you-have-just-experienced-things @yhjet.bsky.social · 08/10/2026yes and it's very spiritually Mansonian as well (the o.g. rebus puzzle in MAZE was never solved without hints) I feel very accomplished for solving #3 properly and accidentally guessing #1 while flipping through letters. (the other two, my spouse reasonably got impatient and googled hints) 130
you-have-just-experienced-things @yhjet.bsky.social · 08/10/2026it was really magical playing Blue Prince after having grown up with MAZE -- think I heard about it from one of your threads earlier this year, so ty for that! there is art from Christopher Manson in the game (the two sketches in each room) which evoked a bunch of nice sense-memories 350
you-have-just-experienced-things @yhjet.bsky.social · 08/10/2026"over 700 files at once is not a demonstration of scholarship" well I'm gonna be using that one in PR reviews 0120
you-have-just-experienced-things @yhjet.bsky.social · 08/10/2026I haven't read it yet, I have fond childhood memories of another minotaur-inspired romp though (Christopher Manson's MAZE) 271
you-have-just-experienced-things @yhjet.bsky.social · 08/10/2026yeah I'm giving the east coast view on (caricature of?) bay area VC, I have no personal experience with VC. my recollection of that view is: "no denying the VC approach generates great returns [in the wonky 0 rates world of early 10s] but they're just flinging cash, not rigorously finding alpha" 130
you-have-just-experienced-things @yhjet.bsky.social · 08/10/2026but when it comes to e.g. lesswrong folks I am torn because they were _really committed_ to one weird position and were at least "directionally" correct about it, but unclear how much to credit them because their track record on the details is not as good 030
you-have-just-experienced-things @yhjet.bsky.social · 08/10/2026personally my sympathies are on the east coast side -- the VC playbook of making a huge number of weird bets in the hope that one of them 1000x's obviously works, but implies little about one's discernment 240
you-have-just-experienced-things @yhjet.bsky.social · 08/10/2026these days I sell shovels (software related to trading mortgage-backed securities) which is nice because I can do it from rural quebec 150
you-have-just-experienced-things @yhjet.bsky.social · 08/10/2026I should probably blog about this but TL;DR I dropped out of my PhD program because I was horribly depressed and joined a small hedge fund trading mortgage-backed securities 280
you-have-just-experienced-things @yhjet.bsky.social · 08/10/2026this feels like an east cost vs west coast cultural difference. my guess is if I pinged those guys and told them about the term "directionally correct" they would laugh 150
you-have-just-experienced-things @yhjet.bsky.social · 08/10/2026in my trading desk days (mortgage bonds, post-crisis), my impression was that getting lucky on a single big bet was enviable but not respectable -- real respect was reserved for the guys with consistent returns above the index. I don't think anyone took Burry very seriously 290
you-have-just-experienced-things @yhjet.bsky.social · 08/10/2026"everyone will live in abundance on UBI" strains plausibility alongside industry statements like "our TAM is the entire labor market" and rising costs at the grocery store, especially in a hollowed-out rural town 160
you-have-just-experienced-things @yhjet.bsky.social · 08/10/2026- "alignment in general is more important than alignment to whom" - "when incidents happen, blameless postmortems are better than flinging lawsuits" - "everyone will live in abundance on UBI, the real final boss is the meaning crisis" these are all DOA 1161
you-have-just-experienced-things @yhjet.bsky.social · 08/10/2026talking to my neighbors (in rural quebec) has clarified for me how quickly narratives from the big AI labs will dissolve on contact with populism 2180
you-have-just-experienced-things @yhjet.bsky.social · 08/10/2026it would be pretty funny if antis get negatively polarized into ultrafinitism 000
you-have-just-experienced-things @yhjet.bsky.social · 08/10/2026Hamming's "Mathematics on a Distant Planet" probably deserves a re-read 100
you-have-just-experienced-things @yhjet.bsky.social · 08/10/2026excited for everyone to suddenly have very strong opinions about what constitutes a human-understandable explanation of a proof 100
you-have-just-experienced-things @yhjet.bsky.social · 07/10/2026like there's clearly a limit to how much productive critique (if any) you can wring out of the scaffolding around the model, you gotta engage with the actual thing dhq.digitalhumanities.org/vol/17/2/000...dhq.digitalhumanities.orgDHQ: Digital Humanities Quarterly: How to Do Things with Deep Learning Code 000
you-have-just-experienced-things @yhjet.bsky.social · 07/10/2026there's an obvious interdisciplinary opportunity here with mechinterp, but, y'know, two cultures... 110
you-have-just-experienced-things @yhjet.bsky.social · 07/10/2026I would love to read "critical prompt studies" and it's disappointing that the CCS folks don't seem to grasp the potential wg.criticalcodestudies.com/index.php?p=...wg.criticalcodestudies.comWhat Can Critical Code Studies Read When Programmers Stop Writing?Something seems to have happened to the language of programming in the past decade. 1111
you-have-just-experienced-things @yhjet.bsky.social · 06/10/2026this will be the first model to get banned by the office québécois de la langue française for improper francisation 1325
you-have-just-experienced-things @yhjet.bsky.social · 06/10/2026the interface: assumeLiability(...) 120
you-have-just-experienced-things @yhjet.bsky.social · 06/10/2026in many ways yeah e.g. everyone reinvents expert systems when their users demand more transparent or more deterministic models. and there's certainly a demand for that with LLMs, I just think the solutions will be very different 000
you-have-just-experienced-things @yhjet.bsky.social · 06/10/2026I don't but an agent in /goal mode with a knowledge base is kinda close to the dream of the BDI interpreter 100
you-have-just-experienced-things @yhjet.bsky.social · 06/10/2026so if I wanted to "do a breakthrough" in LLM system design, I probably wouldn't dust off any of that stuff, except maybe for framing the problem 100
you-have-just-experienced-things @yhjet.bsky.social · 06/10/2026for example LLM subagent interactions are more flexible than anything imagined in a cognitive architecture from the before-times. whole class of protocol problems just gone (extremely online framing: ACT-R was all grug tech, no gorm fluid. but now we know that maybe gorm fluid is all you need) 100
you-have-just-experienced-things @yhjet.bsky.social · 06/10/2026in fact with LLMs we see the exact opposite, harnesses have to evolve in lockstep with the capabilities of the models. the precise details of those jagged edges of capabilities have overwhelming effects on the system design 210
you-have-just-experienced-things @yhjet.bsky.social · 06/10/2026another example: part of the conceit of architectures like ACT-R was that imposing structure on top of trivial cognitive modules could yield useful insights. (problems of symbol grounding and structured learning were out of scope.) in hindsight, they were useful for cogsci, but not for advancing AI 100
you-have-just-experienced-things @yhjet.bsky.social · 06/10/2026(mayyyybe belief-desire-intention could make a comeback, if I had to bet on anything from MAS being useful would be that, but idk, it still feels too prescriptive) 100
you-have-just-experienced-things @yhjet.bsky.social · 06/10/2026so while many of the problems in the MAS literature are relevant (communication protocols, coordination mechanisms, trust, distributed state), the proposed solutions have not been useful, and I expect this to continue 100
you-have-just-experienced-things @yhjet.bsky.social · 06/10/2026even putting the ontology timesink aside, MAS in the 90s focused on agents which were (1) narrow (2) predictable (3) needed formal protocols because they couldn't handle ambiguity (4) operating in well-defined action spaces ...but none of these are true of LLMs 100
you-have-just-experienced-things @yhjet.bsky.social · 06/10/2026when I was in CS academia multi-agent systems and cognitive architectures were no longer in vogue but still publishable. the MAS literature was going nowhere, afflicted with terminal modeling disease (just one more ontology, just one more ontology bro and then we will achieve emergence) 110
you-have-just-experienced-things @yhjet.bsky.social · 06/10/2026(uh, take this thread with a massive grain of salt because I have been out of the field for a long time) 100
you-have-just-experienced-things @yhjet.bsky.social · 06/10/2026it's very true that computer science academia has goldfish memory. I would like to make the case that for multi-agent systems literature specifically, this is good 110
you-have-just-experienced-things @yhjet.bsky.social · 05/10/2026'What “AI” did to machine learning' also a really funny phrase because in the SVM era lots of aging GOFAI guys (including my advisor) were salty about what machine learning had done to "AI" 150
you-have-just-experienced-things @yhjet.bsky.social · 05/10/2026"making only the usual human mistakes" very charitable compression of "shoveling more worthless conference papers into the massive flaming landfill of worthless conference papers" 260
you-have-just-experienced-things @yhjet.bsky.social · 05/10/2026yes I remember claims that scaling up GPT-3 would lead to AGI as early as 2021 -- definitely a fringe position but it was around x.com/ethanCaballe...x.comEthan Caballero (@ethanCaballero) on X#Neural_Scaling_Laws #AGI 020
you-have-just-experienced-things @yhjet.bsky.social · 04/10/2026not sure if there is a good unifying frame--in the 80s and 90s "cognitive architectures" (Soar, ACT-R, etc) and multi-agent systems were popular and you can kinda cram that into a utility maximizer framework. by 2005, it was all about SVMs and gradient boosted trees and stuff. by 2015, it's all DL 011
you-have-just-experienced-things @yhjet.bsky.social · 04/10/2026in 1995 I think it was a defensible framing device, it's very GOFAI-inspired, but keeping it in subsequent editions was questionable 110
you-have-just-experienced-things @yhjet.bsky.social · 04/10/2026sort of related to this, I remember keeping an eye out for whether EY got his belief in nigh-magical sample efficiency from Jaynes, but no, it wasn't discussed at all 160
you-have-just-experienced-things @yhjet.bsky.social · 04/10/2026ok nevermind this rules I'm going to spend all day on this 010
you-have-just-experienced-things @yhjet.bsky.social · 04/10/2026so far this is reminding me how much I hated the quotes threads 220
you-have-just-experienced-things @yhjet.bsky.social · 04/10/2026(sorry for spamming your replies with this grace - maybe i should just go off and blog) 130
you-have-just-experienced-things @yhjet.bsky.social · 04/10/2026there is also a fascinating subthread from a few years later www.lesswrong.com/posts/kXSETK...lesswrong.comThe Level Above Mine — LessWrongComment by Wei Dai - As I asked in the linked comment, if you're the only philosopher in the team, how will others catch your mistakes? I would not trust that when you stop feeling confused, the probl... 240
you-have-just-experienced-things @yhjet.bsky.social · 04/10/2026this was originally posted on OvercomingBias which didn't have threaded replies, he does reply later www.lesswrong.com/posts/kXSETK...lesswrong.comThe Level Above Mine — LessWrongComment by Eliezer Yudkowsky - Manuel, "enroll in a grad program for AI" != "you're smart, you should go to college". Kragen, the short answer is, "It's easy to talk about the importance of effort if... 250
you-have-just-experienced-things @yhjet.bsky.social · 04/10/2026from here www.lesswrong.com/posts/kXSETK... 020
you-have-just-experienced-things @yhjet.bsky.social · 04/10/2026hmm actually Jaynes might be the one place where the sequences do properly acknowledge their influences, but as norvid says above the aura projection here feels...misaligned with the experience of reading the actual source material (ime) www.lesswrong.com/posts/kXSETK...lesswrong.comThe Level Above Mine — LessWrong(At this point, I fear that I must recurse into a subsequence; but if all goes as planned, it really will be short.) … 270
