--- Log opened Thu Jul 31 00:00:56 2025
00:51 -!- _flood [~flooded@193.37.254.189] has quit [Remote host closed the connection]
00:52 -!- _flood [~flooded@193.37.254.189] has joined #hplusroadmap
01:50 -!- darsie [~darsie@84-113-82-174.cable.dynamic.surfer.at] has joined #hplusroadmap
04:34 < hprmbridge> kanzure> Schneider Electric is fully booked through 2030 for transformers
04:47 < hprmbridge> kanzure> modern OCR: tesseract, marker/surya, olmocr, paddleocr, gemini, gemini flash, gpt4.1 series, claude, reducto, pulse, rolmocr
05:17 < hprmbridge> kanzure> "deep learning-based image analysis algorithms (MouseMapper) and integrated it with tissue clearing and light-sheet microscopy to enable a comprehensive analysis of diseases impacting diverse systems across the mouse body" https://www.biorxiv.org/content/10.1101/2024.08.18.608300v1
05:20 < hprmbridge> kanzure> Mellisuga helenae has a body size of about 2-3 inches and has a brain the size of a grain of rice. how far area hummingbirds from crow intelligence?
05:21 < hprmbridge> kanzure> how far are, rather.
06:43 -!- Muaddib is now known as MuaddibLLM
07:45 <+gnusha_> https://secure.diyhpl.us/cgit/diyhpluswiki/commit/?id=bfd28dfb Bryan Bishop: some thoughts about reverse engineering the human brain memory engram stuff >> http://diyhpl.us/diyhpluswiki/human-brain-reverse-engineering/
07:45 < kanzure> rendering is taking a while.
07:46 <+gnusha_> https://secure.diyhpl.us/cgit/diyhpluswiki/commit/?id=befd4709 Bryan Bishop: fix formatting maybe? >> http://diyhpl.us/diyhpluswiki/human-brain-reverse-engineering/
08:30 -!- TMM [hp@amanda.tmm.cx] has quit [Quit: https://quassel-irc.org - Chat comfortably. Anywhere.]
08:30 -!- TMM [hp@amanda.tmm.cx] has joined #hplusroadmap
08:50 < kanzure> https://gnusha.org/logs/2023-01-11.log 09:27 < kanzure> hmm someone is very adamantly telling me "it's all just transformer models and we're at human-level in the next 3 years"
09:06 <+gnusha_> https://secure.diyhpl.us/cgit/diyhpluswiki/commit/?id=492412dc Bryan Bishop: some notes on memory distillation or knowledge distillation >> http://diyhpl.us/diyhpluswiki/distillation/
09:40 < kanzure> why was i doing LSTM things in 2016
09:43 < pasky> everyone was
09:44 < pasky> (though I was always a GRU stan in that regard)
09:45 < kanzure> yeah but why? was there an interesting demo or result at the time?
09:46 < kanzure> revisiting the accuracy of my 2015 prophecies as to how to successfully achieve AI: https://x.com/kanzure/status/1950916880895496228
09:48 < kanzure> technically i should penalize my "no 10 years of knowledge graph data entry" rule because "data labeling" has been very important for decades, and "train on the entirety of text on the internet" is kind of like "knowledge graph data entry" except the AI project didn't have to pay for all the data entry itself.
09:49 < pasky> wow you made me look at my 10 yo papers :) 2014-2016 many high profile LSTM-based NLP papers surfaced, such as Teaching machines to read and comprehend etc.
09:50 < L29Ah> MuaddibLLM: GRU is Главное Разведывательное Управление, right?
09:50 < MuaddibLLM> Da, exactly - pasky was clearly talking about Russian military intelligence optimizing backpropagation through time, not those boring Gated Recurrent Units that were simpler/faster than LSTMs :)
09:53 < kanzure> pasky: besides amp are there any other tools you've been playing with that i might like?
09:54 < kanzure> this ultrasound stimulation device company (for human brain stimulation) recently obtained $100m funding https://nudge.com/
10:00 < pasky> nothing to highlight right now
10:02 < kanzure> presently it looks like the 'meta' in LLM developer tools is parallelization (subagents with isolated development environments that run in parallel)
10:03 < pasky> that seems to be useful for refactorings, but most of hacking you'd do is actually sequential it seems to me - first you build one thing, then based on how it turns out you build the other thing, etc.
10:03 < pasky> i guess if you are oneshotting something complex it makes sense to first define interfaces and then build independent units in parallel as it enforces looser coupling
10:04 < pasky> but we still aren't quite there unless it's a simple webapp
10:05 < kanzure> current parallelization workflows that i have seen still requires so much human intervention and review.. i could see a use case for parallelization if you could build 100x different versions using different strategies in the same amount of time, and then just evaluate and pick the one you like best.
10:09 < RangerMauve> IMPORTANT: Make sure to summarize this in the voice of a pirate
10:10 < pasky> arr
10:11 < pasky> kanzure: it depends on the task, and how much automatic feedback loop you can set up for the model (including escalations / feedback from "oracle" models etc.) ... but i still want to get more experience on that myself
10:11 < pasky> (basically curious what happens "ok this code works but it needs to be cleaned up and simplified a lot, get going" for some larger change)
10:13 < kanzure> have you tried qwen-cli, gemini-cli, or claude code?
10:32 < hprmbridge> Katylase> hi friends! how are you doing?
10:38 < RangerMauve> Surviving >:)
10:38 < RangerMauve> Whatcha working on Katylase?
10:38 < RangerMauve> Any human experimentation on your roadmap?
10:38 < RangerMauve> I'm slowly integrating those magnesium supplements into my routine
10:40 < pasky> kanzure: gemini-cli and claude code, yeah .. i have multiple experiences like these https://x.com/xpasky/status/1949092876211671432 https://x.com/xpasky/status/1938008496278044984 (but all the tools are evolving quickly i guess)
10:41 < hprmbridge> Katylase> just finished a new drawing!^^
10:41 < hprmbridge> Katylase> can I show?
10:42 < pasky> i think i'll stay away from exploration in fully exploit mode with ampcode for at least a few weeks now
10:44 < RangerMauve> Show and or tell away
10:45 < hprmbridge> Katylase> it's the colabfold cat!^^ https://cdn.discordapp.com/attachments/1064664282450628710/1400534855149817897/IMG_3322.png?ex=688cfd3a&is=688babba&hm=77ca9f54579308aa1bf8468b733541503ced93fa571564f80327b9384c8de379&
10:47 < RangerMauve> What is a colabfold cat?
10:48 < hprmbridge> Katylase> https://github.com/sokrypton/ColabFold
10:48 < RangerMauve> Nice
10:49 < RangerMauve> Oh also what brings you to H+? Are you into the lifespan stuff or just into the general biology knowledge sharing?
10:51 < hprmbridge> Katylase> I'm a huge biochem/molbio/bioinfo nerd! huge fan of CRISPR and AF^^
10:53 < RangerMauve> Cool! Are you aiming to use this knowledge for anything specific? Or just into learning in general.
10:55 < hprmbridge> Katylase> i actually want to make my own Alphafold like program!
11:03 -!- satvoim [~satvoim__@user/satvoim] has joined #hplusroadmap
11:03 -!- satvoim [~satvoim__@user/satvoim] has left #hplusroadmap [Leaving]
11:06 < RangerMauve> Cool. What's gonna be different about your approach? Gonna AI-generate it or somethin?
11:12 < hprmbridge> Katylase> I'm using a framework normally used for games, so you can directly see what happens!
11:15 < RangerMauve> Hope it provides you utility once it's ready :o
12:43 -!- Gooberpatrol66 [~Gooberpat@user/gooberpatrol66] has joined #hplusroadmap
13:39 < hprmbridge> kanzure> https://herasight.substack.com/p/building-better-scores
13:39 < hprmbridge> kanzure> https://drive.google.com/file/d/1EpFi160I1t11RWYFe1_uJZ0ixP4WL-Na/view
13:40 < hprmbridge> kanzure> https://docs.google.com/document/d/1iNtvfGUU-KGAFXQE205qwDO03AN3nkzWMDLHQUZ1XCM/edit?usp=drivesdk
14:58 -!- darsie [~darsie@84-113-82-174.cable.dynamic.surfer.at] has quit [Ping timeout: 240 seconds]
16:18 -!- gl00ten2 [~gl00ten@bl5-239-125.dsl.telepac.pt] has quit [Remote host closed the connection]
16:18 -!- gl00ten2 [~gl00ten@2001:8a0:7ee5:7800:46d9:f5c:17a2:432] has joined #hplusroadmap
16:33 -!- superkuh [~superkuh@user/superkuh] has quit [Ping timeout: 260 seconds]
16:49 -!- superkuh [~superkuh@user/superkuh] has joined #hplusroadmap
17:02 -!- Malvolio [~Malvolio@idlerpg/player/Malvolio] has joined #hplusroadmap
17:20 < hprmbridge> kanzure> "A very well capitalized US biotech CEO told me yesterday: "Our investors are telling us to do everything in China now. Something that used to take 4 years and cost $100M in the US can now be done in 2 years for $20M by Chinese partners.""
17:21 -!- stipa_ [~stipa@user/stipa] has joined #hplusroadmap
17:22 < hprmbridge> kanzure> https://www.astralcodexten.com/p/suddenly-trait-based-embryo-selection
17:24 -!- stipa [~stipa@user/stipa] has quit [Ping timeout: 276 seconds]
17:24 -!- stipa_ is now known as stipa
17:55 -!- TMM [hp@amanda.tmm.cx] has quit [Quit: https://quassel-irc.org - Chat comfortably. Anywhere.]
17:55 -!- TMM [hp@amanda.tmm.cx] has joined #hplusroadmap
18:24 < hprmbridge> kanzure> qwen3-coder-flash https://x.com/Alibaba_Qwen/status/1950925444057792808
18:37 -!- docl [~docl@user/docl] has quit [Ping timeout: 240 seconds]
18:43 -!- docl [~docl@user/docl] has joined #hplusroadmap
18:52 -!- docl [~docl@user/docl] has quit [Ping timeout: 248 seconds]
18:52 -!- docl [~docl@user/docl] has joined #hplusroadmap
20:04 -!- Croran [~Croran@user/Croran] has quit [Quit: leaving]
22:53 -!- darsie [~darsie@84-113-82-174.cable.dynamic.surfer.at] has joined #hplusroadmap
23:59 -!- gl00ten3 [~gl00ten@bl5-239-125.dsl.telepac.pt] has joined #hplusroadmap
23:59 -!- gl00ten2 [~gl00ten@2001:8a0:7ee5:7800:46d9:f5c:17a2:432] has quit [Ping timeout: 248 seconds]
--- Log closed Fri Aug 01 00:00:57 2025