidents at the White House ‘Walk of Fame’ | PBS News" https://www.pbs.org/newshour/amp/politics/fact-checking-trum...
What regional interior decorating style is that, with the gold in those shapes on the wall?
Selling US media to foreign ownership?
Who else does that?
EA, FIFA, Madden, WB, CNN, Liv Golf, gold leaf on the walls of the people's house
From https://news.ycombinator.com/item?id=48775482 :
> Waveshare has a E6 full color ePaper/eInk/EPD display in 3.6" and 7.3"
A digital picture frame that's wall-powered could act as a node in a Meshtastic or MeshCore LoRA mesh.
> EPS3.9 also had significant anti-tumor effects in the mice with liver cancer and activated anti-tumor immune responses
“A Novel Exopolysaccharide, Highly Prevalent in Marine Spongiibacter, Triggers Pyroptosis to Exhibit Potent Anticancer Effects” (2025) DOI: 10.1096/fj.202500412R https://faseb.onlinelibrary.wiley.com/doi/10.1096/fj.2025004...
"A Gemma model helped discover a new potential cancer therapy pathway" https://news.ycombinator.com/item?id=45604231 :
> eCPMV VNPs + EPS3.9 + [...]
"Scientists are discovering a powerful new way to prevent cancer" https://news.ycombinator.com/item?id=45474404
Notes re: Kidneys not Livers: https://news.ycombinator.com/item?id=47460486 ; gh/topic/healthcare-ai
"Removing these 50 objects from orbit would cut danger from space junk in half" https://news.ycombinator.com/item?id=45469259 :
> When will it be safe and cost-efficient to - instead of deorbiting toward Earth's atmosphere - Capture and Haul and Rendezvous and gently Land orbital scrap on non-earth locations like the Moon or Mars or a thrust-retrofitted asteroid for later processing?
Is there any amount of time in a higher altitude orbit - or an Earth-Moon lunar cycler orbit - that would sanitize the outer surface of a spacecraft like ISS? Or will all such attempts cause adaptation?
Would ISS be more useful as an oxygen tank in earth-moon orbit than in Earth's atmosphere and ocean?
But for the moon,
From https://news.ycombinator.com/item?id=45237426 :
> What about solar sintering + induction heating of lunar regolith?
> Basalt, Glass,
> Molten oxide electrolysis of regolith would yield Oxygen, Silicon, Iron, Aluminum, and Titanium for 3d printing and for solar panels and semiconductors.
> Is there a sensor for inside of a rover or a vehicle wheel to sample the regolith falling within the wheel before it pours out to self clear?
Does this model include a regolith map or a soil map? Where is there Carbon on the moon?
---
"Lunar soil machine developed to build bricks using sunlight" https://news.ycombinator.com/item?id=45086238
---
From https://news.ycombinator.com/item?id=42543663 :
> I wonder if the ISS could instead be scrapped to the moon. [...]
> Because of my love for old kitchens on the Moon.
---
"Researchers simulated 30 million routes to the Moon and found a hidden detour through L1 that saves fuel and keeps spacecraft talking to Earth the whole way" https://spacedaily.com/sd-researchers-simulated-30-million-r... :
> Most of a deep-space journey is coasting along gravitational contours that physicists collectively call the Interplanetary Transport Network — a web of low-energy pathways that connect orbits around planets, moons and Lagrange points
- "Earth–Moon transfer via the L1 Lagrangian point using the theory of functional connections" (2026) https://link.springer.com/article/10.1007/s42064-025-0297-x :
> The results show a reduced fuel consumption of at least 58.80 m/s as compared to the literature. This difference can be useful for mission designers, while the efficiency of the technique can potentially extend the results to other types of transfers in space.
Where are the L1 and L2 Lagrangian points when you consider the mass of the moon in addition to the earth and the sun?
Lyapunov exponents in SQG/DDF too. This month I've been learning about the differences between SQG and DDF Dilatant Dark Fluid (Fedi) in regards to n-body gravity research.
- "Robots – Comprehensive catalog of 115 humanoid and quadruped robots" re: MJCF in mujoco_menagerie, awesomerobots, robot-descriptions/awesome-robot-descriptions: https://news.ycombinator.com/item?id=45727432
- I can think of at least one museum that already has interactive demo robot arms that are in need of repair or replacement
> low-cost retinal imaging attachment: https://www.google.com/search?q=low-cost+retinal+imaging+att...
low-cost funduscopy mobile attachment: https://www.google.com/search?q=low-cost+funduscopy+mobile+a...
> NIRS fundoscopy is clinically useful FWIU
Though not as low-cost.
> Multispectral imaging (oh and faster OCT) in a portable unit would be clinically useful
Wave field cameras capture phase per pixel. That's probably medically usual.
I have been researching convergent holographic medical imaging and treatment using Continuous Wave (CW) laser with longitudinally-polarized beams, possibly at 1 Watt per laser plus the cost to run the active mask and the (Amplituhedron-applied) holographic transform on a DGX for sub-wavelength scale holography.
From working on optical tweezers, I learned that CP-L + CP-R cancels the transverse and leaves a sort of blurry but narrow focal spot of longitudinal wave, and that CP + HP, and HP-L + HP-R have an even more precise focal spot. (CP: Circularly Polarized, HP: Helically-polarized, L/R: Left/Right).
OpenLIFU (low-frequency ultrasound array) is much further developed as a medical imaging and treatment capability.
Things that can be diagnosed from fundoscopy FWIU: MedicalSignOrSymptom(s), ocular disorders, Diabetes, ASD, Schizophrenia, general psychosis,
How to balance privacy of PHI against the advantage of early low-cost diagnosis?
For most ophthalmologic practices, storing fundoscopy in the chart (EHR) to track conditions over time would be an improvement.
Milkdrop3/projectM can vary visualization parameters according to audio attributes samples in real time.
Could be in an open source bowling alley: https://news.ycombinator.com/item?id=48971134
I've my own particle visualizations, after editing the ThreeJS particle attractor example I've started building a collection of these with R3F ReactThreeFiber and Drei in westurner/particlewavefluid: https://github.com/westurner/particlewavefluid
ThreeJS can use WebGPU if available or fallback to WebGL.
I haven't yet implemented NS; I should prioritize building Fedi's SQG and DDF Dilatant Dark Fluid models of black holes with only just one vortex moisture with GOE Gross-Pitaevskii equation.
I did do (ai) review and try to find extensions from the new NS for Incompressible fluids solution. There are connections from Navier-Stokes to Lagrangian mechanics through the Madelung transformation; https://news.ycombinator.com/item?id=49637058
But SQS and DDF say that the Superfluid of space-time is a dilatant fluid that is both compressible and incompressible, and there's a 4th term. So far, I've used 4x4 normed tensor Gaussian splatters to quantize the fluid simulations and it seems advantageous because FWIU the normed tensors already naturally do conservation of energy in the sim.
Also working with sim optimizations on limited architectures;
Re: the first version of MU-AV (in C++) on an ESP32-S3 with a 1Hz ePaper price tag screen; https://news.ycombinator.com/item?id=49060309
A few weeks ago in the CircuitPython implementation of MU-AV (Mu Animation Video) codec and gaussian splatter scene rep, I implemented loop unrolling for a 14X return (actually 100% to 5% utilization on 1 (one) core)) over nested lists in the CPython on x86-64 emulated version for the 156Mhz TI-84 ARM processor dedicated to running Python. But will it work on a really slow processor with no SIMD or GPU shaders?
Did i386 have SIMD? Re: "SIMD-accelerated computer vision on a $2 microcontroller": https://news.ycombinator.com/item?id=40784199
"Show HN: WebGPU Particles Simulation" https://news.ycombinator.com/item?id=38678703
NewsArticle: "Tiny nanolaser could cut computer energy use in half" (2026) https://www.sciencedaily.com/releases/2026/09/260911003858.h...
NewsArticle: "Scientists discover diamonds can generate electricity" (2026) https://www.sciencedaily.com/releases/2026/08/260829035223.h...
A market economist might remark that perhaps that shouldn't be surprising given the lack of intraday rates and the buy-all-sell-all policies.
> In the Netherlands (and the rest of western europe), we have hourly and quarterly energy pricing (opt-in), next to pricing plans that have one or two fixed tariffs during the day/year
So your policies create incentives for residential customers to become net producers of electricity (with resilient local storage)?