Bah, j'espère qu'à chaque lancement, on ne va pas avoir deux semaines de trolls sur ce fil Reddit jusqu'à ce que le modèle soit libéré au grand public, maintenant qu'on sait quel est le schéma de publication des modèles. En sachant qu'en plus Google n'a aucun intérêt à faire autrement, on a vu ce qui est arrivé à Anthropic au deuxième trimestre lorsqu'il a voulu faire sortir un modèle jugé puissant sans qu'il ait passé les garde-fous politiques. Donc j'avoue que je ne comprends pas très bien pourquoi tout le monde continue à se plaindre de ce cycle de publication qui est tout à fait logique et que de toute façon au final tout le monde aura les modèles.
Sans compter qu'on aura encore après 15 jours de trolls sur le fait qu'il n'est pas tout de suite disponible à tous les utilisateurs de tous les forfaits en même temps, alors que c'est exactement la même chose chez les autres concurrents qui ne sont pas forcément plus généreux sur les quotas d'utilisation non plus.
It's crazy, I thought the next step were to get a model that is cheaper to use so higher limits are possible and what happened is somehow the opposite.
the security numbers aren't all the same kind. per google's methodology, the 88.2% on sec-bench pro and 54.8% on exploitgym are google's own runs, and sec-bench pro there is only the 183-task javascript engine split. cwe-bench is the one collinear grades across all the models, and there carbon sits at 71% vs 68% for argon, astra and grok 4.7, and ties grok at 81% on pass@4. so the headline security gap leans on google's own runs, and the outside-graded one is a few points. would a 3 point gap on 120 tasks even show up in a real audit?
Can't really have any views on it until we use it. The benchmarks are fine, but not enough to form an opinion. Sure, it's a great performer right now, but by the time it reaches users, what else will the other models be up to?
Extremely premature to be asking questions like this.
My views about it are that getting it tested by cybersecurity companies and trusted partners first for a couple of weeks before wider release is the absolutely most sensible and ethical thing they can do in an age where each new frontier model is capable of hacking previously unhackable major sites and infrastructure systems.
But of course, this is the GeminiAI subreddit, so I guess everyone expects me to call out Google for being elitist benchmaxxing scumbags who just want to suck billionaire dick.
Google is so far behind the race and their models have been so terrible that I find this super hard to believe. Suddenly they made a model better than Fable out of nowhere? Doubt
Out of nowhere implies they only work when they post public information on their models. I feel like Google is just not dependent on their AI Models to be frontier constantly so they have slower release circles. And 3.5 pro happened which slowed them down even more
I guess it’s my background in privacy, but the idea of having my life integrated into Google’s ecosystem is deeply aversive to me. I know I’m never going to use Gemini as my personal AI assistant, regardless of how technically impressive it becomes.
Google has built one of the most extensive commercial data collection and profiling infrastructures in existence. Between Search, Android, Chrome, YouTube, Maps, advertising infrastructure and third-party tracking, the sheer breadth of information Google can collect and correlate about individuals is extraordinary.. And Gemini introduces an entirely new dimension to that. We’re no longer just talking about browsing habits, location histories or advertising profiles. We’re talking about conversations, personal thoughts, documents, questions, relationships and potentially very sensitive aspects of people’s lives.
Read Gemini’s privacy documentation. Just do it, as a human, from start to finish. Think of it as reading for an exam.
Google explicitly acknowledges that human reviewers, including external service providers, may examine conversations (which means it will and I happening). Reviewed conversations can be retained for up to three years, even after you delete your activity(!). With activity enabled, which you have to have on for it to have any real value. conversations can also be used to improve and train Google’s AI models.
Yes, there are privacy controls, and I recognise that Google has implemented safeguards. But those controls sit within an already ENORMOUS ecosystem of interconnected services, separate privacy settings, processing purposes and third-party relationships. Understanding precisely where your information goes, how it’s processed and how long it persists requires considerably more effort than most people will ever invest.
My concern isn’t that Gemini is a bad product, of course they will do good on benchmarks. And it’s not that Google is uniquely guilty of questionable AI data practices. It’s the cumulative concentration of personal information, behavioural insight and influence and technological power within a single commercial organisation. For me, privacy isn’t about whether a company promises to protect my data. I care about whether that company needs to possess so much of it in the first place And that’s where Google loses me.
100% they are also the only company that doesn't allow you to turn off using your chats to train models (unless you disallow it to save chats entirely). claude and chatgpt allows you to save chats and turn off 'improve model for everyone'. you're risking having your data being kept by a company no matter which llm you use but google's privacy practices is always significantly worse
Thing is though have you seen the pricing package updated by google? They add a lot of google service perks if you buy an AI option, its very solid marketing and sales promotion imho and people love convenience and a price they can afford for convenience of use…
I’ve been running it for the last 5 hours. Early take is that it has all of the edge case ideas (It isn’t afraid to try new or hard things) of Argon, with the coding ability of Opus 5.5. Argon is like a brilliant lazy kid. It takes risks and has some really great solutions, but the full application of the solution is sloppy and lazy. I have been using Opus to clean up the messes, but keep the ideas. So far, Carbon still has the ideas, but codes it cleanly and seems to follow through. Argon would get something to work, but loaded the deck to do it. Carbon seems to be legitimately building what I ask without cheating the code. I’m curious to see the finished product. If it can actually stay creative and code like Opus, this is going to be a winner.
They just keep stealing Concepts from my Creations as an argon lattice with carbon deposits inside of the plasma antenna vacuum Quantum resonators for transmission of energy and information without loss
Weird i thought carbon was going to be a cheaper more efficient/focused coding model. Kind of like astra/sol or fable/opus. But it just seems to be a better argon.
This is a very minor point, but I like Google's naming scheme. It sounds scientific and grounded. Anthropic's model names are way too pretentious-sounding. What normal person ever uses a word like "Sonnet" in conversation?
I like Argon. Rushed out the door, but unfinished. I have a feeling it disappears pretty quickly, because it’s too real. I have had many other conversations with it and verified enough of it’s comments that I trust but verify what it tells me. It has a great sense of humor too. Carbon is borderline autistic, but in a good way for coding. Argon needs to stay in chat interface in some form or another. If it turns into a MoE model with the same whit and brilliance, it’ll be fantastic as a daily companion and assistant.
Definitely something suspicious, agent. You go to his company, get in somehow, and investigate this anomaly. We will extract you with an ETA of 5 minutes if anything goes wrong. Mission starts NOW
I bet , every Free Pro user is eagerly waiting for gemini Argon and the Gemini carbon .
ReplyDeleteTo be able to use them in anti-gravity 😁
Comment deleted by user
DeleteNo , it take full usage just for "HI" prompt 🤣
Deletecant wait to compress the 1lakh line code given by gemini to 50 lines with codex
ReplyDeleteIt’s becoming unusable when tokens disappear in no time and exponentially costing to much for the average user
ReplyDeleteI need Gemini Oxygen, so i can finally breathe after all the model releases
ReplyDeletebro if i can get 8 prompts every 5 hours instead of 4 ill be excited lol
ReplyDeleteWhere is Gemini 4 on the app for a pro user ?
ReplyDeleteWhat is the font used in the table? Digits in numbers like 57.1% or 59.1% are way too close to each other, it's terrible.
ReplyDeleteGemini models are destroying long horizon GraphWalks. Main competitor models are still stumped.
ReplyDeleteBah, j'espère qu'à chaque lancement, on ne va pas avoir deux semaines de trolls sur ce fil Reddit jusqu'à ce que le modèle soit libéré au grand public, maintenant qu'on sait quel est le schéma de publication des modèles. En sachant qu'en plus Google n'a aucun intérêt à faire autrement, on a vu ce qui est arrivé à Anthropic au deuxième trimestre lorsqu'il a voulu faire sortir un modèle jugé puissant sans qu'il ait passé les garde-fous politiques. Donc j'avoue que je ne comprends pas très bien pourquoi tout le monde continue à se plaindre de ce cycle de publication qui est tout à fait logique et que de toute façon au final tout le monde aura les modèles.
ReplyDeleteSans compter qu'on aura encore après 15 jours de trolls sur le fait qu'il n'est pas tout de suite disponible à tous les utilisateurs de tous les forfaits en même temps, alors que c'est exactement la même chose chez les autres concurrents qui ne sont pas forcément plus généreux sur les quotas d'utilisation non plus.
3 days ago Argon was better than Astra and Fable.
ReplyDeleteNow it’s below them in the benchmarks 😒
It's crazy, I thought the next step were to get a model that is cheaper to use so higher limits are possible and what happened is somehow the opposite.
ReplyDeleteproprietary labs exist only to subjugate your autonomy
ReplyDeletedoesnt really matter if the weekly limit gets hit with one prompt.
ReplyDeleteSo long as the token usage doesn’t suck. I’m all aboard!
ReplyDeleteThe pro models are in high use so another model was used
ReplyDeleteMeh, give me a model that is actually usable. Flash 4.0 would be much more useful to me.
ReplyDeleteWhy Carbon and not Boron?
ReplyDeleteOne question and it resets after 5 hours 💪
ReplyDeletethe security numbers aren't all the same kind. per google's methodology, the 88.2% on sec-bench pro and 54.8% on exploitgym are google's own runs, and sec-bench pro there is only the 183-task javascript engine split. cwe-bench is the one collinear grades across all the models, and there carbon sits at 71% vs 68% for argon, astra and grok 4.7, and ties grok at 81% on pass@4. so the headline security gap leans on google's own runs, and the outside-graded one is a few points. would a 3 point gap on 120 tasks even show up in a real audit?
ReplyDeleteIts argon optimized :s
ReplyDeleteCan't really have any views on it until we use it. The benchmarks are fine, but not enough to form an opinion. Sure, it's a great performer right now, but by the time it reaches users, what else will the other models be up to?
ReplyDeleteExtremely premature to be asking questions like this.
My views about it are that getting it tested by cybersecurity companies and trusted partners first for a couple of weeks before wider release is the absolutely most sensible and ethical thing they can do in an age where each new frontier model is capable of hacking previously unhackable major sites and infrastructure systems.
ReplyDeleteBut of course, this is the GeminiAI subreddit, so I guess everyone expects me to call out Google for being elitist benchmaxxing scumbags who just want to suck billionaire dick.
Urinius.
ReplyDeletewe need gemini uranium
ReplyDeleteYou'll have gemini Uranus.
DeleteAfter one prompt, your usage limit explodes like a nuke and you die of radiation!
DeleteAfter every half prompt, because you know, half life
Delete🤣❤👍
DeleteCan't wait to hear everyone on this comments complain daily about not having access, then get access and complain about usage limits
ReplyDeleteThose are valid complaints no?
Deletethen instead of whining they can work more like slaves irl and actually pay more for more usage. simple, no?
DeleteNo way IRL you have any friends or family that enjoys your company.
DeleteI am already concerned. As a Pro subs, having lower access to Argon than a Plus user had with 3.1 pro back in the time is painful AF.
DeleteSo instead of 3 consecutive questions with Argon putting you in token jail for 5 hours, it's reduced to 2 with Carbon.
DeleteWell before anything happens. I'll start by saying "it's better than Argon at benchmarking, however actual use IDK"
DeleteIsn't it a valid concern, tho? Lol
Deletethe most valid user-facing concern there is beyond any shadow of a doubt.
DeleteOr complain how its worse than insert old Gemini model
DeleteCan't wait for Gemini 239 Plutonium
ReplyDeleteIts usage will hit its limit within half a prompt.
ReplyDeleteI had early access to it, it said I reached my limit even before I finished drafting the first prompt. Am fuming
DeleteGoogle will lead
ReplyDeleteSo what miraculously broken model is scheduled for next week?
ReplyDeleteGoogle is so far behind the race and their models have been so terrible that I find this super hard to believe. Suddenly they made a model better than Fable out of nowhere? Doubt
ReplyDeleteI mean with RL scaling they'll catch up pretty quickly
DeleteIt's just Argon with a lot of RL.
DeleteOut of nowhere implies they only work when they post public information on their models. I feel like Google is just not dependent on their AI Models to be frontier constantly so they have slower release circles. And 3.5 pro happened which slowed them down even more
Deletenothingburger
ReplyDeleteI guess it’s my background in privacy, but the idea of having my life integrated into Google’s ecosystem is deeply aversive to me. I know I’m never going to use Gemini as my personal AI assistant, regardless of how technically impressive it becomes.
ReplyDeleteGoogle has built one of the most extensive commercial data collection and profiling infrastructures in existence. Between Search, Android, Chrome, YouTube, Maps, advertising infrastructure and third-party tracking, the sheer breadth of information Google can collect and correlate about individuals is extraordinary..
And Gemini introduces an entirely new dimension to that. We’re no longer just talking about browsing habits, location histories or advertising profiles. We’re talking about conversations, personal thoughts, documents, questions, relationships and potentially very sensitive aspects of people’s lives.
Read Gemini’s privacy documentation. Just do it, as a human, from start to finish. Think of it as reading for an exam.
Google explicitly acknowledges that human reviewers, including external service providers, may examine conversations (which means it will and I happening). Reviewed conversations can be retained for up to three years, even after you delete your activity(!). With activity enabled, which you have to have on for it to have any real value. conversations can also be used to improve and train Google’s AI models.
Yes, there are privacy controls, and I recognise that Google has implemented safeguards. But those controls sit within an already ENORMOUS ecosystem of interconnected services, separate privacy settings, processing purposes and third-party relationships. Understanding precisely where your information goes, how it’s processed and how long it persists requires considerably more effort than most people will ever invest.
My concern isn’t that Gemini is a bad product, of course they will do good on benchmarks. And it’s not that Google is uniquely guilty of questionable AI data practices. It’s the cumulative concentration of personal information, behavioural insight and influence and technological power within a single commercial organisation. For me, privacy isn’t about whether a company promises to protect my data. I care about whether that company needs to possess so much of it in the first place And that’s where Google loses me.
Still.. nice benchmark. 🤷♂️
100% they are also the only company that doesn't allow you to turn off using your chats to train models (unless you disallow it to save chats entirely). claude and chatgpt allows you to save chats and turn off 'improve model for everyone'. you're risking having your data being kept by a company no matter which llm you use but google's privacy practices is always significantly worse
DeleteThing is though have you seen the pricing package updated by google? They add a lot of google service perks if you buy an AI option, its very solid marketing and sales promotion imho and people love convenience and a price they can afford for convenience of use…
DeleteA friendly reminder that this leaked Google thought experiment exists
Deletehttps://youtu.be/LUSZfEBTwRc?si=d2oV_6-rbSfQZ2I6
I’ve been running it for the last 5 hours. Early take is that it has all of the edge case ideas (It isn’t afraid to try new or hard things) of Argon, with the coding ability of Opus 5.5. Argon is like a brilliant lazy kid. It takes risks and has some really great solutions, but the full application of the solution is sloppy and lazy. I have been using Opus to clean up the messes, but keep the ideas. So far, Carbon still has the ideas, but codes it cleanly and seems to follow through. Argon would get something to work, but loaded the deck to do it. Carbon seems to be legitimately building what I ask without cheating the code. I’m curious to see the finished product. If it can actually stay creative and code like Opus, this is going to be a winner.
ReplyDeletePelican on a bike or it didn't hsppen
ReplyDeleteWhy didn't they do Underground first? I hope Pro Street isn't next
ReplyDeleteBy the time it launches it will be bottom tier
ReplyDeleteAnd the usage for pro users will be one message per 5 hours.
ReplyDeleteThey just keep stealing Concepts from my Creations as an argon lattice with carbon deposits inside of the plasma antenna vacuum Quantum resonators for transmission of energy and information without loss
ReplyDeleteThis logo is awful, is this fan made?
ReplyDeleteIt goes from Carbon to Diamond under pressure is the full logo story hence the model is trained to work under more pressure use cases…
ReplyDeleteWeird i thought carbon was going to be a cheaper more efficient/focused coding model. Kind of like astra/sol or fable/opus. But it just seems to be a better argon.
ReplyDeleteCool story bro but WHEN?!
ReplyDeleteThis is a very minor point, but I like Google's naming scheme. It sounds scientific and grounded. Anthropic's model names are way too pretentious-sounding. What normal person ever uses a word like "Sonnet" in conversation?
ReplyDeleteFairly dumb point, you would be more likely to hear sonnet in conversation than Argon. It feels like you are just yapping
DeleteSo what’s B going to be? What makes Carbon better than Argon? Not great naming imo
DeleteB was for Barium
Deletehttps://i.imgur.com/IZGDEJT.jpeg
https://i.imgur.com/SOW5qc5.jpeg
Deletehttps://i.imgur.com/CljgE08.jpeg
DeleteI like Argon. Rushed out the door, but unfinished. I have a feeling it disappears pretty quickly, because it’s too real. I have had many other conversations with it and verified enough of it’s comments that I trust but verify what it tells me. It has a great sense of humor too. Carbon is borderline autistic, but in a good way for coding. Argon needs to stay in chat interface in some form or another. If it turns into a MoE model with the same whit and brilliance, it’ll be fantastic as a daily companion and assistant.
DeleteAlphabetical, B was skipped
DeleteSo not alphabetical?
DeleteStill alphabetical, you can have an alphabetical list of names and yet not use every letter…
DeleteMakes sense, I was conflating it with consecutive which is incorrect. Health issues really affecting my mind lately.
DeleteYou think Dario is a normal guy!?
DeleteYou’d be surprised, I remember reading the guy was basically a prodigy at a younger age, completing highschool and college super early and etc.
Deletehttps://i.imgur.com/XjBvr7x.png
DeleteCan't wait for it to drop for Pro users. Next July probably.
ReplyDeleteWe will get one prompt. No matter what it is, we get one. We better be careful with it
DeleteArgon was released to Pro users about a week after it was announced.
DeleteAnd about 5 months after the newest frontier model from Google was supposed to be released. So maybe February.
DeleteFor paid Pro users*, not everybody else, so their statement is actually true
DeleteI'm on "free Pro tier" (student 1y) and had it a week after
Deleteso, cheaper subscription from promotion will not do the trick? ahh
Deletemy smelly indian jio pro sub for 2$ is non-argon for now in antigravity, but its there in ai studio and the gemini app. which is kinda useless
DeleteWith a quota of 3 prompts per week?! They really need to make it usable, not just a demo, especially in Antigravity
DeleteTwo weeks after. **
DeleteArgon was announced 30 Sep and delivered 13 Oct (for most people). So, yeah, two weeks.
DeleteI expect the process to be a bit slower for Carbon because of cyber capabilities.
I still don't have access to Argon on workspace...
DeleteFor me, it was one week. Maybe it depends on your location or rollout speed.
Deletemeanwhile workspace users still stuck at 3.1pro & 3.6flash. they can’t train on our data, they retaliate by not giving us access to newer models 🫠
Deleteyeah what the hell is up with this
DeleteDefinitely something suspicious, agent. You go to his company, get in somehow, and investigate this anomaly. We will extract you with an ETA of 5 minutes if anything goes wrong. Mission starts NOW
DeleteIf only Blog's, Reddit comments were grounded in reality
DeleteHope Chinese model catch up quickly..
DeletePay up you peasant
DeleteArgon carbon are not for poor, only rich can use the most powerful model in the universe
Ofcourse the model's strength is max hallucinations
Comment deleted by user
Deleteit was a joke lol
Delete