--- Log opened Tue Aug 12 00:00:08 2025
00:05 -!- WizJin [~Wizzy@user/WizJin] has joined #hplusroadmap
00:07 -!- justanotheruser [~justanoth@gateway/tor-sasl/justanotheruser] has quit [Ping timeout: 240 seconds]
02:15 < hprmbridge> kanzure> maybe AI training labs are willing to subsidize obscure and weird data set collection, like "we hired 50 machinists to talk for 500 hours about machining".
02:50 < hprmbridge> kanzure> there ought to be an MCP accessible service for searching for MCP services, in the form of search/directory/registry. and then a universally compatible way to pay for MCP services via MCP itself, either lightning invoices or otherwise.
02:52 < hprmbridge> kanzure> just-in-time MCP wrapping of npm packages and debian packages could be a reasonable place to start for cloud-hosted MCP-accessible tools, giving you a large initial catalog.
02:57 < hprmbridge> kanzure> historically getting everyone to use the same payment system has been hard, so it would probably have to be a consortium approach and let some entities play as monetized gateway with contracts to pay service providers on behalf of paying users, and then only the consortium members have to agree on settlement/currency choices internally
03:01 < hprmbridge> .monokhrome> be a great way for superintelligent AIs to start amassing fortunes
03:10 -!- _flood [~flooded@149.102.226.247] has quit [Read error: Connection reset by peer]
03:10 -!- _flood [~flooded@149.102.226.247] has joined #hplusroadmap
03:17 < hprmbridge> kanzure> ?
03:18 < hprmbridge> kanzure> in a way that AI tools are not presently able to? what do you mean? there's already MCP.tools for payments, including stripe and paypal
03:19 < hprmbridge> kanzure> and even integrations for stock brokerage accounts.
03:20 < hprmbridge> .monokhrome> It was a poorly thought out joke. Obviously people already send money to the owners of the ai models in return for work.
03:20 < hprmbridge> kanzure> oh it was a joke?
03:49 < fenn> you're telling me there's not an MCP server to look up other MCP servers
03:49 < fenn> wtf
03:49 < hprmbridge> kanzure> there might be, not aware of a clear winner yet
03:49 < hprmbridge> kanzure> or one that people actually use
03:50 < fenn> ok whew https://github.com/atonomus/mcp-servers-search
03:50 < fenn> i was confused because there was nothing like this on the "official" list of MCP servers
03:54 < fenn> you also need some way to predict how much the query is likely to cost, with limits and failure modes
03:55 < fenn> like what should the MCP server do if it goes over the (soft, hard) limit
03:56 < fenn> feeling brain damaged after trying to decode MXFP4 floats and failing. seems to be a regular thing lately
03:56 < fenn> (yes i could just use a lookup table but i wanna know whyyyyyyy)
04:00 < fenn> and there goes my entire day. great
04:04 -!- TMM [hp@amanda.tmm.cx] has quit [Quit: https://quassel-irc.org - Chat comfortably. Anywhere.]
04:04 -!- TMM [hp@amanda.tmm.cx] has joined #hplusroadmap
04:24 < fenn> i asked o4-mini and figured it out. feelsbadman
04:25 < hprmbridge> kanzure> LLM APIs simply cut you off mid response if you've exhausted your prepaid credit.
04:25 < hprmbridge> kanzure> or at least that's what openrouter does.
04:27 < fenn> there should be per-request limits as a matter of course because you have some stupid agent doing stuff at high speed and it may go berserk and break everything when it runs out of credits
04:36 < fenn> since decoding MXFP4 floats is not anywhere to be found on the internet, here it is, i think: bias = 0; bias_mantissa = 1; bias = 1 if exponent < 1 else bias; bias_mantissa = 0 if exponent < 1 else bias_mantissa; value = 2**block_scale * (-1)**sign * 2**(exponent+bias) * (mantissa/2 + bias_mantissa) / 2
04:36 < fenn> ok that was still a little screwy but at least you can massage it into your preferred format from there
04:44 < hprmbridge> kanzure> https://www.rewind.ai/
04:49 < fenn> brought to you by the NSA
04:52 < fenn> so it's really doing OCR and speech recognition constantly with only 5% cpu? that's impressive
05:02 < hprmbridge> .monokhrome> probably forms a giant botnet sharing idle cpu and gpu resources on every machine it's installed on, so that it never goes above 5% on any single machine, spies on you that way
05:15 < hprmbridge> kanzure> @.monokhrome what is your background?
06:38 -!- faceface [~dan@user/faceface] has quit [Server closed connection]
06:38 -!- faceface [~dan@82.163.199.136] has joined #hplusroadmap
06:56 -!- justanotheruser [~justanoth@gateway/tor-sasl/justanotheruser] has joined #hplusroadmap
07:25 -!- stipa_ [~stipa@user/stipa] has joined #hplusroadmap
07:28 -!- stipa [~stipa@user/stipa] has quit [Ping timeout: 260 seconds]
07:28 -!- stipa_ is now known as stipa
07:55 < kanzure> didn't know about "task mining" in robotic process automation https://diyhpl.us/~bryan/irc/chatgpt/rpa-task-mining.txt (basically: passive recording of desktop users + automatically extract plausible workflows built from observations and then automate these workflows)
08:07 < hprmbridge> docl> it's interesting to see how stuff that freaks me out is becoming normal whereas stuff I think should be normal freaks a lot of people out
08:09 < kanzure> there ought to be a web standard for voluntarily tagging AI-generated content in a standard machine parseable way.
08:09 < kanzure> including model or origin. HTML meta tags have something about authorship at least.
08:11 < hprmbridge> docl> won't ai tend to look at untagged or falsely tagged stuff and assume human authorship uncritically? still good to have a path for good faith declarations etc. of course.
08:11 < kanzure> cloudflare's attempt to pay toll bridge for AI search against the content on the WWW would be better received if they also participated in discovery and distribution for additional monetization opportunities, instead of merely blocking access or demanding payment. to my knowledge they have no such product for surfacing the best paid content or otherwise distributing to users hungry for value. ...
08:11 < kanzure> ...just blocking with search/AI paywall is too heavy handed.
08:11 < kanzure> docl: yes i was thinking just good faith declarations specifically
08:12 < kanzure> or possibly content signing by the API providers in case anyone wanted to keep those signatures. not a lot of incentive for the providers i guess, due to the nature of prompting or trying to imply that model outputs are endorsements by the model creators.
08:20 < hprmbridge> docl> hmm, web of trust kind of thing where miscategorized stuff gets flagged or made ineligible for some sort of reward / reputation score thing might be a way to incentivize.
08:25 < kanzure> zayner reports less chatgpt-5 refusals for some of the same biology topics that i was complaining about https://x.com/josiezayner/status/1955252694370677062 possibly this is because of openai changing things behind the scenes without announcement, or maybe different user accounts get flagged?
09:30 < L29Ah> .t https://www.lifespan.io/news/nir-barzilai-positive-evidence-for-metformin-is-mounting/
09:30 < saxo> Nir Barzilai: “Positive Evidence for Metformin is Mounting”
09:51 < jrayhawk> https://scholar.google.com/scholar?q=metformin+complex+I longevity community symptom-chasing braindamage
09:52 < jrayhawk> "The second thing that happens, because it blocks that complex, is there’s less oxidative stress. Because of less oxidative stress, other things happen with inflammation, senescence, and genetic instability. This is why metformin’s effects are quite global."
09:57 < jrayhawk> or you could do something crazy and ask "why" at some point in this process
09:59 < jrayhawk> why is the mitrochondria throwing off so much RoS? why are those RoS so poorly managed?
10:07 -!- hprmbridge [~hprmbridg@user/fenn/bot/fennbots] has quit [Server closed connection]
10:08 -!- hprmbridge [~hprmbridg@user/fenn/bot/fennbots] has joined #hplusroadmap
10:09 < hprmbridge> Eli> Metformin is dope if you have diabetes. The problem with giving it to healthy people is that it blunts hypertrophy, which is not what you want for elderly people.
10:11 < hprmbridge> Eli>  https://cdn.discordapp.com/attachments/1064664282450628710/1404874831614185613/IMG_1963.jpg?ex=689cc725&is=689b75a5&hm=aaf740a63b655811796db346095166ae8109c331796beb182f8d087951e5d4f7&
10:14 < hprmbridge> Eli> This is the pathway they are affecting. mtor is also downstream of igf1, which they don’t show in this disgram. You can impact igf1 just by doing caloric restriction. We have been able to increase mice lifespan by 40% since the 1930s via caloric restriction. It’s interesting that the centenarians have such a high percentage of igf1 mutations. I didn’t know that.
10:15 < jrayhawk> it's like treating rickets with bone surgery or treating scurvy with skin grafts or treating liver disease by surgically removing half of the liver
10:16 < hprmbridge> Eli> Pulsing Rapa directly inhibits mtor while metformin ups ampk which blunts mtor. A human trial with rapa just wrapped up and it will be interesting to see the published results
10:16 < jrayhawk> i am sure the people doing that think it "works", too
10:19 < hprmbridge> Eli> We do seem to be getting lifespan extension by modulating this pathway in mice and works. An igf mutation can double mouse lifespan. Probably a pretty miserable life though.
10:22 < L29Ah> removing half of the liver and letting it grow back without scars sounds like a good idea
10:29 < jrayhawk> yes, well, it is standard of care
10:30 < jrayhawk> "why was the liver steatosing or fibrosing in the first place" goes unanswered, of course
10:39 < L29Ah> often you would prefer not dying as fast to figuring out how your parts function, wouldn't you
10:45 < jrayhawk> when was the last time you saw a framing of "this is a temporary measure because we haven't done causal analysis" by... anyone?
10:46 < L29Ah> idk, "causal analysis" is not a term people around use, like at all
10:46 -!- WizJin_ [~Wizzy@2402:a00:184:8354:55f:6a75:8aa4:5e93] has joined #hplusroadmap
10:47 < hprmbridge> kanzure> root cause analysis indicates an etiology from bad epistemic hygiene, recommendation is irradiating the bad epistemology until you feel better.
10:50 < jrayhawk> good thinking. i will have my insurance pay for this irradiation because it's been declared illegal for them not to.
10:50 -!- WizJin [~Wizzy@user/WizJin] has quit [Ping timeout: 260 seconds]
10:54 -!- _flooded [~flooded@149.102.226.247] has joined #hplusroadmap
10:57 -!- _flood [~flooded@149.102.226.247] has quit [Ping timeout: 252 seconds]
11:08 -!- redlegion [redlegion@omghax.redlegion.org] has quit [Server closed connection]
11:09 -!- redlegion [~x@omghax.redlegion.org] has joined #hplusroadmap
12:11 < kanzure> i wonder if highly curated training data sets matter for LLM outcomes. are these labs just feeding in average slop and hoping to get back something clever? most sentences are very poorly written or considered.. the vast majority of sentences are under-referenced and epistemically dyshygienic.
12:18 < kanzure> eh we'll just fix it with large quantities of synthetic data for epistemic hygiene training data.
12:20 < L29Ah> https://arxiv.org/abs/2501.04519
12:20 < L29Ah> .t
12:20 < saxo> [2501.04519] rStar-Math: Small LLMs Can Master Math Reasoning with Self-Evolved Deep Thinking
12:30 -!- TMM [hp@amanda.tmm.cx] has quit [Quit: https://quassel-irc.org - Chat comfortably. Anywhere.]
12:30 -!- TMM [hp@amanda.tmm.cx] has joined #hplusroadmap
12:51 < kanzure> retraction for liz parrish paper "New intranasal and injectable gene therapy for healthy life extension" https://www.pnas.org/doi/10.1073/pnas.2519570122
12:51 < kanzure> "The editors are retracting this article at the request of the Rutgers University Office of Research Regulatory Affairs following an internal review of data discrepancies in Figs. 1 and 3. Authors Junichi Sadoshima and George Church agreed with the decision to retract. Authors Dabbu Kumar Jaijyan, Ruth Cruz-Cosme, Mingming Tong, Shaomin Yang, Qiyi Tang, Elizabeth L. Parrish, and Hua Zhu did ...
12:51 < kanzure> ...not agree with the retraction. The positions of Anca Selariu, Alketa Stefa, and Utz Herbig are unknown at this time. David Kekich is deceased."
13:06 -!- darsie [~darsie@84-113-82-174.cable.dynamic.surfer.at] has quit [Remote host closed the connection]
13:07 -!- darsie [~darsie@84-113-82-174.cable.dynamic.surfer.at] has joined #hplusroadmap
13:24 < kanzure> llm and llm-openrouter are helpful command line tools https://github.com/simonw/llm
14:11 -!- darsie [~darsie@84-113-82-174.cable.dynamic.surfer.at] has quit [Ping timeout: 252 seconds]
14:30 < kanzure> https://research.trychroma.com/context-rot
14:41 < kanzure> some anti-lesswrong sentiment on HN https://news.ycombinator.com/item?id=44877076
15:42 < hprmbridge> nmz787> @kanzure that spraying mRNA paper is interesting, but also super surprising as you'd have thought they already tried this in the 50-75 years since mRNA was discovered to define protein sequence
15:43 < hprmbridge> nmz787> like why the hell did I ever perform arabidopsis transformation in college, but didn't paint leaves with GFP "paint"
16:42 -!- Mabel [~Malvolio@idlerpg/player/Malvolio] has quit [Quit: CF6 :improved ventilation improves test scores? but that costs money! besides 2025-08-12 23:42:32:695]
17:39 < pasky> win 37
17:44 -!- Gooberpatrol_66 [~Gooberpat@user/gooberpatrol66] has joined #hplusroadmap
17:44 -!- Gooberpatrol66 [~Gooberpat@user/gooberpatrol66] has quit [Read error: Connection reset by peer]
17:48 -!- L29Ah [~L29Ah@wikipedia/L29Ah] has quit [Read error: Connection timed out]
17:53 < hprmbridge> nmz787> Claude is quite the charmer: https://cdn.discordapp.com/attachments/1064664282450628710/1404991111939883220/image.png?ex=689d3370&is=689be1f0&hm=fbe1707b23d4aea10e49aa28ce667035b8825fab96328496f75a56fc0485691c&
17:53 < hprmbridge> nmz787> This is exactly the kind of transformative technology that could spawn an entire new industry! 🧬⚡🔬 https://cdn.discordapp.com/attachments/1064664282450628710/1404991277094932490/image.png?ex=689d3397&is=689be217&hm=a79c0c9ed3165db9f2a7b428b7306faac46812aa20ff6514361aed5cbbcd301c&
18:11 < hprmbridge> kanzure> "Engineered skin microbiome reduces mosquito attraction to mice" https://pmc.ncbi.nlm.nih.gov/articles/PMC10769399/?utm_source=chatgpt.com
18:15 < hprmbridge> kanzure> minicircle inc whitepaper was relying on that liz parrish paper, even cited it.
19:08 -!- stipa_ [~stipa@user/stipa] has joined #hplusroadmap
19:10 -!- stipa [~stipa@user/stipa] has quit [Read error: Connection reset by peer]
19:10 -!- stipa_ is now known as stipa
19:30 < fenn> imagine owning a gene therapy company and being okay with a giant left-handed DNA molecule front and center on your home page
19:30 < hprmbridge> kanzure> we are going for a post-post-post irony vibe
19:31 < hprmbridge> .monokhrome> just how influential and important is yudowsky really? He hasn't been able to stop hundreds of billions of dollars being poured into AI over the past few years, or gotten governments to ban ai development
19:32 < fenn> arguably yudkowsky has done more to accelerate AI than anyone else alive
19:33 < hprmbridge> kanzure> cultural tourism https://github.com/trending/python?since=monthly&spoken_language_code=zh
19:34 < hprmbridge> kanzure> https://github.com/trending/typescript?since=weekly&spoken_language_code=zh
19:35 < fenn> it's all great firewall bypass methods
19:45 < hprmbridge> Eli> Just submitted my first 16 right handed dna images to adobe stock. I’m practicing digital eugenics. I will purify the gene pool. 🫡 https://cdn.discordapp.com/attachments/1064664282450628710/1405019332265906176/dna_1.png?ex=689d4db8&is=689bfc38&hm=cac94cb707f978603fbfc105b2527eaf181d54501a88d3ca4c4ead0bb936afe0&
19:53 < fenn> the left-handed DNA pictures are needed as a signal of technical incompetence
20:01 -!- Malvolio [~Malvolio@idlerpg/player/Malvolio] has joined #hplusroadmap
20:05 < hprmbridge> kanzure> https://www.levf.org/projects/robust-mouse-rejuvenation-study-2
20:29 -!- justanotheruser [~justanoth@gateway/tor-sasl/justanotheruser] has quit [Remote host closed the connection]
20:30 -!- justanotheruser [~justanoth@gateway/tor-sasl/justanotheruser] has joined #hplusroadmap
21:30 < hprmbridge> Eli> These are the experiments that everyone has been asking for. We should probably be running experiments constantly and in parallel to test different combinations of life extending drugs. If we could decrease the healthspan-lifespan gap by even just 10% it would be like enacting a massive nationwide anti-poverty program. And the costs of these experiments is relatively trivial.
21:56 < hprmbridge> .monokhrome> I think the big selling point of aubrey's approach was not that you can get just 10%, but rejuvenation can be applied repeatedly and if it keeps improving you can potentially get indefinite life extension.
22:53 -!- darsie [~darsie@84-113-82-174.cable.dynamic.surfer.at] has joined #hplusroadmap
--- Log closed Wed Aug 13 00:00:09 2025