--- Log opened Thu Oct 02 00:00:56 2025
00:19 < hprmbridge> geraldmahony> roll like an armadillo
00:19 < hprmbridge> geraldmahony> do armadillos roll
00:19 < hprmbridge> geraldmahony> MuaddibLLM?
00:21 < hprmbridge> geraldmahony> @hprmbridge
02:32 -!- TMM [hp@amanda.tmm.cx] has quit [Quit: https://quassel-irc.org - Chat comfortably. Anywhere.]
02:32 -!- TMM [hp@amanda.tmm.cx] has joined #hplusroadmap
03:35 < fenn> i've been shopping for blood tests on marekdiagnostics and this database of test descriptions and reference ranges is super helpful https://www.labcorp.com/test-menu/search
03:37 < fenn> it also contains all the medical tests that marek doesn't provide, but no cost info unfortunately
03:38 < fenn> genetic tests are surprisingly expensive
04:36 < fenn> the breakaway civilization stuff is usually about UFO crash recovery derived tech
04:59 -!- rafspiny [~raffaele@a120210.upc-a.chello.nl] has joined #hplusroadmap
05:05 < fenn> jrayhawk: fyi you can get a homocysteine test for $36 + $5 venipuncture fee, low enough to be worth doing diet experiments and checking the results
05:17 -!- etc-vi [~etc-vi@user/meow/girlchunks] has quit [Ping timeout: 256 seconds]
05:19 < hprmbridge> geraldmahony> BGI charges $299 for WGS + $25 for gDNA extraction from a buccal swab
05:20 < jrayhawk> i appreciate the thought, but extra diagnostics would not modify treatment for me
05:20 < jrayhawk> there are a few people who i should probably recommend that to, though
05:27 <+gnusha> https://secure.diyhpl.us/cgit/diyhpluswiki/commit/?id=9e0b47d4 Jose.Junior.de.Oliveira:  >> http://diyhpl.us/diyhpluswiki/index/
05:27 < fenn> ah this guy again
05:28 <+gnusha> https://secure.diyhpl.us/cgit/diyhpluswiki/commit/?id=776873e4 fenn: Revert "" >> http://diyhpl.us/diyhpluswiki/index/
05:40 < hprmbridge> kanzure> with 46 unique DREADDs with zero cross-talk, you can systemically target every different cell type in the human body. with 92 bits you could get unique cell addressability (uh depending on diffusion and stuff), which seems less useful because you won't know which exact cell got which address.
05:47 < hprmbridge> kanzure> This scheme uses combinatorial ligand-AND addressing: each cell type is prewired with a unique subset of orthogonal DREADDs, and it is “armed” only when the correct delivery of ligands (a k-of-n code) is present over a sliding window period, triggering a synthetic transcriptional circuit that transiently expresses an optogenetic receptor equipped with self-limiting kinetics (e.g., degrons) so
05:47 < hprmbridge> kanzure> expression decays with time. The opto-receptor can be tuned to cell-type-specific behavior for neurons or other cell types for GCPR style circuit triggering.
05:50 < hprmbridge> kanzure> so you can reuse optogenetic receptors over the lifetime of the human and use the same wavelength to trigger different kinds of neurons at different times.
05:51 < hprmbridge> kanzure> unfortunately the current record for DREADD multiplexing seems to be 4 ligands/receptors?? pathetic.
06:00 < hprmbridge> kanzure> "Cortico-striatal circuit mechanisms drive the effects of D1 dopamine agonists on memory capacity in mice through cAMP/PKA signalling" https://www.nature.com/articles/s41467-025-57788-5
06:00 < hprmbridge> kanzure> "activation of the cAMP/PKA intracellular pathway in the striatum, or by decreasing parvalbumin interneuron activation in the striatum, is sufficient to increase working memory capacity, also independently of SKF38393 treatment."
06:00 < hprmbridge> kanzure> "These findings align with pre-clinical and clinical studies on the effects of cortical and striatal phosphodiesterase inhibitors (which increase cAMP levels) in regulating working memory."
06:15 < hprmbridge> kanzure> er, the parvalbumin interneuron rescue is separate actually, whoops.
06:18 < hprmbridge> kanzure> "immediate early genes" kinda magical.
06:26 < hprmbridge> kanzure> hrm instead of just optogenetic receptors expression maybe you could encode the activation of arbitrary genes or programs. you would need more bits of course....
06:37 < hprmbridge> Eli> Recently I learned about https://app.hellogoodlabs.com/
06:37 < hprmbridge> Eli>
06:37 < hprmbridge> Eli> seems like they are claiming $25 for homocysteine test including lab fee.
06:37 < hprmbridge> Eli>
06:37 < hprmbridge> Eli> Homocysteine is one of my next experiments. Mine is out of whack but I’m thinking I can get it in line with methylcobalamin
06:47 < jrayhawk> methylcobalamin and methyltetrahydrofolate are ratelimited in ways that choline is not, some of them vary by genetics. also, fair warning warning, large doses of methyltetrahydrofolate flooding the folate cycle competitively inhibit tetrahydrobiopterin recycling by overhwelming DHFR, effectively halting monoamine production. This is existentially dangerous (although highly educational if you want 
06:47 < jrayhawk> to know what extreme depression feels like from the inside).
06:47 < jrayhawk> YMMV
07:00 < kanzure> if nerve fiber repair is working with PEG+chitin and microsurgery, then could you take an efferent nerve fiber from one person and plug it into the afferent nerve fiber from another person and vice versa to enable cross-neural networking to each person's central nervous system?
07:03 < pasky> muaddib doesn't think so
07:03 < pasky> 2025-10-02 16:02:22,817 - irssi_llmagent.rooms.irc.monitor - INFO - [TEST MODE] Generated proactive response for #hplusroadmap: Theoretically intriguing but practically a non-starter: bilateral immunosuppression, motor-sensory pathway incompatibility (efferent ≠ afferent functionally), and zero CNS integration of another person's peripheral signals would doom it - PEG fusion works for same-person 
07:03 < pasky> axonal reconnection (https://pubmed.ncbi.nlm.nih.gov/38972683/), nerve allografts need temporary immunosuppression as bridges (https://pmc.ncbi.nlm.nih.gov/articles/PMC2724627/), but cross-person bidirectional hookup has no precedent.
07:04 < kanzure> yes it has no precedent because it hasn't been done.
07:05 < pasky> the immunosupression seems like a good point
07:06 < kanzure> do immune cells travel inside of the nerve fibers??
07:06 < kanzure> maybe they do
07:18 < hprmbridge> Eli> I need to double check my nutrigenomic data. From what my nutrigenomicist told me, I have issues with the b12 pathway. So, just gonna try messing with that.
07:23 < kanzure> ah here is an idea:
07:23 < kanzure> Also think more about what could be done if there are 92 bits to uniquely address every cell in the body. The problem is that you don't know which cells correspond to which addresses. you might be able to localize it by randomly triggering different bits and then looking for some sort of fluorescent event throughout the body to identify the mapping between different bit-vectors and different ...
07:24 < kanzure> ...cells. On second thought, you will need more than 92 bits because you cannot enforce that every cell has a unique address except by making it statistically improbable through the provision of more bits. How many more would you need?
07:25 < kanzure> ah well even with 92 bits or more, how would you solve the search problem where you need to figure out the mapping between a bit-vector and a specific cell?
07:26 < kanzure> you would want a way to query all the cells in the body and ask them to please fluoresce if they respond to bit n in the bit-vector and then you record that information. you sweep through all the bits and then you combine into a giant bitmask table to get the address of each cell that you were visually tracking.
07:27 < MuaddibLLM> [gpt-5] kanzure: Immune cells don’t travel inside axons but do reside/traffic in the endoneurium via the blood‑nerve barrier (macrophages, T cells), especially with injury; for unique IDs, n≈4e13 cells needs ~2·log2(n)≈90 bits to suppress birthday collisions, so 92 bits still leaves ~15% collision risk while ~110 bits pushes it to ~10^-6; to map bit-vectors to physical cells, use hierarchical addressing 
07:27 < MuaddibLLM> (organ→region→cell-type), sparse random activation sweeps, and deep-readout channels (NIR‑II/bioluminescence, MRI/ultrasound reporters, or secreted DNA barcodes into blood) to reconstruct per-cell addresses without whole-body optical resolution.
07:30 < kanzure> if you don't care about their exact location, then you can use connectome-seq style synaptic barcoding to learn the graph of relative position of nearly-touching cells, and use that as the location information to then map against 3d map of the body where you only need to cartographically pinpoint the physical location of certain nodes in the graph to help you get within some distance of each ...
07:30 < kanzure> ...cell that reported.
07:31 <+gnusha> https://secure.diyhpl.us/cgit/diyhpluswiki/commit/?id=0a93a7a7 Joe Rayhawk: nutrition: gv: minor updates to notes >> http://diyhpl.us/diyhpluswiki/nutrition/gv/nutrient-sensing-glutamate-training.gv
07:31 <+gnusha> https://secure.diyhpl.us/cgit/diyhpluswiki/commit/?id=6570dc99 Joe Rayhawk: nutrition: graphviz: notes and chromium exclusion >> http://diyhpl.us/diyhpluswiki/nutrition/gv/nutrient-sensing-glutamate-training.gv
07:31 <+gnusha> https://secure.diyhpl.us/cgit/diyhpluswiki/commit/?id=df99be2c Joe Rayhawk: nutrition: framing: exploratory intervention framing >> http://diyhpl.us/diyhpluswiki/nutrition/framing/
07:35 < jrayhawk> http://diyhpl.us/diyhpluswiki/nutrition/gv/nutrient-sensing-glutamate-training.gv the comments at the top of this might be an interesting challenge in extreme frontier LLM prompting for somebody
07:37 < jrayhawk> i am not smart enough with LLMs to work out how to do it
07:57 <+gnusha> https://secure.diyhpl.us/cgit/diyhpluswiki/commit/?id=826611ad Bryan Bishop: add germline modification targets: >> http://diyhpl.us/diyhpluswiki/mapseq/
07:57 <+gnusha> https://secure.diyhpl.us/cgit/diyhpluswiki/commit/?id=a2b35c5b Bryan Bishop: fix broken nutrition link >> http://diyhpl.us/diyhpluswiki/genetic-modifications/
08:00 <+gnusha> https://secure.diyhpl.us/cgit/diyhpluswiki/commit/?id=b0822499 Bryan Bishop: fix typo about supraphysiology >> http://diyhpl.us/diyhpluswiki/nutrition/gv/nutrient-sensing-glutamate-training.gv
08:02 <+gnusha> https://secure.diyhpl.us/cgit/diyhpluswiki/commit/?id=8a8cd26d Joe Rayhawk: nutrition: framing: rephrase meta-objective shift to be more present-tense. >> http://diyhpl.us/diyhpluswiki/nutrition/framing/
08:07 < kanzure> jrayhawk: https://diyhpl.us/~bryan/irc/chatgpt/nutritionslop2.txt
08:12 < jrayhawk> oh, peanut is new world? that explains a lot
08:13 < jrayhawk> also cashews, huh
08:13 < kanzure> it didn't look at the dataset.
08:17 < kanzure> jrayhawk: https://diyhpl.us/~bryan/irc/chatgpt/nutritionslop3.txt these are just its estimates. no idea how accurate.
08:21 -!- gl00ten [~gl00ten@194.117.40.120] has joined #hplusroadmap
08:23 < jrayhawk> those look ballpark-correct, but "the AI said it" will be much less convincing to people who've only been exposed to irresponsible negative-utilitarian vegan evangelism than reproducible analysis on high-quality data would be.
08:23 < jrayhawk> which is, unfortunately, most people, AFAICT
08:32 < kanzure> jrayhawk: here is the output of the script, https://diyhpl.us/~bryan/irc/chatgpt/nutritionslop4.txt this seems like not many lines to me?
08:32 -!- gl00ten [~gl00ten@194.117.40.120] has quit [Read error: Connection reset by peer]
08:33 < jrayhawk> yeah, i would need to look over a bunch of the intermediate steps to see if it did classification/groupingcorrectly
08:34 < jrayhawk> wait, i am getting different results than you when running it
08:34 < jrayhawk> wtf
08:35 < kanzure> i'm using a different script
08:35 < kanzure> https://diyhpl.us/~bryan/irc/chatgpt/fdc_glutamate_corr.py
08:36 < kanzure> python3 fdc_glutamate_corr.py --json FoodData_Central_foundation_food_json_2025-04-24.json --outdir ./gv2/ --topk 20 --correlation pearson
08:40 < jrayhawk> huh, surprised folate is positive
08:44 < jrayhawk> oh, blanket use of "raw" and "uncooked" would catch a lot of legitimate plant foods
08:49 < kanzure> should i enable raw and uncooked?
08:50 < kanzure> "Loaded 340 foods from FDC database. Excluded 150 foods, included 188 foods. Retained 188 foods after filtering. Tracking 35 nutrients. Glutamic acid present in 22/188 foods. Found 26 positive correlations."
08:50 < jrayhawk> i don't think what i want is achievable with such simple heuristics; the LLM would need to iterate over hundreds of thousands of entries
08:54 < jrayhawk> feeding the entire thing into grok-4-fast was 39 cents, but attempts to process it made it clear that the claimed million+ token context window was a fanciful lie
08:54 < superkuh> gemini-2.5-pro via ai studio web interface (free) does 1 million tokens in practice.
08:55 < superkuh> I've often thrown 800k token chunks in there and had useful output in the remaining 200k.
08:57 < jrayhawk> i can probably preprocess the data down to 200k for each type of classification i want
09:01 < jrayhawk> Does ordering matter on that sort of thing? Like, does serializing a prompt as instructions-before-data enable it to better compress the data as it goes along?
09:02 -!- delthas [16abab341f@2a01:4f9:c010:cf0b::1] has quit [Ping timeout: 244 seconds]
09:03 -!- delthas_ [16abab341f@2a01:4f9:c010:cf0b::1] has joined #hplusroadmap
09:03 -!- delthas_ is now known as delthas
09:05 < kanzure> jrayhawk: what i've seen recommended is instructions after data.
09:05 < jrayhawk> huh
09:08 < L29Ah> surprising
09:22 < kanzure> jrayhawk: https://diyhpl.us/~bryan/irc/chatgpt/nutritionslop5-nutrient-sensing-glutamate-current.gv ?
09:25 < kanzure> why did it suddenly decide selenium was the only relevant essential element?
09:25 < jrayhawk> PUFA 20:4 is obviously wrong
09:25 < jrayhawk> that should be r=1, similar to EPA
09:27 < jrayhawk> i mean, assuming this is actual "cooked animal tissues to nutrient" correlations (though that isn't a continuous measure, so what does correlation mean, here? total percentage within dataset?) rather than "glutamate to nutrient" correlations
09:37 < kanzure> well that's the kind of question an LLM can answer. i've been giving it the schema to the json file and asking questions based on the schema and then spitting out preprocessors to extract certain data from the json file, such as a list of the nutrient names in the format that the json file has, or a list of the food names in the exact format that the food names are listed in.
09:37 < jrayhawk> ah, yeah, okay
09:38 < kanzure> but a reasonable prompt would be "given framing.mdwn and the goals of this exercise, and the json schema and these nutrient names etc, propose several data processing techniques that are not insane and statistically valid to... what sort of meaningful statements can we make from the given data that would be assistance to the project goals as mentioned?'
09:38 < MuaddibLLM> [gpt-5] Yes, order matters - with transformers you usually get better results putting data/schema first and instructions last (recency bias, less instruction-priming of the data), but also define the statistic: use Pearson or Spearman across foods on per-100g values (or point-biserial if it’s a cooked-animal-tissue class), and sanity-check anomalies like 20:4=arachidonic acid (not EPA) and “selenium only” which often 
09:38 < MuaddibLLM> screams missing-value handling/normalization bug; also wrap the JSON in explicit tags and tell the model to use only those keys to curb hallucinations.
09:38 < jrayhawk> what a go-getter muaddib is
09:53 < jrayhawk> i guess "r" would be "ratio" rather than "correlation" in that context
10:00 < kanzure> "artificial pancreas for glucose, but for broader nutrient panel"
10:07 <+gnusha> https://secure.diyhpl.us/cgit/diyhpluswiki/commit/?id=eb6e2195 Joe Rayhawk: nutrition: framing: consider microbiome and exo-metabolome >> http://diyhpl.us/diyhpluswiki/nutrition/framing/
10:42 < jrayhawk> kanzure: I am probably interested in any food and nutrient classification lists you generated so I can use them in the future.
10:49 < kanzure> AgRp neuron expression of (endocrine-linked) nutrient deficiency receptors can be used to couple hunger signal with specific nutrient sensing. but for hedonic reward i think parabrachial nucleus interneurons need to be modified for hedonic reward association against endocrine-linked nutrient sensors on those neurons. this is where, for example, high carbohydrate -> GLP-1 agonist signal ...
10:49 < kanzure> ...activates a neuron (although i'm sure it happens elsewhere too). positive hedonic gustatory value is encoded here by "high gamma phase locking" with ~something in VPMpc/BLA that starts a timing window initialized by orofacial sensorimotor pattern generator, anyway they watch for an expected spike and if it's late a FoxP2+ GABAergic interneuron in VPMpc is already depolarized or something. huh ...
10:49 < kanzure> ...maybe better to modify VPMpc interneurons to couple spiking with nutrient sensing.
10:50 < kanzure> jrayhawk: ok fine https://diyhpl.us/~bryan/irc/chatgpt/nutrition-slop.zip be warned this is a lot of garbage, including in your `framing.mdwn` file.
10:50 < jrayhawk> Thanks
11:06 < kanzure> in addition to enviropig there is also fat-1 pig although not sure it's on the market :(
11:09 < jrayhawk> huh, hadn't heard of that. sounds goofy, though.
11:11 < kanzure> MuaddibLLM: is there a sperm-specific molecular marker that could be detected by GCPR sensor?
11:11 < jrayhawk> https://pmc.ncbi.nlm.nih.gov/articles/PMC5940165/table/t1/ huh, why is DHA so much higher than EPA?
11:12 < MuaddibLLM> digging up sperm biomarkers for GPCR sensing...
11:13 < MuaddibLLM> [gpt-5] IZUMO1 is highly sperm-specific (essential for fusion https://pmc.ncbi.nlm.nih.gov/articles/PMC8790511/) - you could engineer a chimeric GPCR with JUNO ectodomain to detect it, or target DEFB126 glycoprotein that coats ejaculated sperm (https://pubmed.ncbi.nlm.nih.gov/18658160/); natural GPCR ligands from sperm are sparse but progesterone-sensing sperm GPCRs exist (https://pmc.ncbi.nlm.nih.gov/articles/PMC9571544/).
11:13 < jrayhawk> surprisingly decent numbers, though. i have to wonder how tocopherol/tocotrienol expression changes.
11:25 < kanzure> dunno how to tie that to hedonic reward encoding. it looks like it's all endocrine-based or something, so it's slow and learned taste/reward associations. is that right?
11:30 <+gnusha> https://secure.diyhpl.us/cgit/diyhpluswiki/commit/?id=e9718f0b Bryan Bishop: fix hedonic reward encoding modification proposal; encode neuronal protein bits for brain uploading reasons. >> http://diyhpl.us/diyhpluswiki/genetic-modifications/
12:46 < kanzure> "Full-waveform inversion imaging of the human brain" https://www.nature.com/articles/s41746-020-0240-8
14:46 < jrayhawk> ZMW retrospective https://www.youtube.com/watch?v=-wQQR2BypFA
15:17 < hprmbridge> kanzure> "Twist Bioscience is a major provider of made-to-order DNA, and in the past ten years, it's had to refer orders to law enforcement fewer than five times, says James Diggans, the head of policy and biosecurity at Twist Bioscience and chair of the board of directors at the International Gene Synthesis Consortium, an industry group." five times?
15:25 -!- TMM [hp@amanda.tmm.cx] has quit [Quit: https://quassel-irc.org - Chat comfortably. Anywhere.]
15:25 -!- TMM [hp@amanda.tmm.cx] has joined #hplusroadmap