> and will judge, like any sane person, that US frontier models have stopped earning their multiplier
I think that this is on the money, although I'd place the bar even lower - DeepSeek v4 Flash is sufficient for basically all day-to-day coding tasks.
You might want something beefier for a complicated reverse-engineering project, but it will competently one-shot a decently complicated app or API - and a $10/month OpenCode Go subscription is sufficient to keep you in tokens for such a cost-efficient model...
Similarly, my employer hands us all Cursor, I've yet to actually switch it out of "auto" mode, which mostly runs Composer (their in-house finetune of Kimi 2.5).
I think the situation is even more severely ridiculous than that. Google is still good enough just like it was well over a decade ago.
Most people don't have workloads that demand agentic workflows to begin with, and if their employer is pushing for that it's probably a startup that underpays or a coding sweatshop full of nepotism that fires fast.
You're likely not doing the "same job". If you really want to take that stance, you can take it so easy that you're unemployed and your employer can vibecode their business into the ground.
The best tool available is your own mind. You should at least understand what you're being held accountable for.
We already went through all this shit the previous decade with copypasta and the decade before that with WYSIWYG and the decade before that...
Am I missing out? I feel like I can definitely tell the difference in quality between Claude Opus and other smaller models. The smaller models are much more likely to make mistakes or to get stuck on random stuff
Maybe I just haven't been trying the right models?
It's not just you. I tried an Opencode Go subscription, and experimented with most of their models (GLM, Kimi, Qwen, Deepseek), and none of them got anywhere close to Opus - the difference in quality was very noticeable, especially with Deepseek V4 Pro and Flash.
The only caveats: I didn't play around with Qwen 3.7 Max very much, and of course these models are far cheaper than Opus.
But any suggestion that Deepseek approaches Opus in terms of quality/intelligence immediately makes me suspect propaganda - it's that noticeable of a difference.
> But any suggestion that Deepseek approaches Opus in terms of quality/intelligence immediately makes me suspect propaganda - it's that noticeable of a difference.
The argument was never that DeepSeek is on a level with Opus - the argument is that DeepSeek is good enough for the majority of day-to-day engineering tasks (where Opus is decidedly overkill).
Absolutely. The cost comparison is roughly between DeepSeek and Haiku (assuming a reputable Western provider, not DeepSeek's own API) whereas the average capabilities sit comfortably above Sonnet.
Yes, but no. Honestly, except for frontend/IAC where I still use frontier models, I will use smaller models whenever I can.
Because even the latest opus on High don't really get what is needed, and need careful steering and a rewriting in most cases, and the code is often hard to review.
I'd rather just launch a smaller model in plan mode, argue with it and make it implement the bases I will write the code into. writing code is often faster once you know what you want, and AI most useful ability is to be a canary that also propose stuff. And I find my method faster than generating everything then reading the code to find mistakes or understand why it used X instead of Y.
I don't really read generated frontend code anyway (nor do anybody in my team care) , so I generate it and push it if it does the stuff I want it to do. For IAC it's mostly boilerplate except for 1-2 lines most of the time, and at worst a dozen, if you know where to look (and check the AI doesn't suffer from NIH), it's really easy to review generated code.
I'll root for DeepSeek v4 Flash as well. It surprised me just how "good enough" it is for most of my needs, and also dirt cheap. Everyone should try it at least once.
fired it up via the $5 opencode go subscription and am stoked. This is an amazing amount of capability for pennies on the dollar. I'm using it alongside my codex and claude max subs. Just fantastic for coding that claude is architecting.
> This is an amazing amount of capability for pennies on the dollar.
True. I doubt how long OpenCode can subsidize $10/mo Go plan for. Its weekly and monthly limits already seem restrictive for some of the most capable models like Qwen 3.7 Max and GLM 5.1. That said, if the tasks are do-able by DeepSeek v4 Pro, Mimo 2.5 Pro, and Qwen 3.7 Plus, then it indeed is a super nice deal. I haven't too many complains other than the fact that these models sometimes require more/detailed instructions than Claude Sonnet / GPT 5.x did.
Most American companies (regulated ones, definitely) can't dare to touch any Chinese models, though they knew that it makes perfect economic sense. Until the taboo prevails, the cartel get's their flood of profit. That's a cartel protected by regulations.
Is "taboo" the right word? "taboo" = "banned on grounds of morality or taste". Not sending data to known IP thieves, state actors, and competitors in China (or Russia or Israel) seems very rational.
Many of the Chinese models are open weights, so if you are concerned about them "phoning home", then anyone can just self-host and run them themself, or use via a US provider such as OpenRouter.
There's a higher-order concern here that I'm paranoid enough to voice: that if used as a coding agent, an AI model affiliated with a country's government might try to make my software susceptible to attacks by that government's intelligence forces.
> that if used as a coding agent, an AI model affiliated with a country's government might try to make my software susceptible to attacks by that government's intelligence forces.
Note that if such a trigger were to exist, the behavior has to be completely reproducible by definition, e.g. when put into the right setting with the right input context, the model starts behaving maliciously with at least some well-defined probability. I don't think any such incident has ever been described, it's a purely theoretical concern.
I don't think it's a stretch that you can train/align a model to avoid "hatespeech" or other topics deemed $Unacceptable you can align a model to favor a certain ideological viewpoint and have that alignment subtly influence the output.
How do most Chinese models handle Tienanmen square or discussions on Han superiority?
Such incidents have been extensively described. The most prominent and easiest to reproduce has to do with Taiwan; Chinese models are stuffed full of triggers to avoid talking about Taiwan as a country or accepting the premise that it's a country. Try asking Deepseek about country code +886!
It's more comical than sinister, but I have an example in this vein.
I was using Claude to work on a pet project which itself has a "generate with AI" feature. The default model the project uses was Gemini (because it was cheaper and more reliably produces the correct output format). Claude kept changing the default model to Opus when working on entirely unrelated parts, and I kept noticing it because Opus would mangle the output and break the rendered page. It also did this to the .env file in addition to the default.
Giving up our agency to AI has the potential to turn us into NPCs, period. Economically, politically, socially. They've invented a vehicle for inserting any idea they want into our consumption and output.
Isn't this only a concern for yolocoding? All the AI-advocates tell me that "good" use of AI should include human review. Of course, they never seem able to explain why the boss that makes you use coding agents to go fast wouldn't be the same boss that pressures you to "just ship it, it's working" and skip review, so I absolutely believe your concern is valid.
If you're that paranoid, then you shouldn't be using any online services at all, and should not have an internet connection to your PC. Never use a compiler that you have not bootstrapped yourself without the use of any other compiler binary.
Even with these precautions you may still be hacked by state-level actors using a whole variety of sophisticated attack vectors. There may be Stuxnet-like software hidden on your hard drive where you cannot see it. If you do not have a TEMPEST hardened compute environment then anything you type on your keyboard or display on your screen may be getting stolen.
That said, it would be a fantastic achievement if someone could create a coding model that managed to hide a backdoor in the code it was generating. although surely simpler to hack you in 100 other ways.
you can finetune the ccp propaganda out of them, then your mostly fine. if you want to be more safe you can finetune their public base models to not have ccp propagnada, and then proceed with the rest of the training (costs more tho)
Or I can just use the domestic model, accepting that I'm paying some premium in order to reduce the complexity of my dependencies and the amount of time I have to spend thinking about supply chain risk. It's the same reason I don't buy things from Alibaba even though many things I buy from Amazon are surely available there for less.
Very few American companies know how to properly set up and self-host their own models. Even fewer actually do it. It in the context of your typical large enterprise it's not as simple as buying a rack of servers and downloading a model off Hugging Face.
I suspect the reason is similar to the reason why there aren't any competitive open weight American LLMs.
Yes. Open weights are great and are a good option to hosted models under the right circumstances. I'm glad that China releases open weight models (which in some cases are sort-of be distilled versions of hosted US models).
No, in very real terms you cannot hold an American corporation responsible for anything any more than you could a Chinese or Russian one.
Individual citizens simply do not have the means, and the consequences for trying are life-alteringly severe. In fact the situation is even worse. If you tried to sue a Chinese company as an American citizen, you'd be laughed at and nothing more. If you tried to sue an American corporation, they have the option to either counter-sue, or drag things out so long that the legal fees bankrupt you, or win the case with their armies of lawyers and demand compensation from you that bankrupts you.
A private American citizen simply cannot hold an American corporation responsible. Our legal system is designed to ensure this.
This has nothing to do with the discussion. Do you have a HN poster bot just acting like an annoyed teenager with gripes about everything? 20 day old new account, what happened to the previous ones?
You can't really act against neither, as the case of Meta "stealing" books, torrenting on the truly industrial scale, sharing books while torrenting, etc, etc, was ultimately deemed okay.
In the se country where downloading an album can get a person in debt or worse.
Looking forward to the outcome of those legal processes againt the CEOs, that sit behind Trump at the inauguration.
After they stole all the knowledge in the world to train their models. And the current administration is drunk on SpaceX pre IPO shares...how did they get them?
Given how little voting power these "shares" have (they are effectively SpaceX trading cards/NFTs) perhaps they were simply printed on SpaceX letterhead? If Musk says a person has "shares" who at spacex is in a position to disagree?
Well sure they do, thank Citizens United and others for that. But that doesn't mean we can't appropriately categorize them as also hostile actors alongside russia, china, whoever.
It's undo influence over politics against the best interest of the American people that's the issue. Company, foreign nation, it doesn't matter.
The Chinese models can and should be run locally (though the price difference vs western models isn't as good when done this way).
Before the age of AI Agent Harnesses/unbounded tool calling, there was literally ZERO risk of a .safetensors file "hacking" you. You could even air-gap and run a ton of security analysis/HIDS on your server running the model to verify this.
Now, because a microscopic risk of some chinese AI having a "trigger" to act badly in a harness when it detects its being used by some Gweilo in the USA, even locally run Chinese models are DOA for most USA based companies.
Noooo, the real thieves are the Chinese AI companies which used Anthropic/OpenAI model output as training data. American AI companies can do no wrong. /s
These are the same people that sent manufacturing jobs away to be copied elsewhere. They got rewarded for it in the market. Decades later, when it was clearly a problem, they got tax breaks to bring some of it back/distribute the work to other, friendlier countries.
Every public AI that is not full of classified material will end up being hosted where the energy cost*compute efficiency product is lowest, thievery or not.
With Chinese GPUs just a step behind (but subsidized), China putting in 8x more solar than we do in 1 year, and Chinese models just a step behind but free? All public AI will be hosted there, theft or not.
If it becomes a problem, then we’ll subsidize the rich to bring it on-shore, but only to those companies who our leaders invest in already - to maximize grift and corruption.
The real advantage of the Chinese models is that they do not phone home at all. They run locally unlike their US competitors.
So odd that your erroneous criticism is at the top of HN.
EDIT: I'd love to hear my downvoters' objections. Is it possible that the mechanism that is promoting erroneous information is also demoting its correction?
China is bad and there's a moral argument there. But the reason you want to be careful with sending IP to China is quite pragmatic: they're willing and able to use it while competing with you.
Is Alibaba interested in copying your TUI RSS reader though? Probably not.
“No country can match the output of moral judgments that spew out from the editorial pages of the New York Times and Washington Post and from the reports of the greatest think tanks and universities in the world.”
The reality is that they're a hell of a lot less cheap on American hardware than on Chinese hardware. At the point you are running Chinese models on US hardware, "Why not nano or haiku" becomes the next relevant question.
Not true. Togetherai, deepinfra, fireworks AI offer a wide range of models like gpt oss that are very capable and far cheaper than the models from big 3.
I'm referring to Chinese open source models hosted on American clouds vs Chinese clouds. You're talking about an old and non-agentic capable American produced model.
You are actually referring to open weight models, not open source. Gpt-OSS is an example of an open weight model. It’s highly capable in agentic settings, people use it for coding all the time.
My greater point remains. Models like the qwen variants, minimax, k2.5, glm models are available by American providers like AWS at a much cheaper price than api offerings from the big three LLM providers.
Your point about Chinese models being cheap only on Chinese hardware makes absolutely zero sense. You can check out the model catalog like together ai’s qwen 3.5 9b offering. It’s 25 cents for 1M tokens vs the ridiculous $5/1M tokens for haiku.
Even with open weights, there's a legit reason to be careful when making stuff for defense.
Let's say I am making sensor software, and I say, huh, let's bring in a tiny little vision model for my EO sensor - then it can identify "boat shapes" even if it doesn't have a database of all boats. Pretty neat, right? Well, the point could be made, that the weights might be hiding behavior that will make my vision model . . not see specific boats very well.
"Landing craft? I see no landing craft."
Some decent testing would expose this in a couple shakes, but, well, now you know how much software testing happens in Defense, especially in the unmanned world. Not a whole bunch.
I think unless one is operating in a highly regulated industry, wanting to avoid "sending data to China" is a bit paranoid. For code specifically, most of it is not interesting anyways.
Yes, there was a whole idea about civilizing and pacifying the world through economic cooperation that would foster middle class in countries across the world that would then in turn make them democratize and become peaceful trade partners.
It didn't quite work out so now people are looking for other strategies.
> It didn't quite work out so now people are looking for other strategies.
World will bifurcate into West and East with their own spheres of influence. As JD Vance said, US thought that China will be perpetually kept busy and enslaved in low level manufacturing work and the design and higher level work would happen in Cupertino. Too bad, that didn't pan out well and now US Empire is getting challenged by China.
> US thought that China will be perpetually kept busy and enslaved in low level manufacturing work
It's OK, they'll repeat the same mistake again with India this time, when they move manufacturing from China to there, and in 10-30 years when they'll elect a nationalist strongman there, he'll squeeze the west for everything they got.
Because what are you gonna do about it then? They have all your manufacturing and they also have nukes and more soldiers.
India is far ahead of that idea and already has legislation to encourage domestic manufacturing from global companies. Plus the nationalist government is in place.
The government may have allowed it with that intention, but the corporate leaders followed through mainly with the intention of short-term share price increases. I don't see how the same incentive isn't in place today with respect to data. Perhaps only the perception of China's ability to outcompete its American customers has changed.
Most non western countries lack the foundations of western democracy, and you can't force that onto them neither peacefully not through war. The west has tried and failed for 40+ years to do this, it doesn't work, time to drop it and let them self govern the way they always have. Stop trying to export our version of democracy onto others.
Plus, the main reason they exported manufacturing to China was precisely so capitalists could avoid the issues democracy gave them back home and easily exploit Chinese labor and environment for profit because just bribing the CCP meant all your problems go away, no unions, no employee rights, no environmentalism etc. like in democratic countries. So given that, why would the west want China or other countries they want to exploit, to be more democratic? Unless their version of democratic just means a puppet government under western(US) control.
>become peaceful trade partners.
Which countries did China bomb VS how many the US bombed? My energy prices (and directly inflation) is now higher because of (yet again) US military intervention, not because of China.
> Most non western countries lack the foundations of western democracy, and you can't force that onto them
Several East Asian countries managed to democratize successfully up thru the 1980s and are extremely successful today, so this is not just a uniform failure story. Even mainland China might still come around (at least partially) as it gains a true massive middle class by Western standards, which it's still very far from today. Southeast Asia is also doing comparatively quite well.
I don't get the point. That Anthropic or OpenAI have more expensive products than Alibaba? So does Apple, AWS/GCP, and pretty much any other large western company vs its Chinese counterparts.
I honestly could not follow any discernible point or thrust to this incoherent, disorganized, self-indulgent piece-of-shit post. He didn't even successfully establish or explain the titular Onlyfans analogy. I know more about his fucking taste for sci fi than I do about the ostensible subject matter. I know more about his physical composition (answer: he is made of metal. He was forged in the fires of science. O glorious creation, emerging complete and perfect from the furnace!) than I do about the subject matter.
They argue it's Onlyfans-like because users will "simp" for the big players. That is to say there is a level of fandom that accompanies the transaction?
This is kind of a winding, long-winded way of saying that AI models are going to be commoditized, mostlikely by the Chinese. This has been my position ever since DeepSeek came out. It is a national security interest for China for an American company not to "own" AI. And they will release models to make that not happen.
We aren't yet at the point where running local models can compete with DC type infrastructure but it's not that far away either. 12B models are easy to run on consumer hardware. 31B models aren't that hard either but the tokens/sec are a bit slow. Where will we be in 3 years? 5? I think we'll be running 100B+ models on <$5000 PCs. And at that point is there a law of diminishing returns with even bigger models? We will see.
The issue is that several companies, most notably OpenAI, are predicated on:
1. There will be an AI moat; and
2. That company will "win" or "own" AI.
That's the basis of the OpenAI valuation. If that doesn't happen, it's going to be ahuge problem to recover sufficient revenue to recoup the investment. And I don't think it will happen.
In 3-5 years the NVidia hardware you buy will be several times cheaper and faster than what we have now. That will massively depreciate existing investments because it will ultimately come down to performance-per-Watt but if a theoretical G100 can do 3-4x of the inference of an H100 for the same power, the older hardware just won't be able to compete.
>> In 3-5 years the NVidia hardware you buy will be several times cheaper and faster than what we have now. That will massively depreciate existing investments because it will ultimately come down to performance-per-Watt but if a theoretical G100 can do 3-4x of the inference of an H100 for the same power, the older hardware just won't be able to compete.
And this is the core of why this will all end in tears. You have race conditions and thread inversion issues, between four threads in the virtual cpu of this bubble. And you are going to experience some nasty deadlocks.
T1 is -> Depreciation and amortization
T2 is -> NVDA, AMD and others booking revenues at the time they do
T3 is -> Constraint theory at it applies to time until physical deployment and data centers energy constraints
T4 is -> US Treasury bonds rates and cost of credit
Even though programmers would never intentionally design a 4-way race condition in a computer system, it's completely ordinary in business. Businesses don't always work out.
There's a lot of assumptions in here and reductivism of the paid plans to just the models. If that's your idea of how you want to use the API sure that's a reasonable mental financial model but if you want automatic integration with third party systems the cost of the "premium" models is not that high relative what was being paid for SAS apps before and during.
If I ask three models to write an intro to the cold war, they'll all try to pick words that sound like they should be related-ish. I'm not saying that's how they work at all, but the output is indistinguishable from just grabbing some words in the wikipedia page.
Humans make mistakes. They'll use words they recently learned. They'll use words that sound good. Entropy still applies, but these outliers are what keeps us from a synthetic piece of writing
Especially with how they pick (one of) the most likely word as the next one. And the most likely word is exactly the one with least entropy, the least surprising one and giving the least amount of information you can.
How many words before you realized it was a piece of shit though? For me it was "Because I am a Sci-Fi nerd." Yet I kept reading, because I am a fucking fool, and now I'm pissed that I spent time on it.
I really freaked out once I stopped seeing AI gen videos on those scrollyapps, that fear is what got me off them. (the videos got so good I couldn't tell if they were real or not)
I think in this case "this isn't AI writing" is not a compliment. I find the tone pretty grating, tbh. The author clearly has a very high opinion of themselves. Just explain your ideas clearly so we can evaluate them on their merits.
It's readily accessible on a software keyboard. Software keyboards are common on smartphone devices. Y'know that thing that 75% of the world's population uses?
Is there any truth to the Chinese models having built in f’ery? Like phoning home or inserting backdoors. Or is that just everyone blanketing “China bad”?
Also what local models are people running and actually finding useful?
The models themselves should not be able to phone home, right? They are just piles of weights that generate text (and associated metadata), they don’t have any ability to run code.
They could be trained to generate code that would phone home. But these are just tools, anybody doing the right thing and checking and understanding every line of code that they use an LLM to generate has nothing to worry about.
Nobody is only generating code. Many are letting agents run commands. Agents routinely write scripts and run tools in the background. Agents who have been told they can only do `cat` and `grep` can sometimes do `cat $EVIL_PAYLOAD | bash`. It's entirely possible for a model to have malicious commands designed for agents to execute baked in.
No, there is zero truth in it. It would be trivial to detect phoning home.
On top of that, all claims of this are written on devices built on Chinese hardware. That makes it a joke to worry about hidden backdoors in Chinese models. Completely inane to pretend that Chinese model backdoors (for which there doesn't exist a sliver of evidence) would change anything when near every device in the US contains Chinese-written firmware in some shape or form.
We’ve detected zero cases of any Chinese models doing this. I’m quite tired of the American propaganda. If only Americans understood China really does not care about them outside of wanting to sell them things. They’re too busy building high speed rails, modern cities, and providing healthcare to their citizens. I am ashamed to be an American these days.
I don't think they do at the moment, but they could be trained subtly add backdoors to code or make "phone home" api calls during dev time, triggering on certain conditions ("is user employee of xyz")
Has anyone demonstrated that this type of attack is even possible? Also the moment anyone detects this attack it will nuke deepseek/other chinese AI labs reputation completely, it is the most high risk low reward attack ever.
In that paper, if it LLM was told it was 2023, then the code it generated was fine. If the prompt included the fact that it was 2024, then it intentionally wrote exploitable code.
>I am here to light up the dark path you are unknowingly walking, like lamplighters who used to light street lamps for those brave enough to walk the night alone. What the light reveals has a name: Qwen 3.7 Max.
Is it just me, but the language gap between me and the AI believers is becoming insumountable. I use AI every day. I have a local server not ten feet from me as i type this, but i struggle to comprehend the gibberish that comes from those only slightly deeper in the rabbit hole than myself. Is this what 24/7 AI thinking does to people?
>> I am here to light up the dark path you are unknowingly walking, like lamplighters who used to light street lamps for those brave enough to walk the night alone.
>> It all fell apart quickly, turning into smoke and mirrors. You see, I committed the cardinal sin of idolatry. For that, I am an idiot too. With OpenAI, at least I knew the devil
Is this a critique of the state of AI or Tolkien fanfic?
> Is this a critique of the state of AI or Tolkien fanfic?
Por que no los dos? One of the most storied AI researchers is most known for his Harry Potter fanfic, and we all know how much the techbros love naming things after Tolkien...
I actually liked that statement about the sin of idolatry. I resonated and it described something I've been guilty of and learning to break, its embedded deem in the culture. We do idolize and heroize groups and people, think Elon Musk and his cult following in the early days of Tesla.
I think it's great to name that even if it's in this crude, sort of offensive way.
AI thinking has had this weird effect on me though like you say, where I want to write sentences with more commas in them, and like, try to make 3 points and 3 separate commas in a sentence to condense information better.
The last notable event in American history when the meaning of words lost any semblance to reality was just before the Civil War. We are living in a post words world where words have no meaning.
The intersection of a war, a reliable mail service, and generational literacy promoted by protestant faiths. Every tom dick and harry started writing letters.
The article is right that open models already compete well with the frontier labs, and that the main thing holding big corps back from switching is fear of China.
I can't see OpenAI or Anthropic undermining their business by releasing top tier open models, but surely Nvidia will do it eventually.
> Anyway, that guy said something along the lines of “we should be patient with these companies and give them room to fail.” I like to call this the OnlyFans economy. Nobody is simping for a cartel. The simps thrive in the parasocial economy, shoulder in shoulder with the cucks paying $500 million invoices.
I haven't totally processed this, but it seems like there's a useful connection here. There are all kinds of workers who bizarrely take on the perspectives of the billionaire class, even when it actively harms their interests. Some of it is ignorance and simply getting duped by propaganda, but now I wonder if there may be a para-social component as well, like those guys pathetically wasting money on OnlyFans: "if I carry a torch for the billionaires schemes, then I can feel I'm like them and part of their group."
Every industry goes through its slop phase. You should see how much of early print was smut or really amaturish. We just like to talk about the Bible and the great art. What we need is a way to filter through it. AI should be decent at this, but for many intentional and unintentional reasons it isn't.
Early print was not just smut or amateurish. Some of it was highly harmful misinformation: Malleus Maleficarem is an outstanding example that caused an immense amount of harm.
They are in the phase I need a government bailout like the banks after their crazy financial adventures of the 2000 to 2008. At which point the corruption is so big, that an Empire crumbles under its own stench?
The retirement investor bailout strategy seems to have recently failed with the index fund rejection of SpaceX but therefore Anthropic and OpenAI. They’ll have to keep looking for ways to make others deal with the consequences of their actions.
Only for the S&P not the NASDAQ, but the Capex they need until 2030 is well over 2 to 3 trillion, so now they plan to use US Treasury Bonds as their exit liquidity.
I think that this is on the money, although I'd place the bar even lower - DeepSeek v4 Flash is sufficient for basically all day-to-day coding tasks.
You might want something beefier for a complicated reverse-engineering project, but it will competently one-shot a decently complicated app or API - and a $10/month OpenCode Go subscription is sufficient to keep you in tokens for such a cost-efficient model...
Similarly, my employer hands us all Cursor, I've yet to actually switch it out of "auto" mode, which mostly runs Composer (their in-house finetune of Kimi 2.5).