how do they do it? and don't give me the "they collect your data" crap, every provider collects everything they possibly can
>>108883900Probably with access to government backed investments to be able to sell at a loss while the winners and losers of this market are determined to ensure that China has at least one champion in that particular industry.
>>108883900They train on chatgpt traces, it's effectively a very efficient distillation method, but totally worthless in terms of advancing things.
>>108883900China can afford it, plus you are training their codebase.
>>108883900By losing customers. Where do you think chatgpt and anthropic's increasing customers are coming from?
>>108883900I'm very seriously considering putting $10 into deepsneed and trying out this coding agent shit I keep hearing about, now that the API price is essentially guaranteed to be dirt fucking cheap
>>108884180the model itself isn't permanent
>>108884170Chatgpt userbase has shrunk considerably, as has gemini's. Only anthropic's is growing, mostly at the expense of the other two. Actually chinese models, including deepseek, have been gaining for the same reasons, but regulatory hurdles have made them less favored.
>>108884190what do you even mean by thatthat they will replace v4 with v4.x and v5 later?
>>108883900Chinese companies optimize inference while Americans just spam data centers.
>>108883900The USA is over. It's finished.
>>108883900The chinks are just trying to rekt the Western Ai market that the juice on the news are always hyping up (OH SHIT GOY, MYTHIC IS SOO POWERFUL WE CANT EVEN SHOW YOU, HURRY AND INVEST GORILLOONS)They know if they hand out shit practically for free there will be no real Ai market
>>108884209>>108884157>>108884114sooooooo what we learnt is that china focuses on whats important, having a metric shit ton of electricity (making it cheap as dirt) and optimizing inference so you need less power (lower cost) and less data centers (lower cost)intredasting... why cant the US or the EU do the same?
>China overwhelms the US with CHEAP decent/good cars>Americans buy cheap china cars>US car market implodes and car makers stop making cars in the US>The US loses car makers>China now has an easy way in ass the US lost a ton of car makers which could be turned into tank makers and moreNow they do the same with AI, outpricing the west and their expensive trash
>>108883900is it significantlly shittier than claude? if not then I'll probably buy it
>>108884228>Why can't the-Stop pretending to be retarded. Of the western world elite offered the plebs an abundance of energy they would risk being over thrown.They give us the illusion of freedom by offering us do what we want but then telling us that all that freedom is pay walled and you have to work harder for it.The chinks on the other hand just use the population as outright, immediate slave labor for their ruling class and the benefit of the western world.If we didn't have them chink insect slaves, we wouldn't have affordable anything them we'd attack our employers for living wages and all hell would break loose.Thank you chinks for dying on liveleak for our affordable living.
>>108884246Tech illiterates like yourselves (why the fuck are you allowed on /g/ anyway) have a hard time differentiating between the harness used and the actual model. Deepseek series of models is generally only second to, and sometimes superior to, openai series of models of a corresponding level. Anthropic is second but quite a bit behind. Gemini is a total joke, mistral is even worse. Grok even lower than that. There is nothing else.However, anthropic's harness is far beyond what anyone else has at the moment and, obviously, they have the opportunity to optimize claude for their harness and vice versa.
>>108884246it's on par with "frontier" model actually.I've even built shitty games/apps/websites/extensions with deepseek flash ,assuming you're ready for the compile&correct&repeat.flash is free so who cares . i've been on the thread being like "lol" when i saw everyone screeching about their daily limits.meanwhile deepseek goes brrrrrrrr. don't know about complex projects/codebase so.
>>108884273>Tech illiterates like yourselves (why the fuck are you allowed on /g/ anyway) have a hard time differentiating between the harness used and the actual model. Deepseek series of models is generally only second to, and sometimes superior to, openai series of models of a corresponding level. Anthropic is second but quite a bit behind. Gemini is a total joke, mistral is even worse. Grok even lower than that. There is nothing else.owever, anthropic's harness is far beyond what anyone else has at the moment and, obviously, they have the opportunity to optimize claude for their harness and vice versaIs this new pasta?
>>108884228Efficiency runs counter to American goals, they need to maximize wasteful spending for the middlemen to profit.Euros are too busy trying to figure how to regulate it.
>>108883900Chinese as a language will offer a vector space thats a magnitude smaller than english. Since LLMs are token predictors shitting out responses based on similarity score criteria ,a denser space will naturally offer superior performance. On the english side of things, even if we train our models to condense to caveman speak or chinese or neuralese, theres always a heftier translation tax to pay when going to or from english compared to chinese.The math is already in their favor, and thats before we start talking planned economies or anything like that.
>>108884246From their release whitepaper>In our internal evaluation, DeepSeek-V4-Pro-Max outperforms Claude Sonnet 4.5 and approaches the level of Opus 4.5.So yeah, they're far behind. >>108884273The models are behind and the agentic capabilities and harness are even further behind. They're basically unusable.
>>108883900Extremely fucking based
>>108884273yeah claude code is just so good that they had to go banning people that used anything else
>>108884194OpenAI hasn't shrunk at all. They're both growing constantly, Anthropic just took a lead in their growth rate. So far I haven't seen any actual evidence that chinese models are seeing any growth. What has seen growth is AI-tool enabled cloud hosting on chinese platforms but that's adjacent, especially since CPC is just throwing money at them.
>>108884228Because after WW2, western leadership was compromised by khazarian mafia. Their only goal is DEATH of west aka amalek, because Hadrian raped them 2000 years ago.
>>108884377fake news.anthropic doesn't care what you use their api for.
>>108884180till a month ago, US models all had even cheaper subscriptions where you never had to worry about any costs per token
>>108884387>revenue
>>108884492Yeah but paying montly subscription is for kikes. If I want to take a 4 month break from coding and I don't use the AI service, I don't want to keep paying for it.Pay per token (especially at the low prices Deepseek offers) is fair and honest. Monthly subscription is a Jewish scam.
>>108884180you can try it out for free in opencode, they have free small models
>>108884510makes no sense.if you take a break for 4 months, you cancel your monthly subscription for those 4 months.
>>108884522>oy vey whats the problem just remember to cancel your subscription>also, you are locked in to our discontinued $15.99 monthly plan for your loyalty, goy, if you cancel and resubscribe we'll change you $29.99 instead!bro there's no fucking way i'm paying $5.99 a month for like 500k tokens when you can just pay $1 per 1 million tokens with DS. 1 million tokens is like the entire lord of the rings or something ridiculous like that. I won't even use that much in a month.Americans' acceptance of subscription based billing is why you end up with people spending $70 a month on netflix, hulu, amazon, disney, chatgpt, onlyfans and all this other money sucking shit>inb4 pooryes
>>108884534>1 million tokens is like the entire lord of the rings or something ridiculous like that.you have no idea what you are talking about if you think 1 millions tokens is much.
>>108883900>>108884114I always find it funny the myopic morons keep getting proven wrong time and time again yet the continue up with there luddite dog shit. There is a fucking reaso china is doing this because it know AI isn't a "bubble" just like every other country with means to do something. The anti-AI fuckheads will just silent move on to some other dumb shit.
>>108884608>its not a bubbleLMAO
>>108884608It's not that deep, it's simply another way to export more, otherwise they'd have to enrich their own people to the point where they could afford the goods they're making and if your view on the economy wasn't a regurgitation of the five year backlog of substack posts you'd understand that none of this is about AI, of course there are useful idiots who think so (like you), but at least you can rest assured that none of them are actually at the levers
>>108884633This of course is cope for useful faggots like yourself. Don't worry though you'll get raped up the ass by AI eventually.
>>108884650Shucks mate, someone's gotta lay copper
>>108884608explain thishttps://www.bloomberg.com/news/articles/2026-04-08/zhipu-hikes-prices-again-as-china-ai-monetization-wave-quickens>Zhipu raised the cost of access to its most advanced AI model by at least 8%, joining other leading Chinese artificial intelligence players in trying to profit off years of research and computing investments.>Zhipu joins rivals from Alibaba Group Holding Ltd. to Tencent Holdings Ltd. in jacking up charges, responding to surging demand for OpenClaw-like agentic services as well as pressure from investors to begin delivering on profits.>The company last month reported a wider loss of 4.7 billion yuan ($688 million) for 2025, but its shares surged as much as 35% with investors upbeat about the prospects for monetization. The stock was up more than 18% Wednesday, buoyed by a broad-based regional rally.
does it still start speaking chinese to you randomly
>>108884608Actually, the ability to undercut implies a valuation mismatch. If you're selling burgers for $50 a pop and I can sell generic knockoffs at $5 and snatch your customer base, there's a supply demand mismatch and all of your investors expecting growth at $50 a burger are now magically in a bubble.You can prevent this from happening by not allowing yourself to be undercut.
>>108884736holy fuck you are stupid
>>108884269dumbass nigger in tech thinks the western world has affordable living, lol. when the 20 year bullrun ends you might finally get a taste of what life truly is you soiboi
>>108884608ai is not a bubble. the ai bubble is a bubble.
>>108884633schizophrenic
>>108884673no retard you prevent by accepting the existence of technological progress as an abstract concept and NOT expect growth at $50 a burger.
>>108883900Dipsy wins again
>>108884273Lol you know you can run Claude code with DS as inference engine. Right?
>>108884242you already lost your "car makers" when the guys making shit for you car makers are already lost to china. the wider implication, if mutts had any capacity for recursive thinking you'd know this means the wider foundations of the pyramid below sustains the stack above, and though it is harder to corner a larger base if you did that for a given layer you own everything above.dumbass niggers like you don't understand what supply chain really means and china is (and I am) banking on it.
>>108885516you missed his point
>>108883900Why should I care if china has my data?I don't live there and they're probably™ not giving it to anyone who I should be worried about.
>>108883900China's LLMs are at most 10% behind American's top LLMs.Meanwhile, they cost 90% less.China's surveillance state is much, MUCH better than America's.At some point, you've gotta ask yourself, why does the US have 5000+ data centers while China only has 400?
>>108885660>China's surveillance state is much, MUCH better than America's.that's an illusion they like to present to the world but it's nothing more than an illusion. since china refuses to share any crime statistics since the 1980s, their word is completely worthless
>>108885660Their shit is also unstable as a motherfucker because their APIs are all run on Frankengpus.
Why would I use Deepseek over say something like Claude, this is just for general chatting (talking about gaming, my day, running etc).
>>108884242>China overwhelms the US with CHEAP decent/good cars>A basic truck is now 90k before all the extra bullshitUs car market did this to themselves.
>>108885892>China's surveillance state is actually dogshit>Americans and Europeans are spied on by their Jewish overlords much betterMakes sense. Arrested in Europe for making mean posts about Jews online or saying Holocaust don't real. Fired and barred from high paying jobs if you are a meanie to Jews in the US.
>>108885953>don't realShouldn't you be scamming boomers ranjeet?
>>108884273can't beat the industry plantopic allegation
>>108883900The real reason is that V4 fucking sucks and no one would use it at the original proposed rates.
>>108885921>this is just for general chatting (talking about gaming, my day, running etc).why the fuck would you use an AI for that
They're literally China's sovereign labTheir model is shit because they completely moved their stack to unfinished Huawei hardwareNo commercial enterprise in their right mind would do something like thatThey're basically CCP tokensWhich is like who cares but it is the truth
>>108883900Inference is cheap as fuck, ai sloppa shops shilling the fact that subscriptions like claude max 20 are hugely subsidized is just pr
>>108884228>why cant the US or the EU do the same?>usIf it doesn’t enrich the ultra-wealthy at the expense of everybody else it’s un-Anerican>euThe whole point of the economic system is to give every last cent to retirees
>>108883900Just algorithmic improvements on the model themselves and the fact that they don't use cuda but wrote their own ptx inference backend.It cost them even less.
>>108884228The entire western world post-WW2 was fooled in to thinking all centralized government must be bad, and so we have our current system of retarded leftists who literally hate humanity affecting the decision making of nations. Anti-natalism by promoting LGBT/mental illness, anti-patriotism by instilling shame, misandry, anti-progress in the name of environmentalism, wealth redistribution based on race in the name of affirmative action, all of it stems from marxism.Meanwhile China didn't buy that, continued with centralized government, and won bigly. They literally uplifted hundreds of millions of people out of poverty in one lifetime, they handled covid much better than the rest of the world, and now they live in a safer and more advanced than the US.Imagine being able to live in a place that when confronted with energy concerns, they just immediately decide to build more to facilitate those demands. Meanwhile the rest of the world debilitates for a decade because of environmentalist traitors who hate quite literally hate humanity so much that they won't want their own people to live more comfortably. They literally fucking hate you. That's our reality.
What the fuck is a "harness"?
>>108886339>They literally uplifted hundreds of millions of people out of poverty in one lifetime, they handled covid much better than the rest of the world, and now they live in a safer and more advanced than the US.through marxism, yeah
>>108884228political grifters
>>108886339this is your brain on americanism
>>108883900E = mc2 + AI
>>108885516Yes I know. Like I said, this isn't ideal because the model and the harness can be co-optimized, so you don't get the full benefits. But you will unironically get better results than you do with claude anyway, just not as much as you could with a ds-specific harness.>>108884346>retard still believes in benchmaxxed results in 2026I based my assessment on the internal evaluation datasets we have been curating for the past couple of years at work and tracking performance over time. They cover a pretty decent range of real-world usage patterns.
>>108884387
>>108884534Generating 100 lines of correct code takes about 80m tokens. Parsing a 450 page pdf document is about 6m token.
>>108886568>Generating 100 lines of correct code takes about 80m tokenswhat kind of bullshit are you smokingis this Saltman shills ITT?
>>108885957Newfag>>108885953> Arrested in Europe for making mean posts about Jews online or saying Holocaust don't real. Correct. It's basically a germ*n and united cUcK land-only thing. In other countries, people more or less openly say kill all kikes in public.
>>108886560great choice of colors
>>108886572It's OK if you've never tried these tools but then what are you doing in this thread and why do you think you should comment?
>>108886590I literally have tried "these tools" you fucking moron. It does not take 80M tokens to generate 100 lines of code unless your definition of "lines of code" means some weird language which the AI is bad at, like assembly for some microcontroller. If you let it get to 80M tokens that means you're a retard. If you don't have something working within 500k tokens then it's never going to succeed and you shoud just give up.
>>108886597You have proven you've never used these tools, so why double down on your hallucinated reality about them?
>>108886597>>108886609CHAT BOT FIGHT!
>>108886609I want to get my hands on whatever tools the doctor used to remove 99.9% of your brain from your skull when you were a baby
Subscriptions (openai/anthropic) are heavily subsidized right now and they come out to a tiny fraction of their API prices, as they are trying to build ecosystem moats to fleece you later. When broken down to tokens, it is not far from what DeepSneed is offering. However they do change your limits ad hoc, and you can only guesstimate the numbers.The real killer is that, if I remember correctly, neither openai nor anthropic permit you to use your subscription tokens in your own programs, and the API pricing is murderous.
>>108884608keep coping, singularitard
>>108886339I don’t think you’ve actually been to chyna, buddy
>>108886661>Subscriptions (openai/anthropic) are heavily subsidized right nowfalse>When broken down to tokens, it is not far from what DeepSneed is offeringanon are you genuinely retarded or what
>>108884242you can't buy chinese cars in the united states government put big tariffs on them and made it really difficult and sometimes but rarely illegal
>>108884228the US is doing that, they just dont want it to be free for the citizens. the people who are spending the money on these killdrone datacenters are getting unbelievable ammounts of money and legislative grease to push things through. normally these guys would be strangled just from having to wait for bureaucracy to let them do even a 10th of what they have so far.
>>108886343the model just takes in bytes and spits out bytes. the harness is the code around it that tries to make that useful, like storing previous input and output, running a loop until the model produced some bytes that match a criteria (agent), a command line interface with funny loading spinners to entertain you because the process is slow as molasses (opencode, claude code) etc.
>>108887124With openai and anthropic you are getting ~5-7k dollars worth of API equivalent tokens by paying for a $200 a month subscription, and that's assuming you are not even quota maxxing 24/7. How is that not a subsidy?$200 is about 6-800 million tokens on DS with generous cache hits, which roughly translates to ~3 weeks worth of quota at oai/anthropic. Thus, "not far".Not to mention, you will end up paying for more DS tokens because it is not GPT 5.5 in quality.
>>108884114Thread answered correctly in the very first reply.
>>108884246Yes
>>108886553I get your point now. I've been using CC with DS API to run coding. It's fast, cheap, and so fair pretty effective.I've looked at other agentic models but not convinced any of them would be better than what Anthropic created since there's the greatest profit motive them to improve / optimized CC. So I guess it's just up to DS to keep in pace with whatever compatibility needs to be in place for it to work.
Is price even the concern right now? Seems like companies want what's best regardless
>>108887421>~5-7k dollars worth of API equivalent tokens by paying for a $200 a month subscriptionEverything I've seen about the OAI and Anthropic subscription models points to opposite: that they're screwing their customers on value. The sub model is so complex (I assume intentionally so) that I've yet to see a compelling breakdown on how you'd come out ahead on it. The biggest tell for me is the outlandish vibe-code spend being posted by corps, that are obv running API and paying per token. If subs were such a cheap option they'd just use that. On a per-token basis, DS is 1-2 orders of magnitude cheaper.
>>108884228Electricity is not expensive in China, but it's not as dirt cheap as you imagine if you were paid local wages.
>how do they do it? and don't give me the "they collect your data" crap, every provider collects everything they possibly canCollecting your data is the likely answer though. If you're not on the frontier, being cheap enough is the only way to attract skilled users, who are more willing than the average slopper to pay for better models. I don't think they particularly care when someone happy with their penis.svg output, but someone with software taste using their service a lot is extremely valuable.
>>108886560>colros don't match between graphsThanks AI!
>>108883900It's almost like DS has been writing papers about how they lower inference costs of serving their models, and maybe taking some of that work and using it themselves. Also, note OpenRouter pricing isn't worlds off either, so even Western providers can get in the ballpark, just can't match it. Given there's probably much less optimization, and the cost basis is higher, that would make sense. It's certainly close to Chinese pricing than that of other SOTA Western models. >>108887529Makes you wonder about the value prop. Or at least it should.
>>108885660>Meanwhile, they cost 90% less.fake news.see >>108884673chinese big tech's ais and chinese six ai tigers all cost roughly the same as american ai. only this obscure tiny ai lab named deepseek is currently undercutting the competition. also not by 90%.
>>108887529>Everything I've seen about the OAI and Anthropic subscription models points to opposite: that they're screwing their customers on value.That's because you are listening to customers with $20 a month "pro" subscriptions who are bitching about their 6 million token code churn getting cut off too early due to the 5h quota on their pack-of-cigarette plan. They have no idea. They have never paid API pricing. The subscriptions are ridiculously cheaper. I've used both. I've spent ~8 grand in a month before through the APIs, with proper caching, and I can rarely saturate the $200 plans with much frivolous workflows. It does not even compare.Compared to the subscriptions, spending $200 on DS API gets you at best the same amount of tokens, except lower quality. I don't mind, I use DS as well, it's good that it's cheap, love it, but it is not killer value UNLESS you want to or have to use API. It's perfect for my own programs. Definitely not better value than CC or Codex with subs.As I've mentioned, from memory, you are only allowed to use the oai/anthropic subscriptions in their own products. Meaning you can't use it in your own program, or hook into it and use it as a gateway. They are trying to build le ecosystems and muh moats. Hence the pricing, for now.Never looked into Enterprise offerings because I do not care. No idea how it works. If I had to take wild guesses I'd say companies are probably paying a fortune for API because they want their own custom sysprompts and heavily programmatic workflows, and/or using their own programs and products, and/or using tools which are not available, and/or want data residency, at ANY time, unrestricted by quota periods. I'm sure there are fine print reasons for it which, as a pleb, I don't care to be aware of. Maybe it's an IP and ownership issue as well, no fucking clue.
>>108887830>They are trying to build le ecosystems and muh moats. Hence the pricing, for now.that's conjuncture.you normally do subscription pricing by assuming most of your subscribers will not use anywhere their limits.if everyone showed up 24/7 to their gym subscriptions, no gym would work. but openclaw meant that subscribers showed up 24/7. so openclaw got kicked out.
>>108887884Either way, if you do saturate your subscription quota to at least ~5% you are already well in the money compared to API pricing. That's why I suspect there is something more to it. Not sure if surrendering your traffic for "training" is opt-in or opt-out, for example.
>>108887935All business subscription plans are opted out of training. The limit on subscription is extremely low to the point it's impossible to do even 1h worth of AI work (which is 10 minutes of human work but nevermind that) per day on it.
>>108887935>~5% you are already well in the money compared to API pricingwrong
>>108887830Good context. You're correct; the only plan I tried to calculate out was the $20/mo anons were bitching about. Haven't even tried exercise w/ $200. So I'll take your word for it. Meanwhile, I'm sticking w/ DS. I do very little with it, still haven't burned through the $20 I put in from Dec 2024. >They are trying to build le ecosystems and muh moats. I agree 100pct. They're attempting classic lock-in. If they manage to discourage enough devs from being devs, they'll have succeeded in cornering the "coding" market as the only way to code at scale and speed will be using their service. >>108887884>conjuncturePerhaps, but from the outside I think the strategy's pretty apparent. It's not like these companies are going to tell you outright what they're doing. They even lie to themselves internally ffs.
>>108887984>they'll have succeeded in cornering the "coding" market as the only way to code at scale and speed will be using their service.that's something completely different than le ecosystems and muh moats
inference margins are over 70% and subs aren't nearly as subsidised any more as people seem to thinkds4 is a terrible model, the only redeeming quality being the low cost. but the token burn and time it takes to get the piece of shit to do anything useful because it's retarded + trips over basic fucking tool calls makes it worthless for real agentic work.ds4 flash is about the same tier as gemini 2.5 flash lite in terms of usefulness
>>108888026>ds4 flash is about the same tier as gemini 2.5 flash lite in terms of usefulnesswhat a weird comparision. don't tell me you used the worst harness, gemini-cli, to make that comparison.
>>108888037pretty sure you can't use either model in that, now deprecated, piece of shit.i hit flash lite on api for baby shit, i don't run that little retard in a harness. v4 flash i run in pi and it's genuinely worse at tool calling than gemini flash 3. pro isn't much better. these feel like models from middle of last year.it's clear that the move to huawei was the primary goal for the model and while that is a worthwhile achievement, model quality did suffer. the other chinese labs have better models.
>>108888024No it IS a moat and ecosystem (centered around programming). Anthropic is also trying to get into the 'expert business' so they buy out companies like converge bio and whatever the other one was, qura or some shit in legal. Their offerings are crap for now but their strategy is to choose silos like these and provide specialized solutions for these services. It's the "right" overall approach, but LLMs aren't enough for this and they're still on copium about it.OpenAI's strategy is... nobody knows. It works worse every day, when we talk to the OpenAI team about it they confirm all our problems with newer models that make them unusable even at a basic level for our product offering have previously been reported also by other clients, they have no plan or intent to address any of this... basically everyone I know is getting off them. We talked to google CS team and they also said they're getting their bulk of their customers from former openai customers who can't use the gpt-5 series of models because it's so unbelievably shit. But then they pulled the same shit openai did which is why everyone's moving to anthropic models now if they can afford it (which is a big if). At least google's strategy is obvious: shove it down everyone's throat to inflate their usage artificially, just like they did with chrome. Soon they'll release 30 different new agent standards to prevent others from catching on, like they did with chrome as well.
>>108887959Anthropic was struggling with a severe compute shortage until May. They kept secretly readjusting their quotas like twice every week, while nerfing Opus 4.6 and releasing the "eco" 4.7. Peak bitching was around March, so I assume that coincided with peak shortage. Meanwhile Openai doubled and then kept resetting the quotas to fish away as many vibeshitters from Anthropic as possible, while putting heavy work into Codex. So one hand giveth, the other taketh, but even Anthropic is very usable nowadays, for the time being, and my estimate of 5% usage would probably be accurate as of today.
>>108888026>inference margins are over 70% and subs aren't nearly as subsidised any more as people seem to thinkmeanwhile your favorite gemini model just had an extreme price increase this week at google io when it switched form per-prompt to per-token pricing and unified their usage limits among all their products. this made google's coding harness completely useless and google had to do do two fucking emergency usage limit increases by 9x.https://9to5google.com/2026/05/21/google-has-tripled-gemini-usage-limits-for-antigravity-twice/google has no fucking idea what their margins are.
/g/ once again gushing about deepseek while in the wild you barely encounter anyone using it.this place is just a bunch of self hating westerners and chicom shills and they're all poor as dirt
>>108886560Market share is purely proportional. The AI market itself is, in fact, growing, despite people claiming they're against it. Part of it is forced though, like google forcing it into all of their products. Same with Tiktok and meta integrating it right into their apps and search.
>>108888082Nobody uses these models only because of liability risk. Most pro use-cases demand data residency requirements be met and people aren't willing to do managed hosting for the most part.
>>108888070it's clear that google's post-training team is terrible.3.5 flash is a token bloated mess.if they had competent people optimising their CoT the prices would be lower. currently that thing burns through enough tokens it ends up more expensive 3.1 pro gpt 5.5 med.goog in general seem clueless, selling tonnes of compute to other people instead of doubling down on deepmind.you get the sense they'll be happy to end up around 6-12 months behind oai and anthropic eventually, not too far off from the chinese labs. which means they will make money in the long run.but the fact that they can 9x usage limits only speaks to:1. complete incompetence on their part2. massive margins on inference allowing them to that reflexively
>>108888100The AI market is computed in $, not seats. AI offerings are greatly increasing in price, which is enough to push the market up even with a reduction in use. Companies like scroogle are forcing people to use the features even when they don't want to, e.g. every search you do in google, even if you always skip the AI overview, counts as increase in gemini usage and the market cap is typically measured as that usage * nominal gemini usage cost (and not, as you seem to expect, only paid use).
>>108888082>i'd rather be rug-pulled than use deepseek
>>108888103>2. massive margins on inference allowing them to that reflexivelyNope. They're burning through money for AI and have negative margins. Their goal is to hold out long enough that all the competition goes belly up. That's the game all the players are playing right now.
>>108888063Imo GPT 5.5 is a surprisingly good* model and works quite well. It just came way too late. Their whole 5 series until this one was a failure, and the general sentiment has become entrenched. Especially since Anthropic properly invested in harnesses and tooling, while Sam was hyping the idea of "adult mode", which brings nothing to the table for them other than liability and PR risk. They are trying to turn the tide with Codex, GPT5.5 and more generous quotas.*given, it is definitely a bigger one than the rest, to stay competitive
>>108888063I keep hearing anons spout off that Codex is the best thing ever. Is that all bs as well? >>108888024Control of a profession is the ultimate moat, anon.
>>108888147what do you mean by codex?codex cli harness is clearly a clone of claude code. claude.md works in codex.
Current Arena ranking: 32nd.
>>108888143>*given, it is definitely a bigger one than the rest, to stay competitivebased on what?opus 4.7 and gpt5.5 are both believed to be 5 trillion parameters.
>>1088881032.5 and 3 were quite good in their own right, I wonder what happened
>>108888169This thing. Which I thought was locked to OAI infrastructure and sub. So it's just another agentic coder then.
>>108886339>They literally uplifted hundreds of millions of people out of poverty in one lifetimelmao
>>108884273>anthropic's harness is far beyond what anyone else has at the momentIt isn't, other harnesses like Pi outperform it in benchmarks while also not being complete spaghettti code.There was a window of time where Claude Code did have the lead but that gap has closed entirely now. Even OpenCode is better.
>>108888225>believedBased on what?A chronic mythomaniac social media-based content for entertainment only?
>>108888270>(((benchmark))), which literally haven't been accurate to real life since basically the beginning>Pi>OpenCode is betterlol, lmao even
>>108888236the codex app? that's just a dashboard app for agent orchestration. anthropic didn't even bother doing a direct equivalent.
>>108888143GPT was good since March with 5.4 + double codex limit while claude had been shitting the bed before XAI deal, most people are just slow to catch up. the best product could change in month or weeks, which is a new thingbtw unlike claude I have never seen "codex deleted my database"
>>108888283the new claude desktop app is almost 1:1 the codex desktop app. and the codex desktop app isn't just a dashboard for orchestration - it's a frontend for the cli + a boatload of other features. it's openai's main focus these days, more so than the traditional chatgpt experience. it's decently made.
>>108888225I meant bigger than 5.4 (and whatever else I remember from the 5+ series, which is not much). Based on latency, cost and vibes. We have nothing else to work from other than rumors which do say it is bigger.
>>108884492This and I never used to pay for AI, I used to work to the token limit then take the output to the next AI and work to the token limit. Shopping like this would normally let me complete full projects before I was truly out of options. Now I hit token limits just watching them take the tooling out. They overestimated greatly people's reliance on specifically their AI when China is now practically giving it away.
>>108888281>(((benchmark))), which literally haven't been accurate to real life since basically the beginningI agree, but would you be happier if I said to just trust me bro? There's no objective measurement of the performance of harnesses + model pairings.Pi allows me to actually manage my context instead of have Anthropic prepend a 40 page system prompt that changes every other day
>>108888334Rumors also say it is bigger than Opus, but that's not what I meant to say originally. Since there won't be a resolution to this for a while now it was pointless from me to bring up, but yeah, it does smell like oai pulled out the heavy artillery with this one, to definitely top their previous ones and compete with Opus.
>>108888334>Based on latency, costgpt-5.5 doubling the price did not go in hand with a doubling of latency
>>108888357>There's no objective measurement of the performance of harnesses + model pairings.tbench exists but obscure harnesses will benchmaxx it
>>108888357>I agree, but would you be happier if I said to just trust me bro?Unironically yes, because at least I could immediately point out that my statements are based on inhouse benchmarks that mimic our real products' customers' use cases, for example. Or we could talk about what makes you feel that one is better than the other e.g. tone or usage pattern, or anything like that.But if you invoke benchmarks, then you are claiming that gemini 3, for example, is what, 2x better than gpt-4o was at math and instructions and hallucinations? This is so far from real life that it becomes doubtful if we even CAN discuss. You know how it is with all the ai brainrotted techlets on the board nowadays surely?
>>108888378>Unironically yescool I'm down with thatmy experience is that the only way in which CC provides a better experience is working OOTB with a wide range of tool calling.aside from running like ass because of an absurdly bloated impossible to reason about codebase, other CC "features" like plan/build modes are bullshit. just ask your model to generate a plan.md file or whatever then execute
>>108884608AI is real and valuable but also a bubble. Both are true and all the discussion that doesn't acknowledge that is retarded and probably being scammed.
>>108885482Correct.
>>108888400That has also been our experience. I think anthropic also agrees, because they are spending a lot on creating specialized sets of "skills" and "tools" and providing them as "addon packages" basically, for different domains. Originally we had been building our own orchestration, but due to budget cuts and providers rugpulling us all the time, we ended up creating infrastructure to let clients use their system of choice instead, including e.g. CC, while still interacting with our platform. Those who report the least problems are those who use claude (either chat or cc). We'll see how this changes over time. Still, there are always nits that need to be fixed, it obviously doesn't work anywhere near as well as it does for ootb etc. It's why we wanted our own harness before, but it is what it is.
>>108883900The chinese guys that subsidize deepseek are evil and propping up the company just to keep other AI models down in relevance while trying to harvest data worldwide. The deepseek team however, very wholesome and loves critical feedback to make their product better.
>>108888400For simple interactions looping over the the llm complete function is not really that difficult. Put some tools, write the loop and exit.It starts getting more tricky when you need to detect cycling behaviour, when some tools might need to abort but do so gracefully, when you need to wait for something to complete while allowing other parts to continue, when you want to protect against too much usage after you hit certain thresholds, when you need to retry whatever, when you need to compact or truncate to maintain good context window and do so with the model in mind (kimi), when messages need to be structured in certain ways to handle various model capabilities, etc.Pi.dev, Hermes and all the other I have seen do not do even half of the things I have enumerated above - not at least out of the box.
>>108888433AI is not real and LLMs are not valuable. It's GoogLeNet v2.But deep learning is real and valuable. Also, the infra and tech developed for distributed training and serving is real and valuable as well.
>>108888451LLMs are obviously valuable, that has been proven.
>>108883900Everybody saw it was worthless and nobody was using, and the number of users would go even lower if they moved to their extremely overpriced planned prices.
>>108888484Like I said, this is GoogLeNet v2. LLMs' only advantage is they are easier to train in a distributed manner. The point of the paper was to show that the trend from the last few years (attention-based hierarchical CNNs with increasingly large receptive fields) could now be turned up by 11 by making the receptive field of these models cover the entire input at every layer (hence, "attention is all you need" = you don't need the convolution anymore). That was always going to happen, research labs just didn't have GPUs with enough vram to do it at the time. It was a very neat and useful result. But we already knew the CNNs don't perform as well as auto-regressive models like LSTMs, we used them because they were a lot faster to train. LLMs are even faster still (since no convolution) with the same tradeoff: LSTMs are still better.Hence, LLMs in and of themselves are not actually valuable. What is valuable is deep learning. It's the same thing with LeNet vs GoogLeNet: all they did was scale the shit out of LeNet and yet it worked very well and much better than base LeNet, but it also got its lunch eaten a couple months later by far smaller models. The problem is that this time around, all the funding across the entire space was completely removed from all research and development not in LLMs. Still, there are much more powerful things than LLMs right now, such as S4 models and their superior descendants like hyena hierarchy, or Schmidhuber's new LSTM variants, that also scale just as well or better.There has been limited attempts to scale LSTMs before, and they've also shown to greatly improve in performance in that regime (obviously), so it is not LLMs, but the scaling, that was useful here. I hope this clears it up.
>>108888527relax, LeCun, time will heal your wound
most of the things I own are made in chinamy car is Chinesemy computer is Chinese my AI is ChineseI even eat Chinese food every other dayUSA is old news, they just can't competeI hope China nukes the US just because they can
>>108888720I don't like any of the 3 head nigs. They all sold out. Worst of them was arguably bengio because he used to pride himself in being the only one not to sell out. Then he sold out WORSE than the others did. He stabbed his own students in the back repeatedly, it was horrible during the transition. Also all he's saying about safety and all that is bullshit he doesn't even begin to believe in t. knowerAnyway. LeCun had only one thing going for him: https://lush.sourceforge.net/I don't know about hinton, I was never exposed to much of his work for some reason. Ultimately all 3 are super overrated, they were really in a right place right time scenario. Had a good idea once and the rest is history. If you look at their output after they sold out, it's literally 0. In 2010-2018ish, bengio was a rockstar. You could namedrop him and everyone would turn their heads. Now? Who even remember him? All the things the MILA has done in the past almost-decade has flopped hard. Hinton's lab has been out of the running since like 2013 or so already. As for LeCun's, about the same time, maybe 2015 or so. Eitherway, if they were so good, they'd have continued advancing things beyond the ONE (1) real contribution they've made each.
>>108888527Better in what way?The linear attention stuff is better in terms of compute time (I think it applies to both train and inference time) and I don't know what Schmidhuber's new LSTM variants are.I suspect LSTMs might be more sample efficient or even compute efficient in single thread scenarios (after all they are used in the Hutter prize) but I'm not sure how that applies to AI at scale which is trained in whole datacenters and needs massive parallelism to actually be trained in a human lifetime.Now of course faster networks should allow bigger networks, but linear layers are already being used in many of the latest models.
>>108886339>leftistspraise marx gweilothe chinese read marx from middle school all the way trough university, it's mandatorymaybe if you also read marx you wouldn't be a dumb gweilo
>>108888837>Better in what way?Good question and I should have expanded on this.It is significantly more data-efficient.>The linear attention stuff is better in terms of compute timeJust not being auto-regressive is far better in compute time in the first place, correct. That is also what I mentioned in my post.It applies during train, but NOT during inference, because inference is performed using an auto-regressive process eitherway (it will be faster for input, but the time there is negligible).>but I'm not sure how that applies to AI at scale which is trained in whole datacenters and needs massive parallelism to actually be trained in a human lifetime.The train time difference is not actually that dramatic. It is sufficiently dramatic to be a problem if you want to train on the entire internet, but it's not a 'whole human lifetime' kinda thing. It's more like, instead of churning out a new model every 6 months, companies would have had to do 3 year cycles. From memory, it was roughly this ratio at scale.
can't wait until I can run this shit locally, qwen 122b and strix halo is nice but after using opus I know how much better it could be.
>>108888864>Just not being auto-regressive is far better in compute time in the first place, correct. That is also what I mentioned in my post.>It applies during train, but NOT during inference, because inference is performed using an auto-regressive process eitherway (it will be faster for input, but the time there is negligible).That's arguable, I think for most use cases most of the cost is going to be input tokens, especially for not cache friendly situations.It's just that it can be parallelized much better.
>>108888875The model in the OP might be worse than Qwen bro. Have you not tried it? It sucks (for code at least).
>>108888976>I think for most use cases most of the cost is going to be input tokens, especially for not cache friendly situations.This and that are totally different matters though. Roughly, for inference, each token you output is computed by taking the entire text so far including the input and all current output, and then outputing the distribution over tokens, then sampling from this, then doing it all over again. This is effectively an auto-regressive process. For an LSTM you compute an internal state as you go through the input, only that internal state and the immediate previous output is being transmitted to the next step, so you don't have to go through the input a bazillion times. The ability to compute this state also means that in principle you have unlimited memory and no hard limit on input and output size. The problem is that the internal state is the entire view of global state, and is of limited capacity, so you end up with cases where the model might forget important information to answer a question, or have trouble retrieving a needle-in-haystack type information. There are many ways around this and it works very well in practice though.Importantly, it means the input size cost for the LLM is effectively reproduced for each output token, but the same is not true for the LSTM, but the LSTM has to "slowly" go through the whole input first, hence why I said it is negligible on balance. This is, of course, greatly simplified on both ends.
>>108889024In practice you use a KV cache, so generating the second token in a response is much, much cheaper than generating the first one.You only have to multiply a single Q vector by each K vector and then do a weighted sum of V vectors. During prompt processing you have to do that for each token in the input. The point of linear attention is that the amount of compute you have to do per token is fixed, both during prompt processing and generation. So yeah, I suppose it pp vs gen stage doesn't matter much, unless the fact that small vecmuls are harder to parallelize than large matmuls means the computer savings show up more in pp.The latest Qwen models use a stateful linear attention layer called Gated DeltaNet which as I understand it is RNN-ish. I'm not sure if during PP you just compute it autoregressively or there is some shortcut to compute it in one shot.
>>108883900They're on the same pareto frontier as everyone else. Nothing special.
>>108889224>score between Gemma and Gemini FlashThat is embarrassing.
>>108887358>Implying the data centers are actually going to finish construction
>>108883900>>108884209Composer 2.5 is 60x cheaper than GPT5.5 or Opus 4.7. Let's not at like China is the only people capable of doing these things.
>>108889195I already explained that this was significantly simplified. Similar optimizations exist on the lstm side as well for example. Read nigga read. What a useless post.
>>108889319>Similar optimizations exist on the lstm side as well for exampleWhat are you referring to?
>>108883900china focuses on long term objectives. america focuses on quarterly profits. a decade ago China was pumping out cheap energy as a baseline for industry, while america was figuring out how to squeeze every penny from American energy consumers.
The power of communism.Capitalist sissies just can't compete.
>>108889299composer 2.5 is a fine-tune of kimi 2.5, a chinese model
>>108889651I know, but lets not pretend that Anthropic, OpenAI or Google couldn't finetune their models to run at equivalent prices if they wanted to.
>>108889668they cannot. that's not how finetunes work.opus 4.6 is a 5 trillion parameter model. finetunes do not change that parameter count. so inference cost will not be lowered.
>>108889700They could launch a distillation attack against themselves and create a version of Haiku distilled from Opus.
>>108886306>Every last EU cents to retirees.Don't forget our muslims. We love our muslims.
>>108884190>the model itself isn't permanentneither is claude or gippity losing billions on inference even when fucking you over with usage cuts every two weeks
>>108884142Absolute retard. They just have a very efficient model using tiles.
>>108891285Wrong
>>108889716>>108889700They could just create new models, stop acting like Chinese researchers are sitting on some magic tech that's not available to the rest of us.
>>108888110What website is that?
>>108891451https://copilot-billing-preview.github.com/
>>108891465I can't get it to read my .csv, anyway this shit sucks balls.I canceled my sub.
>>108886706It doesn't have to be. WHat I said is still true.
>>108887723>all the Americans have to do to get ahead is copy China>they refuse to do it because they think copying is "un-American" o algokek
>>108883900>how do they do it?Iirc power is just cheaper in China, so it probably doesn't cost as much to run the servers there as it does over here. Years of shitting out gigawatts of solarpanels finally paid off I guessAlso probably government subsidies + more people that use it the more, better training data they get to make V5So some combination of those things imo
>>108886339the biggest trick the jews ever pulled on the mutts was making their goys believe the government is always the antagonist.governments, the public sector, work for the people. they are the only form of organized entity in america today that does so. public assets are things you pay for via taxes that work for you. private assets are also things you pay for via profit on goods you buy that do not work for you.
>>108893174Not sure if profoundly retarded or glow-in-the-dark tiers of retarded
>>108884673which extension removes bloomberg paywalls?
>>108893807you are too stupid to know better
I just use Grok for fresh infos from X and Grok Build. It's chill and works for me. Am I retarded ? >>108893820BPW
>>108886306>The whole point of the economic system is to give every last cent to retireesWhat's the average monthly state pension in the EU?
>>108894763Depends a lot on country and salary. For me, the projected amount is US$3000/month, and I also have a private pension surplus I will receive on top of that. Keep in mind that I joined the workforce late so I'm getting less than other people get.
>>108892602Solar is growing, but it's more about their electric grid. Effectively everything that matters is newly built and the system has plenty of spare capacity to use.
>>108894862That's pretty nice at about 2500 EUR, though the average (I have just looked up) is 1443 EUR (1675 USD), which is nothing to sneeze at but nothing to write home about either.If the EU truly is putting every last penny towards pensions, they must not have that much to give in the first place. Your pension is an outlier rather than the norm, and probably not something you should consider all that reflective of EU pension policy overall.
>>108895186>though the average (I have just looked up) is 1443 EUR (1675 USD), which is nothing to sneeze at but nothing to write home about either.Yes, that sounds like it would make sense.>If the EU truly is putting every last penny towards pensions, That's really not the case (I wasn't the other guy, but for example our infrastructure is top tier and the government here is expanding in all the right ways, they've also reduced taxes across the board this year, things like these), but this depends on specific countries. EU is not actually unified in any particular way. For example, the 1443 eur figure is due to 2 factors: 1 is average salary. Mine is high (which is why as you correctly point out, my expected pension is higher than usual). 2 is specific country. Cost of life in portugal is quite a lot less than it is here, and pension essentially follows. In my country, the average public portion of the pension is ~1700 eur, for example.
>>108895186I couldn't find a better map in the two minutes I've spent on it, but this is probably a good approximation (add ~20-30% for inflation). The pension systems themselves differ in the EU, not just the amounts, but a 3kpm pension is not typical.
>>108895221Yes, that's all about what I think too. I understand why the average is lower. It's an average that includes all of the EU countries, and that includes some pretty poor countries, which I know is essentially what you've just said.More what I'm driving at here is that a high pension means that you could say your specific country is putting a lot of money towards pensions, but not that the EU overall is putting a lot of money towards pensions. That was the original claim that my posts are referring to. If we're going to talk about the EU overall and how it allocates money to pensions, we have to talk about all of the EU, not just some subset of it that skews our understanding.
>>108894577>I just use Grok for fresh infos from X and Grok Build. It's chill and works for me. Am I retarded ?considering musk fired the original grok team and rented grok's datacenters to anthropic now, there's not much of a future left.https://techcrunch.com/2026/03/28/elon-musks-last-co-founder-reportedly-leaves-xai/https://www.anthropic.com/news/higher-limits-spacex
>>108884608AI as a concept is no bubble. There is no reason why a machine should not be able to outperform meatbags in anything.The question, for specific tasks, is "merely" when it will happen and at what price. And over-optimistic expectations could be considered a bubble here.
>>108896168The same can be said about the previous 2 AI winters, yet they were clearly bubbles.
>>108883900>how do they do it?they don't wear kippa
