[GUEST ACCESS MODE: Data is scrambled or limited to provide examples. Make requests using your API key to unlock full data. Check https://lunarcrush.ai/auth for authentication information.] [@AIMelGibson](/creator/twitter/AIMelGibson) "@programmer_dude @TheRealAdamG @mathemagic1an @GoogleDeepMind I don't think it is much of a reach. Very little of the training corpus for Open AI's model was from IMO. That has already been confirmed by Jerry. It explains why the formatting is more messy that and the obvious lack of Lean"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1947380149093290297) 2025-07-21 19:36:22 UTC XX followers, XX engagements "@annapanart @chatgpt21 My predication for AGI was 2029 I am looking at 2027 at the latest now. This paradigm shift is huge. This breakthrough is the biggest in the AI field so far imo. Learning in non verifiable domains has always been the ultimate challenge imo"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1946665148292251650) 2025-07-19 20:15:13 UTC XX followers, XX engagements "@_A_Flow_In_Time @brndxix How does killing a bear close up change anything Shooting it from a distance and killing it in CQC results in a dead bear both are bad outcomes. Maybe we should stop killing bears and wild life and just act a bit more civil you know like human being that live in the year 2025"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1947057799881437572) 2025-07-20 22:15:28 UTC XX followers, XXX engagements "@MinuteMovies3 @chatgpt21 An earlier checkpoint of the RL backbone. Chat GPT Agent is also using a similar backbone. It is possible to use such as backbone as a orchestrator that guides the generation. fuzzy reward RL plus mid level chunk planning is what lets Sora maintain storyline and movement"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1947316977489326213) 2025-07-21 15:25:21 UTC XX followers, XXX engagements "@inductionheads It could be the same model can generalize across distinct domains and still have elite level performance over a long task horizon. Not sure how the Open AI team have done it but this is incredible. And the fact the model was not using tools for the IMO problems"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1946520013923291366) 2025-07-19 10:38:30 UTC XX followers, XXX engagements "@programmer_dude @TheRealAdamG @mathemagic1an @GoogleDeepMind Very little is quite easy to quantify especially compared to google saying they used plenty of proofs to back up and help the model solve the problems"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1947380786308714869) 2025-07-21 19:38:54 UTC XX followers, XX engagements "@tanvitabs @btibor91 These frontier math scores alone put it in god tier territory"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1946108831609553111) 2025-07-18 07:24:37 UTC XX followers, XX engagements "@DeryaTR_ I love how all of these idiots completely gloss over the real breakthrough here. And that is having a general model that can learn and explore in a fuzzy reward space. I mean the implications here are bigger than IMO gold. As this model scales it will be elite in multiple domains"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1946908606621708697) 2025-07-20 12:22:38 UTC XX followers, XXX engagements "@EliasHasle @_everythingism @deredleritt3r @pli_cachete @curiousgangsta We are so cooked it isn't even funny at this point"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1946895287617790418) 2025-07-20 11:29:42 UTC XX followers, XX engagements "@FriesIlover49 @Angaisb_ I guess we are going to find out pretty soon think GPT 5's release and this new coding model are imminent. Hoping that Open AI are going to give us at least an early checkpoint and a taste of what this new reasoning core can do. Expect the IMO level performance months from now"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1947045116918505734) 2025-07-20 21:25:04 UTC XX followers, XX engagements "@FriesIlover49 @Angaisb_ I am XX percent sure it is using a similar reasoning core but an older checkpoint. Open AI staff have confirmed that Chat GPT Agent is also using an older checkpoint of the reasoning core"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1947042217446572333) 2025-07-20 21:13:33 UTC XX followers, XXX engagements "@natolambert @nvidia Lmarena is hot garbage. The best model is chat gpt agent by far. It is SOTA on HLE and Frontier Math"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1946115106208637352) 2025-07-18 07:49:33 UTC XX followers, XXX engagements "@VraserX @koltregaskes I imagine some sort of autoregressive audio generator would be the answer similar to what open AI did with their latest image generator. Moving away from diffusion and layering a sophisticated layering core into the architecture might be the answer. It would be expensive though"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1946876590178152835) 2025-07-20 10:15:24 UTC XX followers, XX engagements "@programmer_dude @TheRealAdamG @mathemagic1an @GoogleDeepMind Open AI most likely just have a better system than Google that is not to take away from Google's achievement it is also incredible"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1947381535667282071) 2025-07-21 19:41:53 UTC XX followers, XX engagements "@lacker @davikrehalt @littmath This does show that both google and open AI have their own cook for learning from fuzzy reward signals. Which is even more bullish for AGI in the very near future"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1946573134506676518) 2025-07-19 14:09:35 UTC XX followers, XX engagements "@MinuteMovies3 @chatgpt21 Open AIs bet imo is a reasoning first backbone + hybrid renderer that yields lower inference cost and better long horizon storytelling at the price of an extra AR refinement stage. Still significantly cheaper than Google's approach but also nets higher quality video gen"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1947322736499474651) 2025-07-21 15:48:14 UTC XX followers, XX engagements "@Angaisb_ Yes but it shows that there will be a huge push on this front. And GPT X is bound to gain from it. Not saying they will have it built into GPT X on release but I would bet my ass GPT X will be updated with this reasoning core"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1947026443063443602) 2025-07-20 20:10:52 UTC XX followers, XX engagements "@VraserX @btibor91 @swishfever GPT XXX will be AGI. Going to take a few iterations to get to a truly agentic and cross domain system that can perform at or above human level"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1946533267605184585) 2025-07-19 11:31:10 UTC XX followers, XXX engagements "@SerendipitousOr @daniel_mac8 GPT X isnt the finish line its the on ramp to rapid plug and play upgrades driven by dense reward RL. A model or series of models that can learn across domain and within a fuzzy reward space. O3 Alpha is one of these models"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1946917181981126849) 2025-07-20 12:56:42 UTC XX followers, XXX engagements "@BenPielstick So expect this kind of performance around winter. Probably looking at a November/December release. GPT X should still be releasing this summer. But it won't be close to the generality and performance of what was on display at IMO and AT coder"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1946656789564453008) 2025-07-19 19:42:00 UTC XX followers, XX engagements "@FriesIlover49 @Angaisb_ I imagine the end goal is to have one general model built on this reasoning core which can switch reward heads on the fly instead of a more full blown fine tune for each domain. I think this core could be the groundwork for a real AGI system or something close at least"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1947046791834181870) 2025-07-20 21:31:44 UTC XX followers, XX engagements "@FriesIlover49 @Angaisb_ Plus OpenAIs distillation pipeline historically cuts XX latency before release"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1947041626389381387) 2025-07-20 21:11:12 UTC XX followers, XX engagements "@MinuteMovies3 @chatgpt21 How it would work. Step X. Takes the Director LLMs shot plan and produces low res jitter tolerant keyframes (or 2fps latent video chunks). Global look & feel lighting palette coarse motion locked in here"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1947319911837020312) 2025-07-21 15:37:01 UTC XX followers, XX engagements "@giorgiottolina @basedjensen True but I think this still points towards a reasoning core that is capable of learning in a fuzzier reward domain. The only way I can imagine you could get gold on this math set. If both google and Open AI have stumbled across the same discovery that would be truly incredible"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1946593821069910306) 2025-07-19 15:31:47 UTC XX followers, XX engagements "@chatgpt21 @deredleritt3r Could potentially be the SEAL framework plus the new reasoning backbone that both of the Open AI models where using at IMO and AT Coder"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1947051968309965009) 2025-07-20 21:52:18 UTC XX followers, XXX engagements "@perceptron_elf @bindureddy This model got gold on IMO without tool use and no internet connection. More than likely we will be using this approach and tool use scaling. Plenty of gas left in LLM's"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1946988069074120918) 2025-07-20 17:38:23 UTC XX followers, XX engagements "@macosiossafari @iruletheworldmo @flowersslop This has a big model smell to it. My guess is an o3 finetune. Better RL and tool calling. If this is the OS model than it is fucking incredible because no other model comes close to this in terms of coding capabilities"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1946243685911290181) 2025-07-18 16:20:28 UTC XX followers, XXX engagements "@flavioAd @sama I mean I remember people saying that LLM's can't do math only X years ago"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1946579905216430209) 2025-07-19 14:36:29 UTC XX followers, XXX engagements "@perceptron_elf @bindureddy Did you both just completely miss the new open AI reasoning model the one that got gold on IMO and placed 2nd in an elite coding tourney That reasoning core is going to be deployed into all of the new models. Massive gains with this new approach"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1946986808425808047) 2025-07-20 17:33:23 UTC XX followers, XX engagements "@idontexist_nn @pixel_dreams01 Mr Satoshi. Is O3 Alpha running the reasoning core that we have seen in the AT Coder IMO Chat GPT Agent models My best guess is that it is using an earlier checkpoint of this reasoning core just like your Chat GPT Agent but fine tuned for code"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1946942257237565531) 2025-07-20 14:36:21 UTC XX followers, XXX engagements "@VraserX @btibor91 @swishfever I am guessing GPT X will be with us sooner than most of us think could be released within the next X weeks. Open AI staff have stated the the IMO model will take several months to make it to a public release"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1946534573610877184) 2025-07-19 11:36:21 UTC XX followers, XXX engagements "@snowmaker Maybe that is what is missing from mastering the art of the cold email"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1946663835877683580) 2025-07-19 20:10:00 UTC XX followers, XX engagements "@PromptInjection @sama @grok He could be talking about o3 alpha the new coding model that is on web arena he also could be talking about the model that recently got gold on IMO and placed 2nd in an elite XX hour long coding competition. It held 1st place for about X of them hours"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1946581357963981015) 2025-07-19 14:42:16 UTC XX followers, XXX engagements "@j4nve @kimmonismus Yes. but built on an older reasoning core. They have already completed GPT X training. This new reasoning model working as well as it has came as a surprise for everyone at Open AI. So it is going to take them sometime to cook it into GTP X. Probably a latter iteration of it"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1946595745211027747) 2025-07-19 15:39:26 UTC XX followers, XXX engagements "@Sal_Cataudella @VraserX @btibor91 @swishfever To me this is the core of any AGI system. I don't even truly think general intelligence exist in humans. But if we want a system that is better than or at human level in all domains then we need a model that can generalize across domains and this new model shows that is possible"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1946548197754568767) 2025-07-19 12:30:30 UTC XX followers, XX engagements "@EMostaque @zjasper666 @GoogleDeepMind XXX percent chance this is the last year that humans will beat AI at mathematics. When you look at what was achieved with the Open AI model it is much bigger than people are giving it credit for. This is a general model that can learn in a fuzzy reward space"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1946661184666554450) 2025-07-19 19:59:28 UTC XX followers, XXX engagements "@slow_developer @polynoamial @OpenAI Most likely will be including in an iteration of GPT X but most likely not release GPT 5"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1946538920512000055) 2025-07-19 11:53:38 UTC XX followers, XXX engagements "@flowersslop This is AGI in a sense. And once this update rolls out to a future iteration of GPT X that is when the world will change imo"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1946537613621342596) 2025-07-19 11:48:26 UTC XX followers, XX engagements "@Spiralizing @littmath It also appears that Google have their own cook for a model to learn in fuzzier reward domains. They also are going to announce a gold placement in IMO (very likely)"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1946573855075471701) 2025-07-19 14:12:27 UTC XX followers, XX engagements "@HisiDIssy @selini0 At most it is a different fine tune same base system is behind it and even the latest chat gpt agent. Although the chat gpt agent reasoning model is very old look at Jerry's thread"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1946869513443553683) 2025-07-20 09:47:17 UTC XX followers, XX engagements "@_W99ZY @mrpyo1 No shame in it but don't be stubborn. Get yourself GTA X and enjoy yourself"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1946907041030709428) 2025-07-20 12:16:25 UTC XX followers, XXX engagements "@perceptron_elf @bindureddy It is capable of learning in fuzzier reward spaces and even non verifiable domains. And is a general purpose reasoning model. An older version of it is being used for their new Chat GPT Agent the latest version will be with us in a few months probably through GPT 5"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1946987024835072022) 2025-07-20 17:34:14 UTC XX followers, XX engagements "@SIGKITTEN Who cares they are still GOATED. They have gifted us the reasoning paradigm and now models that can learn from a fuzzy reward signal. They can be as closed source as they want. Just keep delivering the sauce"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1946609063275512086) 2025-07-19 16:32:21 UTC XX followers, XXX engagements "@ChaseBrowe32432 @kalomaze Being able to learn in a fuzzy reward space and potentially non verifiable domains could lead to gigantic scaling gains. Something like this could act as a multiplier for every frontier in every research space. So yeah this is a much bigger event than people realise"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1946836613675221157) 2025-07-20 07:36:33 UTC XX followers, XX engagements "@perceptron_elf @bindureddy GPT X will be iteratively improved with this new approach"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1946987140027437097) 2025-07-20 17:34:42 UTC XX followers, XX engagements "@flowersslop GPT XXX will be the model that most of us can call AGI. GPT X will be impressive but whatever Noam and the team are working on is truly groundbreaking and I don't think it will be ready for the GPT X release. Check out his thread alot of juicy details in there"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1946537177573105948) 2025-07-19 11:46:42 UTC XX followers, XXX engagements "@MossyPathways @aidan_mclau @YouJiacheng But what we do know is that the same reasoning backbone was used across three domains computer use coding and Math. Albeit an older checkpoint for the Chat GPT Agent. Still incredibly impressive that both models are general but Open AI have the edge thus far"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1947372003704385838) 2025-07-21 19:04:00 UTC XX followers, XXX engagements "@flowersslop I think it gets us most of the way to AGI being able to work cross domain and learn in a fuzzy reward space is most of the battle. I would say this reasoning core is the biggest breakthrough in modern AI so far. You do know it is the same core that was also behind AT Coder"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1946874463854719336) 2025-07-20 10:06:58 UTC XX followers, XX engagements "@SIGKITTEN That is fine personally I don't mind being abused. They can treat me like dirt just deliver the next frontier models"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1946610766800462264) 2025-07-19 16:39:07 UTC XX followers, XX engagements "@FavidJandourek @hardmaru We have entered a new era of RL now models can learn in non verifiable domains. And they can reason with near perfect accuracy over much longer time horizons. It is a good time to be an optimist"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1946948259685011952) 2025-07-20 15:00:12 UTC XX followers, XX engagements "@axttimm @Angaisb_ Basically GPT X will ramp up to full power over the coming months. It won't be at it's full potential on release that is not to say it won't be impressive. Open AI have much bigger moat than people realise"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1947029506683777229) 2025-07-20 20:23:03 UTC XX followers, XX engagements "@programmer_dude @TheRealAdamG @mathemagic1an @GoogleDeepMind They can be made exactly verifiable but the proofs that come out of an IMO contestant (or these new models) arent in that machine checkable format. That fuzziness is why OpenAI and Google had to invent a some sort of step level reward model:"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1947381175732863226) 2025-07-21 19:40:27 UTC XX followers, XX engagements "@VraserX @btibor91 @swishfever Not to downplay GPT X it will be impressive. But I think it is going to take a bit more time to implement what we have seen in the IMO and AT coder model into GPT X. That is why I think it will be GPT XXX that truly shows us cross domain generalisation"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1946534309382246676) 2025-07-19 11:35:18 UTC XX followers, XXX engagements "@TomMcMurtryNZ @grok @theakxis @Neuralithic Yep it is apples and oranges. USAMO is a difficult math benchmark but frontier math offers questions that only genius mathematicians can solve or large groups of mathematicians over a long period of time we are talking about weeks of effort"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1946077741675119002) 2025-07-18 05:21:04 UTC XX followers, XX engagements "@philtrem22 @fchollet Open AI's model was also trained on way less IMO specific data we are talking 1-2 percent of it's corpus. It appears Open AI have created the backbone of a AGI level system. That can learn in non verifiable domains and generalize to an almost human level"  [@AIMelGibson](/creator/x/AIMelGibson) on [X](/post/tweet/1947371360340045876) 2025-07-21 19:01:27 UTC XX followers, XXX engagements
[GUEST ACCESS MODE: Data is scrambled or limited to provide examples. Make requests using your API key to unlock full data. Check https://lunarcrush.ai/auth for authentication information.]
@AIMelGibson
"@programmer_dude @TheRealAdamG @mathemagic1an @GoogleDeepMind I don't think it is much of a reach. Very little of the training corpus for Open AI's model was from IMO. That has already been confirmed by Jerry. It explains why the formatting is more messy that and the obvious lack of Lean"
@AIMelGibson on X 2025-07-21 19:36:22 UTC XX followers, XX engagements
"@annapanart @chatgpt21 My predication for AGI was 2029 I am looking at 2027 at the latest now. This paradigm shift is huge. This breakthrough is the biggest in the AI field so far imo. Learning in non verifiable domains has always been the ultimate challenge imo" @AIMelGibson on X 2025-07-19 20:15:13 UTC XX followers, XX engagements
"@_A_Flow_In_Time @brndxix How does killing a bear close up change anything Shooting it from a distance and killing it in CQC results in a dead bear both are bad outcomes. Maybe we should stop killing bears and wild life and just act a bit more civil you know like human being that live in the year 2025" @AIMelGibson on X 2025-07-20 22:15:28 UTC XX followers, XXX engagements
"@MinuteMovies3 @chatgpt21 An earlier checkpoint of the RL backbone. Chat GPT Agent is also using a similar backbone. It is possible to use such as backbone as a orchestrator that guides the generation. fuzzy reward RL plus mid level chunk planning is what lets Sora maintain storyline and movement" @AIMelGibson on X 2025-07-21 15:25:21 UTC XX followers, XXX engagements
"@inductionheads It could be the same model can generalize across distinct domains and still have elite level performance over a long task horizon. Not sure how the Open AI team have done it but this is incredible. And the fact the model was not using tools for the IMO problems" @AIMelGibson on X 2025-07-19 10:38:30 UTC XX followers, XXX engagements
"@programmer_dude @TheRealAdamG @mathemagic1an @GoogleDeepMind Very little is quite easy to quantify especially compared to google saying they used plenty of proofs to back up and help the model solve the problems" @AIMelGibson on X 2025-07-21 19:38:54 UTC XX followers, XX engagements
"@tanvitabs @btibor91 These frontier math scores alone put it in god tier territory" @AIMelGibson on X 2025-07-18 07:24:37 UTC XX followers, XX engagements
"@DeryaTR_ I love how all of these idiots completely gloss over the real breakthrough here. And that is having a general model that can learn and explore in a fuzzy reward space. I mean the implications here are bigger than IMO gold. As this model scales it will be elite in multiple domains" @AIMelGibson on X 2025-07-20 12:22:38 UTC XX followers, XXX engagements
"@EliasHasle @_everythingism @deredleritt3r @pli_cachete @curiousgangsta We are so cooked it isn't even funny at this point" @AIMelGibson on X 2025-07-20 11:29:42 UTC XX followers, XX engagements
"@FriesIlover49 @Angaisb_ I guess we are going to find out pretty soon think GPT 5's release and this new coding model are imminent. Hoping that Open AI are going to give us at least an early checkpoint and a taste of what this new reasoning core can do. Expect the IMO level performance months from now" @AIMelGibson on X 2025-07-20 21:25:04 UTC XX followers, XX engagements
"@FriesIlover49 @Angaisb_ I am XX percent sure it is using a similar reasoning core but an older checkpoint. Open AI staff have confirmed that Chat GPT Agent is also using an older checkpoint of the reasoning core" @AIMelGibson on X 2025-07-20 21:13:33 UTC XX followers, XXX engagements
"@natolambert @nvidia Lmarena is hot garbage. The best model is chat gpt agent by far. It is SOTA on HLE and Frontier Math" @AIMelGibson on X 2025-07-18 07:49:33 UTC XX followers, XXX engagements
"@VraserX @koltregaskes I imagine some sort of autoregressive audio generator would be the answer similar to what open AI did with their latest image generator. Moving away from diffusion and layering a sophisticated layering core into the architecture might be the answer. It would be expensive though" @AIMelGibson on X 2025-07-20 10:15:24 UTC XX followers, XX engagements
"@programmer_dude @TheRealAdamG @mathemagic1an @GoogleDeepMind Open AI most likely just have a better system than Google that is not to take away from Google's achievement it is also incredible" @AIMelGibson on X 2025-07-21 19:41:53 UTC XX followers, XX engagements
"@lacker @davikrehalt @littmath This does show that both google and open AI have their own cook for learning from fuzzy reward signals. Which is even more bullish for AGI in the very near future" @AIMelGibson on X 2025-07-19 14:09:35 UTC XX followers, XX engagements
"@MinuteMovies3 @chatgpt21 Open AIs bet imo is a reasoning first backbone + hybrid renderer that yields lower inference cost and better long horizon storytelling at the price of an extra AR refinement stage. Still significantly cheaper than Google's approach but also nets higher quality video gen" @AIMelGibson on X 2025-07-21 15:48:14 UTC XX followers, XX engagements
"@Angaisb_ Yes but it shows that there will be a huge push on this front. And GPT X is bound to gain from it. Not saying they will have it built into GPT X on release but I would bet my ass GPT X will be updated with this reasoning core" @AIMelGibson on X 2025-07-20 20:10:52 UTC XX followers, XX engagements
"@VraserX @btibor91 @swishfever GPT XXX will be AGI. Going to take a few iterations to get to a truly agentic and cross domain system that can perform at or above human level" @AIMelGibson on X 2025-07-19 11:31:10 UTC XX followers, XXX engagements
"@SerendipitousOr @daniel_mac8 GPT X isnt the finish line its the on ramp to rapid plug and play upgrades driven by dense reward RL. A model or series of models that can learn across domain and within a fuzzy reward space. O3 Alpha is one of these models" @AIMelGibson on X 2025-07-20 12:56:42 UTC XX followers, XXX engagements
"@BenPielstick So expect this kind of performance around winter. Probably looking at a November/December release. GPT X should still be releasing this summer. But it won't be close to the generality and performance of what was on display at IMO and AT coder" @AIMelGibson on X 2025-07-19 19:42:00 UTC XX followers, XX engagements
"@FriesIlover49 @Angaisb_ I imagine the end goal is to have one general model built on this reasoning core which can switch reward heads on the fly instead of a more full blown fine tune for each domain. I think this core could be the groundwork for a real AGI system or something close at least" @AIMelGibson on X 2025-07-20 21:31:44 UTC XX followers, XX engagements
"@FriesIlover49 @Angaisb_ Plus OpenAIs distillation pipeline historically cuts XX latency before release" @AIMelGibson on X 2025-07-20 21:11:12 UTC XX followers, XX engagements
"@MinuteMovies3 @chatgpt21 How it would work. Step X. Takes the Director LLMs shot plan and produces low res jitter tolerant keyframes (or 2fps latent video chunks). Global look & feel lighting palette coarse motion locked in here" @AIMelGibson on X 2025-07-21 15:37:01 UTC XX followers, XX engagements
"@giorgiottolina @basedjensen True but I think this still points towards a reasoning core that is capable of learning in a fuzzier reward domain. The only way I can imagine you could get gold on this math set. If both google and Open AI have stumbled across the same discovery that would be truly incredible" @AIMelGibson on X 2025-07-19 15:31:47 UTC XX followers, XX engagements
"@chatgpt21 @deredleritt3r Could potentially be the SEAL framework plus the new reasoning backbone that both of the Open AI models where using at IMO and AT Coder" @AIMelGibson on X 2025-07-20 21:52:18 UTC XX followers, XXX engagements
"@perceptron_elf @bindureddy This model got gold on IMO without tool use and no internet connection. More than likely we will be using this approach and tool use scaling. Plenty of gas left in LLM's" @AIMelGibson on X 2025-07-20 17:38:23 UTC XX followers, XX engagements
"@macosiossafari @iruletheworldmo @flowersslop This has a big model smell to it. My guess is an o3 finetune. Better RL and tool calling. If this is the OS model than it is fucking incredible because no other model comes close to this in terms of coding capabilities" @AIMelGibson on X 2025-07-18 16:20:28 UTC XX followers, XXX engagements
"@flavioAd @sama I mean I remember people saying that LLM's can't do math only X years ago" @AIMelGibson on X 2025-07-19 14:36:29 UTC XX followers, XXX engagements
"@perceptron_elf @bindureddy Did you both just completely miss the new open AI reasoning model the one that got gold on IMO and placed 2nd in an elite coding tourney That reasoning core is going to be deployed into all of the new models. Massive gains with this new approach" @AIMelGibson on X 2025-07-20 17:33:23 UTC XX followers, XX engagements
"@idontexist_nn @pixel_dreams01 Mr Satoshi. Is O3 Alpha running the reasoning core that we have seen in the AT Coder IMO Chat GPT Agent models My best guess is that it is using an earlier checkpoint of this reasoning core just like your Chat GPT Agent but fine tuned for code" @AIMelGibson on X 2025-07-20 14:36:21 UTC XX followers, XXX engagements
"@VraserX @btibor91 @swishfever I am guessing GPT X will be with us sooner than most of us think could be released within the next X weeks. Open AI staff have stated the the IMO model will take several months to make it to a public release" @AIMelGibson on X 2025-07-19 11:36:21 UTC XX followers, XXX engagements
"@snowmaker Maybe that is what is missing from mastering the art of the cold email" @AIMelGibson on X 2025-07-19 20:10:00 UTC XX followers, XX engagements
"@PromptInjection @sama @grok He could be talking about o3 alpha the new coding model that is on web arena he also could be talking about the model that recently got gold on IMO and placed 2nd in an elite XX hour long coding competition. It held 1st place for about X of them hours" @AIMelGibson on X 2025-07-19 14:42:16 UTC XX followers, XXX engagements
"@j4nve @kimmonismus Yes. but built on an older reasoning core. They have already completed GPT X training. This new reasoning model working as well as it has came as a surprise for everyone at Open AI. So it is going to take them sometime to cook it into GTP X. Probably a latter iteration of it" @AIMelGibson on X 2025-07-19 15:39:26 UTC XX followers, XXX engagements
"@Sal_Cataudella @VraserX @btibor91 @swishfever To me this is the core of any AGI system. I don't even truly think general intelligence exist in humans. But if we want a system that is better than or at human level in all domains then we need a model that can generalize across domains and this new model shows that is possible" @AIMelGibson on X 2025-07-19 12:30:30 UTC XX followers, XX engagements
"@EMostaque @zjasper666 @GoogleDeepMind XXX percent chance this is the last year that humans will beat AI at mathematics. When you look at what was achieved with the Open AI model it is much bigger than people are giving it credit for. This is a general model that can learn in a fuzzy reward space" @AIMelGibson on X 2025-07-19 19:59:28 UTC XX followers, XXX engagements
"@slow_developer @polynoamial @OpenAI Most likely will be including in an iteration of GPT X but most likely not release GPT 5" @AIMelGibson on X 2025-07-19 11:53:38 UTC XX followers, XXX engagements
"@flowersslop This is AGI in a sense. And once this update rolls out to a future iteration of GPT X that is when the world will change imo" @AIMelGibson on X 2025-07-19 11:48:26 UTC XX followers, XX engagements
"@Spiralizing @littmath It also appears that Google have their own cook for a model to learn in fuzzier reward domains. They also are going to announce a gold placement in IMO (very likely)" @AIMelGibson on X 2025-07-19 14:12:27 UTC XX followers, XX engagements
"@HisiDIssy @selini0 At most it is a different fine tune same base system is behind it and even the latest chat gpt agent. Although the chat gpt agent reasoning model is very old look at Jerry's thread" @AIMelGibson on X 2025-07-20 09:47:17 UTC XX followers, XX engagements
"@_W99ZY @mrpyo1 No shame in it but don't be stubborn. Get yourself GTA X and enjoy yourself" @AIMelGibson on X 2025-07-20 12:16:25 UTC XX followers, XXX engagements
"@perceptron_elf @bindureddy It is capable of learning in fuzzier reward spaces and even non verifiable domains. And is a general purpose reasoning model. An older version of it is being used for their new Chat GPT Agent the latest version will be with us in a few months probably through GPT 5" @AIMelGibson on X 2025-07-20 17:34:14 UTC XX followers, XX engagements
"@SIGKITTEN Who cares they are still GOATED. They have gifted us the reasoning paradigm and now models that can learn from a fuzzy reward signal. They can be as closed source as they want. Just keep delivering the sauce" @AIMelGibson on X 2025-07-19 16:32:21 UTC XX followers, XXX engagements
"@ChaseBrowe32432 @kalomaze Being able to learn in a fuzzy reward space and potentially non verifiable domains could lead to gigantic scaling gains. Something like this could act as a multiplier for every frontier in every research space. So yeah this is a much bigger event than people realise" @AIMelGibson on X 2025-07-20 07:36:33 UTC XX followers, XX engagements
"@perceptron_elf @bindureddy GPT X will be iteratively improved with this new approach" @AIMelGibson on X 2025-07-20 17:34:42 UTC XX followers, XX engagements
"@flowersslop GPT XXX will be the model that most of us can call AGI. GPT X will be impressive but whatever Noam and the team are working on is truly groundbreaking and I don't think it will be ready for the GPT X release. Check out his thread alot of juicy details in there" @AIMelGibson on X 2025-07-19 11:46:42 UTC XX followers, XXX engagements
"@MossyPathways @aidan_mclau @YouJiacheng But what we do know is that the same reasoning backbone was used across three domains computer use coding and Math. Albeit an older checkpoint for the Chat GPT Agent. Still incredibly impressive that both models are general but Open AI have the edge thus far" @AIMelGibson on X 2025-07-21 19:04:00 UTC XX followers, XXX engagements
"@flowersslop I think it gets us most of the way to AGI being able to work cross domain and learn in a fuzzy reward space is most of the battle. I would say this reasoning core is the biggest breakthrough in modern AI so far. You do know it is the same core that was also behind AT Coder" @AIMelGibson on X 2025-07-20 10:06:58 UTC XX followers, XX engagements
"@SIGKITTEN That is fine personally I don't mind being abused. They can treat me like dirt just deliver the next frontier models" @AIMelGibson on X 2025-07-19 16:39:07 UTC XX followers, XX engagements
"@FavidJandourek @hardmaru We have entered a new era of RL now models can learn in non verifiable domains. And they can reason with near perfect accuracy over much longer time horizons. It is a good time to be an optimist" @AIMelGibson on X 2025-07-20 15:00:12 UTC XX followers, XX engagements
"@axttimm @Angaisb_ Basically GPT X will ramp up to full power over the coming months. It won't be at it's full potential on release that is not to say it won't be impressive. Open AI have much bigger moat than people realise" @AIMelGibson on X 2025-07-20 20:23:03 UTC XX followers, XX engagements
"@programmer_dude @TheRealAdamG @mathemagic1an @GoogleDeepMind They can be made exactly verifiable but the proofs that come out of an IMO contestant (or these new models) arent in that machine checkable format. That fuzziness is why OpenAI and Google had to invent a some sort of step level reward model:" @AIMelGibson on X 2025-07-21 19:40:27 UTC XX followers, XX engagements
"@VraserX @btibor91 @swishfever Not to downplay GPT X it will be impressive. But I think it is going to take a bit more time to implement what we have seen in the IMO and AT coder model into GPT X. That is why I think it will be GPT XXX that truly shows us cross domain generalisation" @AIMelGibson on X 2025-07-19 11:35:18 UTC XX followers, XXX engagements
"@TomMcMurtryNZ @grok @theakxis @Neuralithic Yep it is apples and oranges. USAMO is a difficult math benchmark but frontier math offers questions that only genius mathematicians can solve or large groups of mathematicians over a long period of time we are talking about weeks of effort" @AIMelGibson on X 2025-07-18 05:21:04 UTC XX followers, XX engagements
"@philtrem22 @fchollet Open AI's model was also trained on way less IMO specific data we are talking 1-2 percent of it's corpus. It appears Open AI have created the backbone of a AGI level system. That can learn in non verifiable domains and generalize to an almost human level" @AIMelGibson on X 2025-07-21 19:01:27 UTC XX followers, XXX engagements