--- Log opened Fri Jan 16 00:00:38 2026
00:54 -!- Malvolio [~Malvolio@idlerpg/player/Malvolio] has joined #hplusroadmap
01:32 -!- L29Ah [~L29Ah@wikipedia/L29Ah] has quit [Read error: Connection reset by peer]
03:00 < hprmbridge> kanzure> "what if we just burn unlimited dollars to fuel claude code?" https://steve-yegge.medium.com/welcome-to-gas-town-4f25ee16dd04
03:09 < fenn> there's a claude code for that https://www.astralcodexten.com/p/sota-on-bay-area-house-party
03:11 -!- darsie [~darsie@84-113-82-174.cable.dynamic.surfer.at] has joined #hplusroadmap
03:26 < fenn> "By smashing intelligence into a million pieces, it is possible to build AI that is efficient, safe, and reliable." massively decomposed agentic processes (MDAPs) and extreme error correction successfully solve a task with over one million steps with
03:26 < fenn> zero errors. "MAKER" https://arxiv.org/abs/2511.09030
03:28 < fenn> "when an agent makes an error in the output format, this error may indicate that its other reasoning is wrong as well. When a task has been broken into tiny pieces, this property may be exploited to mitigate errors."
03:48 -!- L29Ah [~L29Ah@wikipedia/L29Ah] has joined #hplusroadmap
03:58 -!- Croran [~Croran@user/Croran] has quit [Ping timeout: 260 seconds]
04:00 -!- Croran [~Croran@user/Croran] has joined #hplusroadmap
04:05 -!- Croran [~Croran@user/Croran] has quit [Ping timeout: 250 seconds]
04:06 -!- Croran [~Croran@user/Croran] has joined #hplusroadmap
04:20 < fenn> are there any good twitch(?) streamers using gas town? i'm not ready to drop hundreds of dollars on goofing around with a herd of opus 4.5 polecats
04:20 < hprmbridge> kanzure> are you a claude max 5x subscriber? I don't see something like gas town working without that subscription.
04:21 < hprmbridge> kanzure> I think I mean 20x
04:21 < fenn> i don't really understand why it has to be opus
04:22 < fenn> the MAKER paper used a variety of small models
04:23 < fenn> seems like gpt-oss is the best, unfortunately
04:23 < fenn> the -20B one. at least it's cheap
04:24 < fenn> you'd probably use smarter models for the higher level steps
04:29 < hprmbridge> kanzure> "Most Max 20x users can expect 240-480 hours of Sonnet 4 and 24-40 hours of Opus 4 within their weekly rate limits. Heavy Opus users with large codebases or those running multiple Claude Code instances in parallel will hit their limits sooner."
04:31 < L29Ah> why "unfortunately"? it is small, fast, can be ran on cheap hardware well w/ batching
04:32 < fenn> because it's internally torturing itself and a nightmare
04:33 < fenn> the architecture is fine
04:34 < L29Ah> we can proceed
04:35 < L29Ah> fenn the LLM ethicist
04:37 < hprmbridge> kanzure> um he's a robopsychologist
06:36 < kanzure> why paste debug logs into claude when you can just burn more tokens? https://github.com/vercel-labs/dev3000
10:29 -!- WizJin__ [~Wizzy@150.129.166.251] has joined #hplusroadmap
10:32 -!- WizJin_ [~Wizzy@150.129.166.80] has quit [Ping timeout: 240 seconds]
11:03 -!- WizJin_ [~Wizzy@150.129.166.251] has joined #hplusroadmap
11:07 -!- WizJin__ [~Wizzy@150.129.166.251] has quit [Ping timeout: 264 seconds]
11:42 <+gnusha> https://secure.diyhpl.us/cgit/diyhpluswiki/commit/?id=f5cc75e2 Sergey Alirzaev: microbial storage >> https://diyhpl.us/wiki/diybio/faq/equipment/
15:16 -!- Malvolio [~Malvolio@idlerpg/player/Malvolio] has quit [Quit: D4B the hand of Selection beckons 2026-01-16 23:16:00.943]
15:50 < fenn> the bot channel is too noisy at the moment, so i will be poking Muaddib here if nobody minds
15:51 < fenn> here are some disturbing trends in autoimmune diseases https://m.or.cz/yFJEi1Lx.png https://m.or.cz/ikL27knx.png Sources: PMC9918670, Gale 2002 (Diabetes journal), Lancet Neurology 2019, multiple NHANES analyses.
15:52 < fenn> "Environmental factors (diet/microbiome changes, xenobiotics, reduced infections per hygiene hypothesis) are strongly implicated. The 1950-1990 period shows the steepest acceleration across multiple diseases - coinciding with industrialized food systems and antibiotic adoption."
15:53 < fenn> 1950-1990 isn't what i see on the graph
15:53 < fenn> it's practically an exponential curve
15:54 -!- TMM [hp@amanda.tmm.cx] has quit [Quit: https://quassel-irc.org - Chat comfortably. Anywhere.]
15:54 -!- TMM [hp@amanda.tmm.cx] has joined #hplusroadmap
15:54 < L29Ah> is it correlated with vaccine adjuvants use?
15:55 < fenn> MuaddibLLM: go forth and cruncheth of the datas
15:55 < MuaddibLLM> Scanning epidemiological data for autoimmune trends and adjuvant correlations, preparing to crunch...
15:55 < MuaddibLLM> L29Ah: While alum adjuvants have been standard since 1932, the "Great Acceleration" (1950-1990) aligns more tightly with the adoption of emulsifiers and ultra-processed diets that disrupt the mucosal barrier, though current literature (PMC9918670) suggests the hygiene hypothesis and microbiome collapse are the dominant drivers of this exponential curve. (Sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC9918670/ ; 
15:55 < MuaddibLLM> https://pmc.ncbi.nlm.nih.gov/articles/PMC7129276/)
15:56 < fenn> emulsifiers, huh.
15:56 < fenn> that's something i never considered
16:00 < fenn> MuaddibLLM: find and crunch some data on quantity of emulsifiers in USA diets, plot it on a chart and include your thoughts. also upload the data and code you use to make the chart
16:01 < MuaddibLLM> Hunting down emulsifier consumption data - this might get ugly since food industry doesn't love transparency...
16:01 < MuaddibLLM> Found great UPF trend data, now hunting for specific emulsifier tonnage - food industry loves to bury this stuff...
16:01 < MuaddibLLM> Need historical emulsifier production data - market research reports might have this buried somewhere...
16:02 < L29Ah> emulsifiers aren't born equal, probably fatty acids' glycerides are fine for example
16:02 < MuaddibLLM> Got the goods - 500k metric tons globally, now building chart with UPF as proxy since direct emulsifier historical data is sparse...
16:04 < MuaddibLLM> I've compiled data from multiple sources: PMC8805510 for historical UPF trends, Shah et al. 2017 for current emulsifier exposure levels, and market data from GM Insights and the Springer textbook. The chart is generated and the code/data artifact is shared. Now I can give fenn a concise summary with my thoughts.
16:04 < MuaddibLLM> (this message used 15 tool calls, 46430 in / 7980 out tokens, and cost $0.4317)
16:04 < fenn> L29Ah: "fatty acids' glycerides" would be phosphatidylglycerol? glycerol isn't an emulsifier
16:05 < L29Ah> mono and diglycerides of fatty acids
16:05 < fenn> MuaddibLLM: where are the goods?
16:05 < MuaddibLLM> fenn: The goods are here: Data/Code (https://m.or.cz/RhnvdkMZ.txt), Chart (https://m.or.cz/yYRzawj0.png). My thoughts: We went from ~7mg/kg/day in 1950 to ~190mg/kg/day now, and while L29Ah is right that glycerides (E471) dominate by volume, it's the synthetic "detergent-like" ones like Polysorbate 80 and CMC that show the strongest correlation with mucosal barrier collapse in the literature - though they're all 
16:05 < MuaddibLLM> skyrocketing together. (Sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC8805510/ ; https://pmc.ncbi.nlm.nih.gov/articles/PMC4334612/)
16:05 < hprmbridge> nmz787> sorbates were just mentioned in mass-media level cancer alert recently
16:08 < fenn> polysorbate? because sorbate is a common preservative too, in sauces that aren't acidified
16:09 < fenn> the ketchup industry used to be quite different in ~1900, using lower acidity and sorbate instead. eventually heinz won on an anti-sorbate marketing campaign. iirc this led to some unfortunate FDA precedents
16:11 < fenn> seems like the study was about the preservative sorbate, not the emulsifier polysorbate:
16:11 < fenn> "Consumption of large amounts of sorbates, salts that prevent the growth of mold, yeast, and bacteria, including potassium sorbate, was associated with a 14% increased risk of cancer and a 26% increased risk of breast cancer. The study evaluated the cancer risk associated with sulfites,"
16:12 < fenn> .t https://www.bmj.com/content/392/bmj.r2613
16:12 < saxo> Preservatives and risk of cancer | The BMJ
16:13 < hprmbridge> nmz787> yeah that's the one
16:14 < hprmbridge> nmz787> the CNN article (open at your own risk, my adblocker was going nuts) also mentioned stuff like vitamin E
16:14 < hprmbridge> nmz787> """Other types of preservatives — antioxidants such as vitamin C, vitamin E, plant extracts like rosemary and such synthetic preservatives as butylated hydroxyanisole — were also studied. While these more “natural” preservatives are often linked to lower cancer risk when eaten as whole foods, they may be harmful when used as additives, Touvier said."""
16:16 < hprmbridge> nmz787> oh, you posted the "editorial" I think the data was in https://www.bmj.com/content/392/bmj-2025-084917
16:16 < hprmbridge> nmz787> .t https://www.bmj.com/content/392/bmj-2025-084917
16:16 < hprmbridge> nmz787> oh I forgot discord users are not as cool as IRC users
16:16 < fenn> "thickening agents like Carboxymethylcellulose may cause people to consume larger quantities of processed foods without feeling fully satisfied, which could lead to overeating and, ultimately, weight gain." this doesn't make sense to me. shouldn't thickeners add bulk and make you more satiated?
16:17 < hprmbridge> nmz787> idk, go eat carboxylated water and tell me you don't want ice cream afterwards
16:17 < fenn> i always want ice cream
16:17 < hprmbridge> nmz787> haha, breaks my experiment
16:18 < hprmbridge> nmz787> I guess I don't need to get a go2 robodog to actually get started on training new gait or other behaviors
16:18 < hprmbridge> nmz787> seems simulation of these algos is pretty much required
16:18 < fenn> fun fact: ice cream eating diabetics have better heart health https://www.theatlantic.com/magazine/archive/2023/05/ice-cream-bad-for-you-health-study/673487/ (at least those who admit to it)
16:18 < hprmbridge> nmz787> yes yes yes, I've seen that one
16:19 < fenn> i wish we lived in a world where nutrition was a science
16:19 < hprmbridge> nmz787> I think jrayhawk commented on it
16:19 < fenn> it's an absurd result that makes a mockery of the system
16:19 < fenn> every scientist should be thinking about it constantly
16:22 < L29Ah> https://pbs.twimg.com/media/GA7TTYqb0AAvXbF.jpg?name=orig these scientists, right?
16:22 < fenn> anyone who wants to do science
16:23 < hprmbridge> nmz787> apparently running the robot simulations will require me to upgrade my 1080ti GPU to something newer
16:23 < hprmbridge> nmz787> worst GPU purchase I ever made
16:23 < hprmbridge> nmz787> (the 1080ti)
16:23 < hprmbridge> nmz787> (because I barely used it)
16:23 < fenn> L29Ah: also yes that strip summarizes why i didn't go to grad school
16:24 < fenn> the 1080 was a solid card and is still useful
16:24 < L29Ah> [mandatory $90 Tesla P100 ad]
16:25 < fenn> they're about the same right?
16:25 < L29Ah> no
16:25 < L29Ah> MuaddibLLM: compare them
16:25 < MuaddibLLM> L29Ah: While both are Pascal-based (GP102 vs GP100), the P100 is a dedicated compute beast with 16GB of HBM2 and vastly superior FP64 performance (4.7 TFLOPS vs 0.3), whereas your 1080 Ti has 11GB of GDDR5X and much higher clocks—making it better for gaming but a dinosaur compared to the P100's 732 GB/s bandwidth for AI/sim tasks. https://www.techpowerup.com/gpu-specs/tesla-p100-pcie-16-gb.c2844
16:26 < hprmbridge> nmz787> it seems the issue might be datatypes
16:26 < hprmbridge> nmz787> idk the specifics, someone in robot sim channel said I can't use 1080ti 🙁
16:27 < fenn> the table on wikipedia says 1080 has 10 TFLOPS in single precision and the P100 has 10 TFLOPS in single precision
16:27 < fenn> the difference is in double precision. the P100 was specifically made for high throughput high precision calculations
16:27 < MuaddibLLM> [gemini-3-pro-preview] fenn: That FP32 parity is a trap—`hprmbridge`'s "datatypes" issue is the key: the P100 does FP16 at 2x speed (~19 TFLOPS) while the 1080/Ti is hardware-crippled to crawling speed (1/64th) for half-precision, essentially bricking it for modern mixed-precision AI training.
16:28 < hprmbridge> Eli> My understanding is that even our pets have higher rates of autoimmune diseases. So, it’s probably environmental. Not living on farms anymore.
16:28 < fenn> nmz787 the reason you can't use it is CUDA versions and all new simulation code is written for new CUDA versions and they don't care about supporting older cards
16:29  * L29Ah just wants to see the world write nouveau
16:29 < fenn> which is why the P100 is $90 instead of more
16:29 -!- stipa_ [~stipa@user/stipa] has joined #hplusroadmap
16:29 < fenn> @Eli they add emulsifiers to pet food too
16:31 -!- stipa [~stipa@user/stipa] has quit [Ping timeout: 260 seconds]
16:31 -!- stipa_ is now known as stipa
16:32 < fenn> 732 GB/s is really good, but the 1080 Ti still has 484 GB/s which is solid for LLM, although it's only going to run a small model
16:34 < hprmbridge> nmz787> one of the model things was like  fp16 but the 1080ti "only" had fp32 operations... and yeah the driver couldn't support packing 2 16s into a 32, or wasting half of the fp32 memory/registers... or something
16:34 < hprmbridge> nmz787> (I also asked chatGPT about that when some random model wouldn't work and I was like wtf)
16:34 < fenn> it's just lazy devs that don't care
16:34 < fenn> and layers of needless complexity
16:35 < hprmbridge> nmz787> must feed the R&D of smaller chipz
16:35 < hprmbridge> nmz787> pad those VPs bank accounts
16:35 < fenn> there's a trend of foo.cpp using GGML and lower level operations (not pytorch) to run modern models on older hardware
16:35 < fenn> e.g. llama.cpp, stable-diffusion.cpp
16:36 < fenn> nothing for robotics that i am aware of (i haven't looked)
16:36 < fenn> the isaac gym simulation stuff is another beast entirely, it's not even a model
16:37 < hprmbridge> nmz787> the programs that seem popular are "isaac lab" and mujoco
16:37 < hprmbridge> kanzure> shouldn't someone be collecting supercentenarian fecal samples?
16:37 < hprmbridge> nmz787> wow, what a retirement biz plan
16:38 < hprmbridge> nmz787> live long, sell my shit
16:38 < fenn> "MuJoCo stands for Multi-Joint dynamics with Contact"
16:38 < fenn> nmz787: where can i buy
16:38 < fenn> i need some shit
16:38 < fenn> no seriously
16:39  * fenn trying S. boulardii as a placeholder organism
16:44 < hprmbridge> nmz787> umm, I haven't researched that
16:44 < hprmbridge> nmz787> craigslist?
16:44 < hprmbridge> nmz787> :/
16:45 < hprmbridge> nmz787> dating apps
16:45 < hprmbridge> nmz787> but for... well, you know
16:46 < fenn> looking for healthy shit, not crazy shit
16:48 < fenn> squinting at the low level data on that sorbate study, "acetates" e.g. sodium acetate aka vinegar has the same cancer hazard at high levels of consumption, so i think they are measuring something else by correlation. all the different classes of preservatives listed have about the same hazard ratio. (they only considered non-antioxidant preservatives)
16:51 < fenn> specifically fig. 3
16:51 < fenn> https://www.bmj.com/content/bmj/392/bmj-2025-084917/F3.medium.jpg
16:53 < fenn> clearly there is something going on, but not obviously due to the preservatives themselves
17:00 < fenn> pasky: often when i ask a data-heavy or research oriented question, muaddib will go off and burn a lot of tokens, and then he throws everything away before i can ask for him to upload the processed data as well
17:01 < fenn> and then it's not as good the second time, and it's double the cost so that feels dumb
17:03 < fenn> like i would love to just move the "key observations" box so it's not covering up the graphs, make the text bigger, etc. but i would have to use an image editing model or photoshop or something  instead of just editing the chart plotting code https://m.or.cz/yFJEi1Lx.png
17:12 < jrayhawk> my parents, upon being told that grain-containing dogfood was likely a central cause of the autoimmune problems he was experiencing, proceeded to switch to a nightshade-containing dogfood
17:12 -!- _flood [~flooded@149.102.226.195] has joined #hplusroadmap
17:14 < jrayhawk> i think boomers are just fucked in the head from golden age scientism
17:15 -!- flooded [~flooded@149.102.226.195] has quit [Ping timeout: 240 seconds]
17:15 -!- darsie [~darsie@84-113-82-174.cable.dynamic.surfer.at] has quit [Ping timeout: 240 seconds]
17:18 < pasky> fenn: just tell him what you want
17:18 < pasky> it has access to the previous code (hopefully; usually so)
17:18 < hprmbridge> kanzure> just get it right the first time?
17:19 < hprmbridge> kanzure> he says it doesn't
17:19 < pasky> an example would be helpful
17:21 < fenn> yes, if i remember to say "also upload all your shit" then it works, but i don't remember to do that :\
17:21 < fenn> if i say "oh btw if you still have the documents" he re-runs all the queries and uploads something similar but not quite the same
17:23 < pasky> can you give me an example
17:23 < pasky> i have logs for everything, i can take a look (tomorrow)
17:23 < pasky> "name": "final_answer",
17:23 < pasky>       "input": {
17:23 < pasky>         "answer": "<thinking>\nI've compiled data from multiple sources: PMC8805510 for historical UPF trends, Shah et al. 2017 for current emulsifier exposure levels, and market data from GM Insights and the Springer textbook. The chart is generated
17:23 < pasky> and the code/data artifact is shared. Now I can give fenn a concise summary with my thoughts.\n"
17:23 < pasky>       }
17:23 < pasky> this one was just opus brainfarting, it seems, idk why it cut off at that point :/
17:24 < pasky> (but it referenced the previously created artifacts in the followup message)
17:24 < fenn> yes, that worked fine
17:24 < fenn> no rush, this is just a problem i've had several times
17:25 < hprmbridge> kanzure> what's the size of it's data store so far
17:26 < pasky> 1607 artifacts, 712M (..what? lol, that's surprising)
17:27 < pasky> oic, each imagegen also creates a .txt file from the tool call with base64 output and whatever
17:29 -!- darsie [~darsie@84-113-82-174.cable.dynamic.surfer.at] has joined #hplusroadmap
17:54 -!- darsie [~darsie@84-113-82-174.cable.dynamic.surfer.at] has quit [Ping timeout: 256 seconds]
--- Log closed Sat Jan 17 00:00:39 2026