turbolib @megalib.bsky.social · 1hConcluding from someone not getting a date that they are a misogynist seems like concluding from someone being unemployed that they are a bad worker 150
turbolib @megalib.bsky.social · 1hThat said, very sympathetic to the broader point about how LLMs may replace many existing tools and transform what we currently regard as best practices 000
turbolib @megalib.bsky.social · 1hI would also observe that "temperature 0" or "fixed seed" is not enough for LLM reproducibility. Even setting aside the inherent stochasticity of serving MoE in multiuser environments, the tool calls themselves may be nondeterministic in extremely complex ways 100
turbolib @megalib.bsky.social · 1hInstead of talking about determinism, I would frame it in terms of reproducibility. I want to know that if I change some stuff and it breaks, I can go back to the old version. Obviously there can also be pitfalls in traditional software systems for that kind of reproducibility 110
turbolib @megalib.bsky.social · 11hWhat's a concrete example of a basic math problem that a modern agent does less reliably than a human? 110
turbolib @megalib.bsky.social · 13/10/2026Boosting a Monte Carlo algorithm to an arbitrarily high success probability has been a standard technique for at least half a century 000
turbolib @megalib.bsky.social · 12/10/2026It is very clearly the case that for many meaningful budgets, and for many important codebases, the AI or assisted human finds more exploits and of greater severity. Not sure how you can read the Mythos report without seeing this 050
turbolib @megalib.bsky.social · 12/10/2026You're the one who made the universal claim. I'm glad you're walking it back. One way to quantify "better" would be to say: given a wide range of time and money budgets, and the goal of finding an exploit of a particular severity, the [AI/human+AI/human] will be more likely to succeed within budget 190
turbolib @megalib.bsky.social · 12/10/2026That's clearly false. For example, AI is better than humans at producing code satisfying formal specifications (as in lean), and better than humans at finding and writing exploits against exploitable bugs. It's also IME better at other things, but those two are pretty undeniable at this point 1110
turbolib @megalib.bsky.social · 12/10/2026Sure. Of course, people are also inconsistent, and can also produce code that looks right but actually has subtle errors. So I still wonder in what cases you expect unassisted humans to outperform AI-assisted humans. 1150
turbolib @megalib.bsky.social · 12/10/2026What's a concrete example of a coding task where an AI-assisted human would be expected to produce a worse outcome than an unassisted human, in your experience? 1100
turbolib @megalib.bsky.social · 12/10/2026TIL that mainstream theory of computing as practiced since the 70s is a malicious exploit of mathematical traditions. 1110
turbolib @megalib.bsky.social · 12/10/2026Yes, including the result this post complains about. It is a huge and surprising result, even though the algorithm itself is galactic (I.e., not practical) 110
turbolib @megalib.bsky.social · 11/10/2026In fact KV weights can be a *large majority* of the total model weights, depending on model and context length. And agentic harnesses can entail quite complex, arbitrarily long-term persistence and renewal of these weights. 110
turbolib @megalib.bsky.social · 11/10/2026Modern LLMs do this. KV weights are weights - on large contexts, the are an appreciable fraction of total model weights. Both MoE and CSA create dynamic, interaction-dependent network topology 110
turbolib @megalib.bsky.social · 10/10/2026The post suggests junior career (late PhD, postdoc, new Prof) are in trouble and recommends funding to help shield them. Doesn't that also imply that curtailing new PhDs would be the responsible thing to do? 100
Reposted by turbolibDominik Schnaus @schnaus.bsky.social · 09/10/2026DINOv2 has never seen a caption, and Qwen3 has never seen an image. We still aligned their embedding spaces without a single image-caption pair. It even works when the images and the captions come from different datasets. Project page: dominik-schnaus.github.io/unpaired-rosetta ⬇️ 428460
Reposted by turbolibQuantian @quantian.bsky.social · 07/10/2026Putting Descartes before the whores, one might say 2894
turbolib @megalib.bsky.social · 06/10/2026The fact that companies release Apache-licensed open models trained at great expense, but often avoid copyleft training data (Apertus isn't unique in that respect) is IMO yet another indictment of the FSF model 110
turbolib @megalib.bsky.social · 05/10/2026One knob that controls how fast the fan blows, and one that controls how hot the air coming out of the vent is. I don't want to specify a temperature target, I want to either be hotter or colder 3270
Reposted by turbolibTed Underwood @tedunderwood.com · 05/10/2026It's okay to learn from everything, and train on everything. That's everyone's birthright, and we shouldn't apologize for it. Prohibitions on copying were a late invention, and a narrowly tailored one. They do not express a general ethical law that all information ought to be priced and paid for. 717627
turbolib @megalib.bsky.social · 05/10/2026Americans will reject socialist regimes that take direct ownership stakes in private firms 010
turbolib @megalib.bsky.social · 05/10/2026What is the moral difference between training a model on libgen books and training a model on blog posts? 100
turbolib @megalib.bsky.social · 05/10/2026I mentioned olmo because I knew off the top of my head it was scraped. Just checked Granite and Apertus, and of course they also use scraped data. Apertus respects robots.txt and avoids copyleft, but they certainly use unlicensed copyrighted data 220
turbolib @megalib.bsky.social · 05/10/2026That isn't true. Olmo's dataset is scraped from the web, like every other text corpus of similar size 250
turbolib @megalib.bsky.social · 05/10/2026I can't even imagine how much people on here would lose their shit if Dems tried messaging like that 000
turbolib @megalib.bsky.social · 05/10/20263% of identifying as Harris -> Paxton in a poll is not in fact particularly good evidence that even a single such voter exists. People just say stuff. That isn't to say the broader point isn't accurate, the mind of the median voter is an unknowable mystery, like the qualia of a bat 020
turbolib @megalib.bsky.social · 04/10/2026MSF is a fine organization. But I don't think there is a coherent definition of "localization" that lauds MSF while denigrating AMF as elitist, top-down, Western 000
turbolib @megalib.bsky.social · 04/10/2026Right, aside from GiveWell and Giving What We Can and Coefficient Giving, what have the EAs ever done for us? 000
turbolib @megalib.bsky.social · 04/10/2026You're complaining that EAs don't care enough about localization, and holding up MSF and OxFam as superior alternatives? 110
turbolib @megalib.bsky.social · 04/10/2026I think this general line of criticism would be more effective if you identified a specific opportunity that would be a better use of funds than what EA orgs currently fund, that EA orgs are ignoring because of elitism or whatever 110
turbolib @megalib.bsky.social · 04/10/2026Yeah that's just not accurate. A cursory glance at GiveWell research shows that they are intensely aware of their limitations (hence the prominent Our Mistakes page) and deeply engaged in incorporating prior research into their recommendations 220
turbolib @megalib.bsky.social · 04/10/2026You are offering socialism as an antidote to problematic "centralization and top-down enlightened guidance"? 110
turbolib @megalib.bsky.social · 04/10/2026I think even the most cursory review of GiveWell research would disabuse you of the idea that they are "not interested in the nuances and side effects" 000
turbolib @megalib.bsky.social · 04/10/2026There is not enough information to evaluate the second claim, but it doesn't sound like a typical EA cause so I am skeptical 100
turbolib @megalib.bsky.social · 04/10/2026The complaints of the author of the article seem to be (1) some EA-adjacent guy didn't promote his cause because he wasn't able to explain why it would be a better use of funds, and (2) some unnamed international fund promoted by EA (who?) is doing maize work instead of cash crop work 200
turbolib @megalib.bsky.social · 04/10/2026As an example, they have directed several grants to ALIMA www.givewell.org/research/gra... headquartered in Senegal with a Senegalese CEO.givewell.orgALIMA — Primary Healthcare and Malnutrition Treatment in Far North Cameroon (June 2025)Note: This page summarizes the rationale behind a GiveWell grant to ALIMA. ALIMA staff reviewed this page prior to publication. 200
turbolib @megalib.bsky.social · 04/10/2026That isn't accurate. As I point out above, GiveWell does give grants directly to local grassroots organizations 100
turbolib @megalib.bsky.social · 04/10/2026I'm sure it's true that they don't regard localization as an end in itself. Why would they? They care about the effectiveness of the aid they provide, and when aid is more effectively provided via local partners, they prefer those local partners 300
turbolib @megalib.bsky.social · 04/10/2026I don't think this is accurate. There are documented examples of bed net diversion. But I am not aware of any RCT suggesting this was ever a "huge issue" endangering overall effectiveness 120
turbolib @megalib.bsky.social · 04/10/2026I don't think this is accurate. GiveWell is the quintessential EA org, and the main thing they do is exhaustively study funding opportunities in order to direct aid. That very much includes local orgs, as you can see from their grants page. 180
turbolib @megalib.bsky.social · 03/10/2026Unions are fundamentally illiberal. The right to a safe workplace, decent working conditions, a secure retirement, and a financial safety net should be universal, not contingent on private association. 110
turbolib @megalib.bsky.social · 03/10/2026Because when your toddler is in school, it will be very helpful either to your family or your neighbors that school (including bus schedules, and before/after care) are compatible with work schedules, rather than every business and school choosing arbitrary hours 001
