--- Log opened Thu Aug 14 00:00:10 2025
01:10 -!- srat3 [~srat3@user/srat3] has quit [Ping timeout: 272 seconds]
01:10 -!- srat3 [~srat3@user/srat3] has joined #hplusroadmap
03:30 -!- juri_ [~juri@implicitcad.org] has quit [Ping timeout: 252 seconds]
03:42 -!- juri_ [~juri@implicitcad.org] has joined #hplusroadmap
04:29 -!- darsie [~darsie@84-113-82-174.cable.dynamic.surfer.at] has quit [Quit: Avoid fossil fuels and animal products. Have no/fewer children. Protest, elect sane politicians. Invest ecologically.]
04:43 < hprmbridge> kanzure> https://p2pfoundation.ning.com/ is no more, a piece of bitcoin history gone
04:50 < hprmbridge> kanzure> mbauwens, who operates p2pfoundation, is promising a memoir of his Satoshi correspondence. This is good because he alleges to be possibly terminally ill, which might also explain why he's quote tweeting a scam coin.
04:54 < hprmbridge> kanzure> someone else is claiming >200,000 tokens/second. this must be false; maybe they mean parallel tokens, not linear tokens/second?
05:23 < fenn> large batch size vastly increases throughput. yes, it must be parallel, but that still seems high, especially with such large context
05:24 < fenn> for batch size = 1 you have to transfer the entire model (many gigabytes) from VRAM to GPU for each token
05:25 < L29Ah> from GPU to GPU you mean?
05:25 < fenn> for batch size = 1000 you have to transfer the entire model from VRAM to GPU for 1000 tokens, and so on
05:25 < fenn> eventually the time to actually do the multiply accumulate operation overtakes the time to transfer, and you start getting diminishing returns
05:26 < L29Ah> or is it about the PU-local SRAM?
05:26 < hprmbridge> kanzure> maybe groq and cerebras keep the weights on chip between each batch
05:27 < fenn> but you run out of VRAM for KV cache long before that, on large models like this. you can stretch it to the max with only one layer per GPU, and even splitting across multiple GPUs, but then inter-GPU bandwidth becomes a constraint
05:27 < fenn> groq does, cerebras doesn't
05:31 < fenn> hmm actually i think you're limited by bandwidth between different levels of cache, not "arithmetic bandwidth" but the result is the same
05:38 < hprmbridge> kanzure>  https://cdn.discordapp.com/attachments/1064664282450628710/1405530904352260126/image.jpg?ex=689f2a29&is=689dd8a9&hm=aeb9e367ffa9cef8bb05801a4d265758381b6c73bf84464f0a6930240bdfffa4&
06:02 < hprmbridge> kanzure> https://docs.voyageai.com/docs/contextualized-chunk-embeddings
06:31 < hprmbridge> kanzure> have you a reinforcement learning blogpost https://tokenbender.com/post.html?id=avatarl
06:40 < hprmbridge> kanzure> "gpt-oss-120b, RTX Pro 6000 Blackwell, LM Studio on Windows 11, 169 tokens/sec, 0.09s to first token"
06:46 < hprmbridge> kanzure> baez 2024 "What is entropy?" https://arxiv.org/abs/2409.09232
06:52 -!- gl00ten [~gl00ten@2001:8a0:7ee5:7800:46d9:f5c:17a2:432] has joined #hplusroadmap
07:06 -!- gl00ten [~gl00ten@2001:8a0:7ee5:7800:46d9:f5c:17a2:432] has quit [Ping timeout: 272 seconds]
07:20 -!- gl00ten [~gl00ten@bl5-239-125.dsl.telepac.pt] has joined #hplusroadmap
07:22 -!- gl00ten [~gl00ten@bl5-239-125.dsl.telepac.pt] has quit [Remote host closed the connection]
07:26 -!- stipa_ [~stipa@user/stipa] has joined #hplusroadmap
07:29 -!- stipa [~stipa@user/stipa] has quit [Ping timeout: 276 seconds]
07:29 -!- stipa_ is now known as stipa
07:37 -!- gl00ten [~gl00ten@bl5-239-125.dsl.telepac.pt] has joined #hplusroadmap
09:16 -!- TMM [hp@amanda.tmm.cx] has quit [Quit: https://quassel-irc.org - Chat comfortably. Anywhere.]
09:16 -!- TMM [hp@amanda.tmm.cx] has joined #hplusroadmap
09:41 < kanzure> https://github.com/jj-vcs/jj "Jujutsu internally it abstracts the user interface and version control algorithms from the storage systems used to serve your versioned content. This allows it to serve as a VCS with many possible physical backends. [...] Jujutsu keeps track of conflicts as first-class objects in its model; they are first-class in the same way commits are, while alternatives like ...
09:41 < kanzure> ...Git simply think of conflicts as textual diffs. While not as rigorous as systems like Darcs (which is based on a formalized theory of patches, as opposed to snapshots), the effect is that many forms of conflict resolution can be performed and propagated automatically."
09:41 < kanzure> "designed to be safe under concurrent scenarios; simply using rsync or Dropbox and then using that resulting repository should never result in a repository in a corrupt state"
09:42 < kanzure> https://radicle.xyz/2025/08/14/jujutsu-with-radicle
09:46 < kanzure> wonder if jujutsu could offer a reasonable alternative for the bitcoin hack of adding a sha512 hash into the commit message. alternatively, a separate branch of signed commits with metadata in the commit messages, so that normal user experience doesn't have to see that.
09:47 < kanzure> or does git have a plan for incorporating other cryptographic hash functions? several years ago there was no plan IIRC.
10:09 < L29Ah> oh they replaced the joystick with something hopefully nicer in Twiddler 4
10:10 < L29Ah> please tell me how good it is :3
10:29 < RangerMauve> I'm _thiiis_ close to getting the Twiddler 5. The optical mouse and USB-C charging is really sealing the deal. Just gotta budget vs other stuff
10:44 -!- gl00ten [~gl00ten@bl5-239-125.dsl.telepac.pt] has quit [Ping timeout: 245 seconds]
10:51 < jrayhawk> You just need those and some HUD glasses and you'll be living the 90's cyberpunk dream
10:59 < L29Ah> XReal Air so far seems like the best HMD deal
10:59 < L29Ah> RangerMauve: have you ever tried such a thumb position input?
10:59 < L29Ah> i've never seen it anywhere
11:00 < L29Ah> is it just a mouse sensor turned around?
11:07 < RangerMauve> jrayhawk I already am! :P I use head mounted displays for coding on the reg. Viture XR pro mainly
11:08 < RangerMauve> L29Ah From what I can see it's equivalent to the pointer on my GPD Win 4. I don't use it as much as the joystick since it can be finnicky but I think I can get used to it on the twiddler
11:09 < L29Ah> cool
11:09 < L29Ah> how does it compare to other thumb input methods like TrackPoint and PSP Thumbstick?
11:12 < RangerMauve> I like trackpoint the most tbh but I haven't had a device in years. IDK about psp thumbstick but controller joysticks in general are so-so. Depends on the software that captures velocity. SteamOS is decent with most thumbsticks, the GPD thumbsticks only do 45 degree angles which sucks ass.
11:12 < RangerMauve> I'd say the steam deck touchpad thingies have been my favorite mouse analogs but the steam deck sucks for typing.
11:14 < RangerMauve> The the optical pointer can be hard to use for small precise movements in my experience. Changing thumb orientation to use just the tip helps
11:16 < L29Ah> maybe some sculpted around a thumb pointy thimble could help
11:16 < L29Ah> or not pointy, hmm
11:17 < L29Ah> perhaps something that attaches rigidly to the thumbnail?
11:17 < RangerMauve> IMO a TrackPoint that you can touch with your nail would be ideal
11:20 < RangerMauve> Me at a recent hackathon: https://media.discordapp.net/attachments/1363157429885337620/1391095853745307658/IMG_1563.jpg?ex=689f6278&is=689e10f8&hm=58319b39fb166581fbffb31cf9dec6dcb7a576b6b5f9b2d48fcbce3958358c04&=&format=webp&width=1102&height=826
11:21 < RangerMauve> The joystick mouse on the GPD might not be ideal but TBH it's good enough. I can't draw with it but I just whip out a drawing tablet if I really need it. It's good enough to press buttons when a GUI is a pain to use with just keyboard events
11:23 < jrayhawk> https://www.instagram.com/reel/DJ-M-xPh4nX/ it's a very versatile I/O paradigm
11:26 < RangerMauve> Fuck maybe I should get two twiddlers and map them to different output modalities
11:26 < RangerMauve> No, I should have a free hand to do stuff with
11:27 < jrayhawk> that's what robots are for
11:27 < L29Ah> https://www.mytwiddler.com/assets/tgimages/sushi-dragon2.png
11:30 < RangerMauve> jrayhawk IDK if you saw me ramble about this before but I'm aiming to use EMGs to map muscle movements to control robot arms eventually. Currently focused on making neural nets to detect events from EMGs, then I'll control a mouse pointer, then 3D point control, then some cheap robot arms
11:31 < jrayhawk> oh wow
11:33 < RangerMauve> Eventually I want to be full time in mixed reality with my digital arms + physical arms. IDK if it's feasible or how long it'll take to get there, but I'm slowly making progress
11:35 < jrayhawk> are there existing productized EMG arrays you like for that, or are you stuck prototyping the hardware yourself?
11:35 < kanzure> vim keybindings for simonw's llm tool:
11:35 < kanzure> nnoremap gl :%!sh -c 'tmp=$(mktemp); cat > "$tmp"; cat "$tmp"; printf "\n--- Command output ---\n"; cat "$tmp" \| llm; rm "$tmp"'<CR>
11:35 < jrayhawk> or is there a blog or a github page i should be looking at
11:35 < kanzure> vnoremap gl !sh -c 'tmp=$(mktemp); cat > "$tmp"; cat "$tmp"; printf "\n--- Command output ---\n"; cat "$tmp" \| llm; rm "$tmp"'<CR>
11:49 < kanzure> https://www.mbi-deepdives.com/golden-age-of-digital-ads-llm-p-l/ "6.7% increase in CTR for the advertisers using AdLlama compared to the ones using the old imitation model" that's a huge CTR improvement. what the hell?
11:50 < RangerMauve> @jr
11:50 < kanzure> maybe they are comparing to a particularly bad prior model. you're telling me they were earning hundreds of billions of dollars per year from ads and they had no other tricks for roughly similar CTR boost?
11:51 < RangerMauve> jrayhawk Nothing published yet but I'm working with these cheap EMGs for now: https://www.aliexpress.com/item/1005005272063208.html
11:51 < kanzure> RangerMauve: have you calculated a hypothetical bits/second rate that you can extract using this modality?
11:53 < RangerMauve> kanzure I'm thinking of focusing more on continuous values actually. A gesture to enable tracking along an axis and muscle flexing / unflexing gradually to control a value between 0/1. I don't have data yet to know if I can actually generate that value
11:54 < RangerMauve> bits/second would be very low sadly. I wouldn't use this to type for example
11:55 < kanzure> something something cam encoder.. something something.. even continuous motions have a bit rate if you can't control your muscles below some resolution during continuous movement... or something. sorry for the unformulated slop thinking.
11:55 < kanzure> ah okay.
11:55 < kanzure> well, wake me up when thre's a high bandwidth bits/second method of extracting data from human bodies
11:57 < RangerMauve> Yeah I'll defs send an update on how fine grained I can get the resolution to be. My "pulling numbers out my ass" guess is no more than a couple bits per second
12:03 < RangerMauve> Just ordered the twiddler 4
12:10 < kanzure> monkey king says he can control every single follicle of hair to transmit data into any computer via piloerection of arrector pili muscle contraction.
12:11 < kanzure> he tells me there are several methods for this:
12:11 < L29Ah> sus
12:12 < hprmbridge> setecastronomy1891> Sun Wukong said this?
12:12 < kanzure> micro-lever cantilever tactile‐surface sensors can treat hair as a spring arm and it can trigger a few nano-newtons of force
12:13 < kanzure> depending on what brand of hair spray you use, the hair follicle acts as a moving electrode
12:13 < kanzure> other hair coats can increase conductance but perhaps more interesting is reflective coating for visual monitoring by camera
12:14 < kanzure> probably something something laser scanning over the reflective surfaces, optical fiber sensor, uhh you wouldn't be able to move around much
12:15 < L29Ah> that's great but i find the capability to trigger individual hair-muscles at will dubious
12:16 < kanzure> monkey king has practiced for 1,000 years
12:16 < kanzure> he also innovated a method of cloning himself from a single hair follicle, although often he used much more than one.
12:17 < RangerMauve> That's a fun thought but also sounds like something a toddler would say on a playground
12:17 < RangerMauve> Who is monkey king?
12:19 < kanzure> sun wukong monkey king was a buddhist monk named xuanzang
12:21 < RangerMauve> I'd trust him with viability of various computer I/O modalities 🫡
13:06 -!- Netsplit *.net <-> *.split quits: delthas, Gooberpatrol_66, Malvolio, Jenda, nsh, drmeister, srk, yuanti, TMM, _flooded,  (+43 more, use /NETSPLIT to show all of them)
13:09 -!- Netsplit over, joins: @ChanServ, srat3, WizJin, andytoshi, jrayhawk, s0ph1a, nsh, FelixWeis__, stipa, TMM (+35 more)
13:10 -!- Netsplit over, joins: streety, flyback
13:10 -!- Netsplit over, joins: kanzure, dartmouthed, balrog, alethkit, soundandfury, archels
13:10 -!- flyback [~flyback@2601:540:c700:2380:b783:3959:4032:9fee] has quit [Max SendQ exceeded]
13:10 -!- flyback [~flyback@2601:540:c700:2380:b7af:5add:a79b:d03e] has joined #hplusroadmap
13:35 -!- Gooberpatrol_66 [~Gooberpat@user/gooberpatrol66] has quit [Quit: Konversation terminated!]
15:37 < kanzure> kimi k2 is a great model https://diyhpl.us/~bryan/irc/chatgpt/hair-display.txt
15:38 -!- Gooberpatrol66 [~Gooberpat@user/gooberpatrol66] has joined #hplusroadmap
15:47 -!- Gooberpatrol66 [~Gooberpat@user/gooberpatrol66] has quit [Quit: Konversation terminated!]
15:47 -!- Gooberpatrol66 [~Gooberpat@user/gooberpatrol66] has joined #hplusroadmap
15:51 < kanzure> kimi says "the minimum spatially addressable voxel of vasoconstriction is likely on the order of 30–100 μm per side" based off of arteriole neuroanatomy
15:51 < kanzure> "aggregate channel capacities in the range of kbps to maybe 50 Kbps across a whole body,"
15:57 < kanzure> "roughly a few million independently tunable vascular segments throughout the human body"
16:04 < kanzure> however, claude disagrees about 50 Kbps and thinks it should be up to 50 megabits throughout the whole human body.. who knows. it would be cool if the sympathetic system does parallel simultaneous control of all vasoconstriction voxels (and other terminal actuators), but for all i know it has high switching costs to stimulate/signal different actuators.
16:13 -!- Gooberpatrol66 [~Gooberpat@user/gooberpatrol66] has quit [Quit: Konversation terminated!]
16:23 < kanzure> weird vasculature-based vasodilation/vasoconstriction imaging technique https://gnusha.org/logs/2017-01-02.log "Molecular imaging with engineered physiology" http://www.nature.com/articles/ncomms13607
17:14 -!- Gooberpatrol66 [~Gooberpat@user/gooberpatrol66] has joined #hplusroadmap
18:01 -!- L29Ah [~L29Ah@wikipedia/L29Ah] has quit [Ping timeout: 260 seconds]
18:03 -!- L29Ah [~L29Ah@wikipedia/L29Ah] has joined #hplusroadmap
18:15 -!- Gooberpatrol66 [~Gooberpat@user/gooberpatrol66] has quit [Ping timeout: 244 seconds]
18:21 < hprmbridge> kanzure> "Targeted deletions of large syntenic regions in Arabidopsis thaliana" https://www.pnas.org/doi/10.1073/pnas.2419744122
18:35 -!- TMM [hp@amanda.tmm.cx] has quit [Quit: https://quassel-irc.org - Chat comfortably. Anywhere.]
18:35 -!- TMM [hp@amanda.tmm.cx] has joined #hplusroadmap
18:37 < hprmbridge> kanzure> syn61 seems to be owned by https://www.constructive.bio/
18:40 < hprmbridge> kanzure> "capability to produce strains with resistance against phage infections and intolerance" yep that's industrially useful.
18:52 < hprmbridge> kanzure> "Systemic factors in young human serum influence in vitro responses of human skin and bone marrow-derived blood cells in a microphysiological co-culture system" https://www.aging-us.com/article/206288/text
19:06 < hprmbridge> kanzure> "Photophoretic flight of perforated structures in near-space conditions" https://www.nature.com/articles/s41586-025-09281-8
19:31 < hprmbridge> kanzure> "Vascular response to embryo implantation in anterior chamber of eye following interspecies embryo transfers between the rat, mouse, and guinea-pig" https://link.springer.com/article/10.1007/BF00239463
19:54 -!- Gooberpatrol66 [~Gooberpat@user/gooberpatrol66] has joined #hplusroadmap
20:28 < hprmbridge> kanzure> "Application of the quail-chick chimera system to the study of brain development and behavior" https://www.science.org/doi/10.1126/science.3413496
20:29 < hprmbridge> kanzure> "Hatched chicks with chimeric brains containing cells from both the domestic chicken (Gallus gallus domesticus) and the Japanese quail (Coturnix coturnix japonica) have been produced by transplantation of various regions of the neural tube at the 8- to 15- somite stage. The positions of host and donor cells relative to graft boundaries observed throughout embryonic development and after hatching
20:29 < hprmbridge> kanzure> implicated both radial and tangential cell movements in brain morphogenesis. In addition, transplants containing the entire quail mesencephalon and diencephalon resulted in the transfer of certain aspects of species-typical crowing behavior."
22:11 -!- superkuh [~superkuh@user/superkuh] has quit [Ping timeout: 245 seconds]
22:26 -!- superkuh [~superkuh@user/superkuh] has joined #hplusroadmap
--- Log closed Fri Aug 15 00:00:11 2025