toni @tkukurin.bsky.social · 30/09/2026I have no skin in the game here (not a prof, not a student, just "concerned public" :)) based on some of my recent interactions it'd probably be beneficial to broadcast this message via twitter/X as well where a ...shall we say, "hustle"... sentiment might need regularization. 000
toni @tkukurin.bsky.social · 22/09/2026I think they are required by some econometric law to do this no? :D even figure descriptions already identify managers as attending most meetings which might clearly indicate a different plausible mechansim 101
toni @tkukurin.bsky.social · 21/09/2026I truly admire how well the simple "nerdsnipe + provide robust infra" recipe scales :) still have to test out (2) for the link but this seems great. 010
toni @tkukurin.bsky.social · 21/09/2026what's your option space / openness / budget towards LM reviews? 000
toni @tkukurin.bsky.social · 18/09/2026prescriptively: IMO mathematicians might just need to eventually trifurcate into (1) teachers of higher lvl concepts, (2) "big math" (SAIR, abstraction levels over abstraction levels? category theorists rejoice), (3) data science (better real world -> models mapping). seem realistic / sensible? 000
toni @tkukurin.bsky.social · 06/09/2026this is how we've always thought about engineering, right? I'm more curious what you think about trust layers. what I mean is e.g. I would delegate linux kernel tests outside of my ML r&d team even in presence of coding agents pytorch library, clearly. sqlite, yes. sqlite utils? hm, now we talk. 000
toni @tkukurin.bsky.social · 05/09/2026nice effort; I hope the "reading market revolt" does ultimately force writers' hands to higher standards. at least for now evading AI detection is to me a valid proxy of the writer's thinking + more likely correlates w/ my time investment. related discussion on the other bsky x.com/tkukurin/sta...x.comToni Kukurin (@tkukurin) on X@Thom_Wolf @patrickc is that how you read it? I think repetitive oral tradition served diversification, thinking through alternate views of a single topic seems we increasingly offload this process ... 030
toni @tkukurin.bsky.social · 05/09/2026yep, imo part of consistent globalization since age twitter (before that internet>hollywood>radio>books>...) eg how many news articles do you know the true source of? I own maybe few of my ideas, the competition just gets slightly narrower (ps you might enjoy work of @kennethstanley.bsky.social) 001
toni @tkukurin.bsky.social · 04/09/2026thx! fwiw I'm not convinced (all?) founders are naive in that sense, it's just hard to resist the "moloch" of game theory. 010
toni @tkukurin.bsky.social · 04/09/2026well fwiw I subscribed. re/today's piece, the way I've previously seen it is almost like collusion, am I off base? enough big name investors means distribution to big name buyers (usually involved w/same investors) which is revenue, "big VC" buys employees and early product buildout 110
toni @tkukurin.bsky.social · 02/09/2026all good I was just wondering. you did some really cool demos, is that recent or just stuff accumulated over time? 110
toni @tkukurin.bsky.social · 02/09/2026previously saw on twitter, it's a nice demo but is this not domain dependent? the OP pubmed.ncbi.nlm.nih.gov/7569981/ replicates in ideal conditions. fm materially matters in even low noise settings pubmed.ncbi.nlm.nih.gov/15677723/ eg. github.com/tkukurin/hea... or did I misunderstand your setup? 110
toni @tkukurin.bsky.social · 30/08/2026I see, thanks for expanding. I didn't mean to imply his or my taxonomy was nearly precise or complete. 020
toni @tkukurin.bsky.social · 30/08/2026I assume you'd meant overloading concepts such as attention and memory (eg slide 16 of neurips.cc/virtual/2025...) but just to point out, there was a "morally" similar ML (well, LLMs...) paper published last neurips on the datasets&evals track: proceedings.neurips.cc/paper_files/...neurips.ccNeurIPS Invited Talk The Oak Architecture: A Vision of SuperIntelligence from ExperienceNeurIPS 2025 110
toni @tkukurin.bsky.social · 12/08/2026probably should have anticipated this, sorry :) I will have a look. 010
toni @tkukurin.bsky.social · 12/08/2026congrats on the launch! quick feedback: iirc the app requested somewhat lenient access to my account (full edit ability over profile). is this required for core functionality? might be worth interrogating Claude, I am not overly familiar w/ atproto ACLs 110
toni @tkukurin.bsky.social · 11/08/2026added to reading list, this seems exactly the framing I've recently been thinking of. iiuc a "Bayesian agent embedded in a symbolic" world would perfectly fit here as well? bsky.app/profile/tkuk... 110
toni @tkukurin.bsky.social · 10/08/2026the hopeful / (maybe?) positive equilibrium is filtering on AI-sounding content via automated methods. eg. instead of "linguistic flourish" effort to write begins to mean "avoid pangram/synthID detection" I am not yet completely convinced, but can see this developing into a win-win global outcome. 000
toni @tkukurin.bsky.social · 09/08/2026absolutely, precisely why "annealing" recall-to-precision is core to the process. I see it as shifting UX research forward into the deployment phase. "just work" defaults + dynamic config to peculiarities of an individual's workflow. far from solved, still plenty of room for me to be wrong :) 000
toni @tkukurin.bsky.social · 09/08/2026I have some signal you could train (some of) this anticipation for a significant number of regular user workflows - start as explicit high-recall escalation to HITL, over time backs off essentially "learned permission mgmt." it's not trivial, interesting engineering design problem for sure. 100
toni @tkukurin.bsky.social · 09/08/2026nice! checks out that I find most of them familiar yet well worth the re-read :) 010
toni @tkukurin.bsky.social · 09/08/2026true but generally entropy messes up everything I have ~strong opinions on moving escalation to agent layer (which I guess implicitly is alignment) given societal trajectory it's likely the only long term feasible solution (in addition to standard observability methods) bsky.app/profile/tkuk... 010
toni @tkukurin.bsky.social · 09/08/2026what's the generative process behind these posts? do you read them weekly or do they come from some personal archive? one of the better bsky series :) 100
toni @tkukurin.bsky.social · 09/08/2026hm ok I was more referring to what happened prior to the hacking if we assume by "intelligence" and "agents" they mean "use an LLM" I can see your conclusion too. otoh takeaways "continuous red teaming", "automated remediation" and "automated incident response" are just obvious reliability advice. 010
toni @tkukurin.bsky.social · 09/08/2026flip side is the transparency should clearly be acknowledged/encouraged bsky.app/profile/tkuk... 000
toni @tkukurin.bsky.social · 09/08/2026I assume non-malice but I truly don't get whether to characterize as naive, lazy, or otherwise. this is 101 systems-level negligence. 210
toni @tkukurin.bsky.social · 09/08/2026sentiment is valid but tbh not sure re/framing (my understanding thereof at least) seems clear a reward hacking agent exists regardless of the substrate (eg. minor delta between a tool composition and code, complex systems) training and explicit escape hatch ("escalate") seems a more feasible impl 010
toni @tkukurin.bsky.social · 08/08/2026"tire of (...) transcripts of professors having conversations with ChatGPT" even terry tao's are difficult to read :'( I hope there's a push for some kind of "math to application" strategy with his sair.foundation 50 "irrelevant" thm papers to 1 useful outcome. toni.leaflet.pub/3ms7fxgheak2...toni.leaflet.pubten AIdvances in nothing reallythe hopium! 010
toni @tkukurin.bsky.social · 02/08/2026certainly. unlikely to find a universe where I believe otherwise. 000
toni @tkukurin.bsky.social · 01/08/2026true; I could be wrong but the delta I still see is in one case explicitly modeling a formal system (deduction under full observability, incomplete or inconsistent). in the other it is unclear (we understand to be modeling partial observability, aleatoric uncertainty, probabilistic decision making) 000
toni @tkukurin.bsky.social · 01/08/2026funnily enough, I took it to mean the decision theory small/big world concept which I suppose also fits 010
toni @tkukurin.bsky.social · 01/08/2026it's a sensible ontology albeit might become increasingly aliased over time (e.g. prompt v config). did you deliberately decide against some of the existing NLP et al eval suites? e.g. I am personally a fan of inspect.aisi.org.ukinspect.aisi.org.ukInspectOpen-source framework for large language model evaluations 010
toni @tkukurin.bsky.social · 31/07/2026nice sentiment at the end. the reward hacking scientist cures type-1 diabetes with a starvation diet. science is the process. 010
toni @tkukurin.bsky.social · 31/07/2026but this is at least WAI? labs stopped publishing capabilities research ok but build on each others' infra shortcomings. morbidly amusing, not grounds for denigration. 000
toni @tkukurin.bsky.social · 29/07/2026afraid it's also emerging as a spontaneous order charliebecker.substack.com/p/is-an-ai-c...charliebecker.substack.comIs an AI company buying up all the used books? (I don’t think so, not this time.) | #103But no matter what, the Library of Alexandria is still burning. 000
toni @tkukurin.bsky.social · 28/07/2026IMO theoretically Abel/Schuurmans seem to have the ~right frame; practically I love @andymatuschak.org's work. none do exactly what I had in mind. def >>1 unified views (query expansion, self-curriculum, more compute vs memory ... en.wikipedia.org/wiki/D%C3%A9...), will look into Nathan's work thx!en.wikipedia.orgDéformation professionnelle - Wikipedia 010
toni @tkukurin.bsky.social · 28/07/2026I don't see it explicitly spelled out -- how do you see the interaction of environment design vs symbol manipulation or do you know good reference work to set this framework? ie. symbol manipulation as a special case of environment design (which also goes to the pretrain-v-in-context line of work) 110
toni @tkukurin.bsky.social · 28/07/2026arguably worse, instead of a remote server you're wasting other humans' time. I think this makes LM-mediated review passes largely non-negotiable. 000
toni @tkukurin.bsky.social · 24/07/2026+1, we already have social media as a model organism. some synchronous thinking: xcancel.com/tkukurin/sta...xcancel.comVerifying your browser… 020
toni @tkukurin.bsky.social · 24/07/2026A-la this? github.com/mitchellh/vo...github.comGitHub - mitchellh/vouch: A community trust management system based on explicit vouches to participate.A community trust management system based on explicit vouches to participate. - mitchellh/vouch 100
toni @tkukurin.bsky.social · 19/07/2026re/ terminology I'll just reference this lovely web-piece of art. agents2001.csc.liv.ac.uk/about.html agents as terms du décennies; the interpretation had not even drifted too far from its origins :)agents2001.csc.liv.ac.uk Autonomous Agents 2001 000
toni @tkukurin.bsky.social · 19/07/2026I'd assumed that might be it; the last part I still ~ contest. X layer uses intelligence and vice versa; eg some "intelligence" is the impl layer for my app. I constrain access layer, hyrum's law leaks through nevertheless. prob worth clarifying at another pt rather than offhand bsky comment tho. 100
toni @tkukurin.bsky.social · 18/07/2026hm. at first glance I liked it but on 2nd thought is this not conflating mechanisms? it seems to me "intelligence" is the core of what humanity does. knowledge, the dynamic societal system. the whole reason you are creating some artefact. lower layers (platform/format) are representation mechanism 100
toni @tkukurin.bsky.social · 28/05/2026abstracting the setup slightly ("highly general minimax with a goal function and human-niche heuristics"), some ai researchers did have this view as path to AGI right (cyc flavor)? + was alphafold not a result of this "overfitting", showing ai research has an appropriate regularization mechanism? 120
toni @tkukurin.bsky.social · 27/04/2026repo-specific: put superset of skills in a git repo, bootstrap per project by asking your coding agent to appropriately distribute (subset of) files. branch and upstream any updates. 110
toni @tkukurin.bsky.social · 30/03/2026futurist riff: agents do async alignment humans only sync? :) 020
toni @tkukurin.bsky.social · 29/03/2026ha. they promised decentralization, they delivered. with agents, get ready for thrice :) bsky.app/profile/jay.... 020