04:18:44 https://words.filippo.io/seeds-solved/ 04:42:35 (and this AI costs 200$ per month unlike real humans) 11:19:25 <321bob321> Can they find satoshi ? 11:24:15 after reading it, it seems like a really easy problem. I am more surprised that people were looking but couldn't find this. 11:24:35 @monero.arbo:matrix.org: 🤷‍♂️ 11:24:53 it used a common prng and a unix time stamp as seed, not exactly complex stuff 11:25:55 "Claude tells researchers answer was in the room with them the whole time" heheh 11:28:14 this is what i think too, the actual generation process was clearly trivial, but what's interesting to me isn't how complex the solution was, it's the fact that an AI independently found an approach that researchers hadn't successfully explored before 11:31:58 yeah I'm not even dissing Claude so much as I am being disappointed in the humans. it sounds like it could have been brute forced with a pretty simple script any time in the last decade plus 11:32:12 if anyone had had the right hunch to try 11:32:55 @ap4sh:matrix.org: independently? 11:33:04 it says there were like 7 people guiding it 11:33:37 > This work was conducted at Anthropic by John Wood, Henry de Valence, Ava Howell, Zack Voase, Lucy Sweet, Amir Omidi, Paul Crowley, Iain McGinniss, Nicholas Carlini, and Steve Weis. 11:33:45 10* 11:34:19 they're not saying that people actively guiding Claude through the discovery 11:34:40 so those 10 people wrote the report 🙃 11:35:14 we don't really know how much human interaction was involved, that's a problem on all these problems solved by openai/anthropic 11:35:23 probably guided claude into using codex 11:35:27 lmao 11:36:01 and its an internal research model, so is it really claude? 11:36:28 @ofrnxmr:xmr.mx just know that basically all devs at anthropic have stopped typing any line of code for years now 11:36:44 saw this like 1 year ago already on some interview 11:37:01 yeah, they use Actual Indians now 11:37:07 or maybe Serious Indians 11:37:13 @ofrnxmr:xmr.mx: ptolly that it's a model specifically optimized for research and quite different from the publicly available models 11:37:35 btw regarding anthropic OSS it's not even clear which model they'll use 11:37:44 maybe it's even better than what I we have rn 11:37:45 @ap4sh:matrix.org: its like elons flying car. do we even know that is a car that has seats in it? that it wsnt a $200lb shell that only contained vacuums and boosters 11:37:45 internal mythos? 11:38:00 @thomasbuilds:matrix.org: "maybe" lol 11:38:04 @ap4sh:matrix.org: no idea but they didn't say which model 11:38:14 @thomasbuilds:matrix.org: wdym? 11:38:15 its clearly not any model that they let us use 11:38:19 yeah lol 11:38:25 that's not going to be opus 4.8 11:38:30 the government doesnt use the same shit that we use 11:38:30 @ap4sh:matrix.org: like something better than opus 5.5 11:38:40 internal anthropic doesnt use the same shit we use 11:38:51 can't do any vuln research with opus 5.5 or fable without CVP+++ access for VIP++ red team+++++ 11:38:54 so yeah that's some internal anthropic model 11:38:57 they give us 5.5 when they are on 9.5 11:39:05 u can if it's unrestricted 11:39:21 question is if it's opus 5.5 unrestricted or something entirely different 11:39:38 also u can still do a lot with opus 5.5 11:40:19 (check my recent randomx work plz) 13 commits in 4 days which are all basically bug fixes. no human could do that without AI 💀 11:40:21 i think it's gonna be some mix with opus and mythos, but without restrictions and internal to anthropic. they won't use just some opus 5.5 even unrestricted imo 11:40:40 @thomasbuilds:matrix.org: not a lot in vuln research :( 11:40:52 @ap4sh:matrix.org: they never released mythos 5.5 11:41:02 @ap4sh:matrix.org: well I had no issues with randomx 11:41:12 because it was pretty secure already 11:41:13 why would they use mythos/fable when they "perform worse" than opus 5.5 atm? 11:41:21 @thomasbuilds:matrix.org: doing bug bounty for living and it's hard af to use latest anthropic or openai models to hunt 11:41:21 they have at least 6.0 done internally 11:41:28 @ofrnxmr:xmr.mx: yeah 11:41:40 youd have to be the most naive person on the planet to think that the us government lets them release the same model that the government uses 11:41:57 @ap4sh:matrix.org: ofc bug bounty is different yeah, you'd want to kyc and need all those verifications 11:42:11 @ofrnxmr:xmr.mx: ofc I didn't say that 11:42:29 my point is that internal anthropic is ahead of what they give to the government 11:42:49 yeah so anthropic OSS could give us some crazy reports 11:42:50 why would they use opus 5.5 when they ahave access to bleeding edge on dedicated compute? 11:42:51 had to do it yeah 11:42:54 https://mrelay.p2pool.observer/m/matrix.org/WIxBfhoBIFUujLWwZSOtBovB.png (image.png) 11:43:23 but even with CVP, get a lot of safeguards and going back to opus 4.8 :d so i use 5.6 Sol 11:43:29 nice, u kyc 'd? 11:43:34 yeah 11:43:45 I didn't yet lol 11:43:59 did you get unrestricted mythos? 11:44:06 even stnby doesnt have the highest level of mythos 11:44:14 @ap4sh:matrix.org: trick is to fork conversation and switch to sonnet 5.5 when doing the most sensitive 11:44:15 "highest" in quotes 11:44:40 @ofrnxmr:xmr.mx: i think only some really really big companies can get red team/offensive access 11:44:40 only got defensive access (same with openai) with the kyc etc.. 11:44:43 peasant 11:44:59 (i had that pre-typed, because i knew youd say no :P) 11:45:38 and again, id be willing to bet that the highest lvl of mythos is still below internal and government 11:45:46 way below 11:46:42 gov prob had opus 5.5 quality over a year ago 11:46:57 @ap4sh:matrix.orgare the relaxed safeguards on all models or just opus? 11:47:02 @ofrnxmr:xmr.mx: US gov for sure 11:51:16 @thomasbuilds:matrix.org: i think it's all models for a while, but as soon as it starts doing stuff like bypasses, writing PoCs, etc, it quickly downgrades to an older model, usually Opus 4.8 :/ 11:51:16 i only have defensive access, so i guess having enterprise access that's more focused on offensive security would give you access to Mythos 5.1 with fewer safeguards 11:51:16 i tried using Mythos with my current access and it instantly killed a 30 hour opus session lol 11:51:22 https://mrelay.p2pool.observer/m/matrix.org/djgVNTyzKTFSPrzGEGbmjsoA.png (image.png) 11:52:10 @ap4sh:matrix.org: mythos is fable 5.1? which is worse than opus 5.5 11:52:39 mythos is fable without the wheelchair and handcuffs 11:52:50 ok yeah well opus 5.5 is better 11:53:12 mythos is always better than fable, because fable lies to you and hides info 11:53:15 @thomasbuilds:matrix.org: maybe, i don't really know, can't really use it anyway for my use case 😔 11:54:06 while fable and mythos are supposed to be the same model, when you chop the balls off, the output suffers. mythos should still be better than opus 5.5 in every way 11:54:18 fable suffers because of the heavy guardrails 11:54:43 @ap4sh:matrix.org: funny how people think that with AI you got no work to do but basically you spend your time dealing with this above bullshit 11:55:34 @ofrnxmr:xmr.mx: ok don't know never tried mythos 11:56:04 not many people have access to it (stnby has access to a lower tier of it) 11:56:21 just saying that when you audit you want to spawn many agents and you'd kill your 200$ plan in 1 audit. 11:56:46 opus 5.5 is super goated imo 11:57:10 @thomasbuilds:matrix.org: on a consumer level 11:57:17 yeah 11:58:26 the thought of how much further ahead they are internally is the worrisome part. there needs to be some competitive open source models 11:58:35 deepseek allegedly scores similar to opus 11:58:58 and api is like 1/20th the price 11:59:03 GLM apparently is good? 11:59:10 saw this on X 12:00:12 yep! it's pretty good! actually free on opencode btw (same for deepseek, and he got updated today) 12:00:15 https://livebench.ai/#/ 12:01:05 seems wrong to me for opus 5.5 / fable 5.1 12:01:09 also glm 5.3 flash can run locally, need a pretty big hardware but you can do it 12:01:09 qwen abliterated runs very good on mine also, can ask him whatever i want and let him iterate on things, pretty conveniant 15:10:27 Boris Cherny? > <@thomasbuilds:matrix.org> saw this like 1 year ago already on some interview 15:11:08 No couldn't say interview is wrong world, Instagram reel probably 18:10:25 @ofrnxmr:xmr.mx: I think this was true like on first deepseek and opus 3.6 or whatever, but I think opus 5.5 most people think is cream of the crop rn 18:10:25 I like opus 5.5 but he burns a lot of credits, think sol 6.1 almost as good and seems like openai giving you a lot more mileage rn 18:11:15 @jwinterm:matrix.org: opus burns a lot of credits? 18:11:54 opus 5.5 on high or max thinking seems to drain my credit pool much fastger than sol 6.1 yes 18:12:06 not nearly as bad as le fable 18:12:18 but guess maybe next week I might try sonnet 6 sometimes :D 18:12:28 on what plan? 18:12:37 I have $200 per month on both 18:12:50 hmm 18:12:50 same 18:12:50 not too mention openai hands out usage resets like candy 18:12:53 opus seems almost unlimited 18:12:55 but even disregarding that 18:13:39 @monerify:matrix.org: try running couple parallel ultracodes, they will go overnight or 24h sometimes, and get a lot done, but I can use whole week in like 2-3 days 18:14:20 @jwinterm:matrix.org: what type of tasks? 18:32:33 @monerify:matrix.org: game stuff or wallet stuff primarily 18:33:20 working on making swaps work directly from smirk wallet with basicswap counterparties via nostr rn... 18:48:24 @jwinterm:matrix.org: "Code GTA 6, make no mistakes " 18:48:43 Just wait like two weeks 💀 18:49:04 Getting GTA 6 before FCMP++ will be crazy 18:49:26 @thomasbuilds:matrix.org: I meant like u can't be surprised these types of prompt drain ur entire usage 18:50:42 @jpk68:matrix.org: Nah FCMP++ will happen before GTA6 18:50:56 (the official version) 18:52:13 Highly doubt it 18:53:15 you know about the GTA6 meme lol ? 18:53:48 FCMP won't beat that meme lmao 18:55:39 Not quite sure what you mean :/ 18:56:22 we won't have fcmp before gta6 18:56:35 Yes, that's what I'm saying 18:57:55 well thats what he means too 18:58:13 now is time for you to cheer and kiss 18:58:27 @thomasbuilds:matrix.org ❤️ @jpk68:matrix.org 18:59:01 sure, I guess part of it is that new opus seems willing to work on his own for much longer time I guess, I dunno > <@thomasbuilds:matrix.org> I meant like u can't be surprised these types of prompt drain ur entire usage 18:59:01 I am not doing GTA6 type stuff but some 3d games stuff for sure, and it seems new to me that he will iterate for so long (which I guess helps eat more credits, so maybe not strictly a "pricing" thing) 19:00:33 I mean creating a 3d game is complex and lenghty task, it's fair it consumes usage credit very fast imo 19:00:46 what is super annoying is those mf safeguards 19:05:36 yea, I am not complaining, was just commenting that opus 5.5 seems to use them up faster for same tasks compared to older opuses (even if maybe that is just because he is willing to just go for longer without me typing or clicking) 19:05:59 safeguards are annoying tho...just need to get 800 GB VRAM GPU... 19:06:10 doesn't you see improvement in output though? 19:06:21 (huge improvement) 19:06:42 yeah, for 5.5 they lowered the defaul from high to medium, and then said it uses less :D > <@jwinterm:matrix.org> opus 5.5 on high or max thinking seems to drain my credit pool much fastger than sol 6.1 yes 19:06:46 yes, I think opus 5.5 does better or as good as fable 5.1 19:06:54 anways I'll stop praising claude this is monero research lounge 19:07:20 let's go in #monero-slop:monero.social for sloppy conv 🤣 21:41:54 there is clearly an overlap between AI and proof of work computing 21:42:38 The effort should be used to train a custom model, which could accelerate monero 21:44:36 Nah, decentralized competiton and centralized computations are going in different directions, 1st one is inefficient, 2nd one is super efficient 21:45:03 there is no reason not to make the work useful 21:45:29 the early adopters will have to accept that they were prototypes 21:46:25 it's purpose is to prove that you buned resource in exchange for reward, useful would mean you burned less for the same reward 21:46:44 humans cant defend the village alone, time to create a golem. 21:48:04 fvk, feels like talking with dumb AI 21:48:09 LOL 21:48:30 you've never trained a model before, have you 21:49:49 you are attacking person instead of argument, don't even try to reverse engineer my perspective, it's not needed for the sake of discussion 21:49:58 I don't even know who you are, LOL 21:50:02 context is important 21:52:31 you suggested to use decentralized computing for AI training, I said that it's inefficient, and any existing company that trains state of the art AI model will confirm it 21:53:24 free Google AI is smarter than you, LOL 21:55:14 @interestingband:matrix.org: can train model with gpu, I understand random x dosent do that - something to think about 21:55:25 @interestingband:matrix.org: debateable, claude is for sure 21:55:35 there is a new google model havent used that much yet 21:56:54 if it's your best efforts than you should do something else where critical thinking isn't needed, something like art, etc 21:59:51 😏 22:00:13 Art is really hard, communicating ideas in abstract ways. It's not always successful. Kind of like, encryption. 22:01:02 good art uses simple forms that reaches anyone, it's more about expression of something, rather thna critical thinking or analyses 22:03:57 I would argue the art is good if it successfully communicates the idea