I certainly wouldnt have predicted that 10 years ago.
Very glad to see Mistral still in the game even after some big stumbles with Large 3. I deeply hope that this model is 'good enough' that it becomes the European go-to, giving them the resources to keep the pace up.
I'm excited to try this out today.
Deepseek essentially releases instruction manuals in paper form.
DeepSeek is basically a research lab founded by a hedge fund guy, more than anything else.
Would you say you could be serious in stating this?
They have other issues as well, but among other things, I'd classify them as somewhat charitable too.
Even ignoring monetary subsidies, there are the non-monetary ones: not being sued into oblivious by the government for their countless hacks of other companies and countries, the slaps on the wrist for massive piracy, the waving of environmental (and other) regulations in order to allow their data centres to be built an operated.
That's what I'm asking - which monetary subsidies?
> not being sued into oblivious by the government for their countless hacks of other companies and countries
That isn't normally how enforcement works, and it hasn't been very long since they disclosed those breaches. If the victims want to pursue legal action, they can, and they still may!
> the slaps on the wrist for massive piracy
So judges and juries are involved in the subsidization conspiracy, too?
> the waving of environmental (and other) regulations in order to allow their data centres to be built an operated
Sure, though if you think this isn't happening in China too, I have a bridge to sell you.
> That's what I'm asking - which monetary subsidies?
I am listing for you the non-monetary subsidies, which are just as real and equally important.
> not being sued into oblivious by the government for their countless hacks of other companies and countries
If it was one of the Chinese labs doing this hacking, the government would be stepping in. If it was European labs they'd be stepping in. If it was you or I the government would be stepping in. That is a massive subsidy (they don't have to worry about the same legal fees and exposure) and being allowed to continue doing business is in fact priceless.
> So judges and juries are involved in the subsidization conspiracy, too?
There is no conspiracy. They are objectively operating by a different set of rules than you or I could operate in this market.
> Sure, though if you think this isn't happening in China too, I have a bridge to sell you.
I never said it wasn't, I'm saying that it's happening here and it's a very real subsidy.
They really aren't. Nothing in the development of these systems, absolutely nothing, is as important as money. Stop these silly false equivalences.
All the major labs would be dead in the water if the government acted on the things I mentioned. If they treated the labs the way they would have treated us for all those hacks. Or if they enforced pollution measures (or god forbid ban on-prem turbines because of the climate damage). Or rule that training on data is not fair use, or that ingesting GPL3 code and then turning that into weights counts as a derivative work. Or. Or. Or.
Those are all just as important as cash.
The hacks are pretty bland as long as there is no mens rea and the hacked victims do not want to press charges. This isn't a "subsidy", it's just our normal legal process at work.
The pollution question is winding its way through court systems and getting a lot of pushback. I wouldn't call this a subsidy, just a slow bureaucracy. And at any rate this isn't special to the US; certainly China's environmental policy is far more lenient.
With the training data, again there's nothing specific to US corporations here. Anyone can use that dataset. It's just how our legal system works. You might not like it, but it's the accepted law right now. This doesn't meet any reasonable definition of "subsidy".
On the other hand, shoveling billions of dollars of taxpayer funds into corporations is by definition a subsidy. Full stop.
That's just the ones i know off the top of my head in the US. Those programs account for more than 60 billion in committed spend.
[1] Where "Bad Things" would be the typical interest, taxes, depreciation, amortisation plus the Anthropic specific employee compensation, LLM training (you know, for the LLM lab), revenue sharing agreements (which is a form of paying for infrastructure), etc.
For LLMs, marginal cost is just electricity
Assuming you meant to reply to me and not someone else are you saying under GAAP accounting standards Anthropic is a profitable business because under GAAP accounting their only expense is electricity?
(I see what happened you skimmed the conversation and didn't follow what was being discussed.)
2. The marginal cost is electricity until the hardware overflows. So it's continuous for electricity and a step function for hardware capacity.
Plus you don't address the core point. Frontier AI is capital intensive. A lot more capital intensive than regular software. The existing clouds supporting the entire internet (!!!!) were built for a fraction of the cost of AI infrastructure, and there isn't even an end in sight to this continuous hardware investment.
Let alone hardware refresh cycles.
AI is basically investment into roads.
Software used to be a magical place, closer to selling music albums but even better.
AI is a much worse business margin wise than regular software.
Non-punishment is also not a subsidy; again, words have meanings. Let's use the correct words.
Not arbitrarily banning companies from buying a product from a supplier doesn't meet my definition of the term "subsidy".
I don't see how. Anyone can buy electricity without any intervention from the government. An intervention is an action that changes what would happen by default.
Microsoft did not pay with money - it paid (mostly) with Azure cloud computing credits. MSOFT is then able to write this off as a loss against tax.
It is generally far more tax-efficient in the US to write a loss in this way than it is to write a loss for a cash investment.
In this case, I believe the difference was ultimately highly significant. When including MSOFT eventually writing-off the deprecating Azure hardware it had used to buy the OpenAI equity, the result was MSOFT's tax reduction being either close-to or exceeding the actual cash value of MSOFT's investment in OpenAI - IIRC.
These examples represent taxes that the US chooses not to collect - the US could choose to make investments like these less tax-efficient. Instead, by making them extremely tax-efficient, the US subsidises the transaction hugely.
Its very xenophobic of you to say China has zero intention of helping humanity, and just wants to "wage economic warfare".
Last time I checked, it was ourselves (USA) waging economic warfare on 2/3rds of the world.
I dont get this cope people have where people have this idea that its impossible for a Chinese company (that make billions of dollars) to have done something by their own merit, but instead its always some Chinese Communist Party conspiracy where the main goal is to destroy America.
Lay off twitter for a bit.
>Last time I checked, it was ourselves (USA) waging economic warefare on 2/3rds of the world.
Up until recently the USA Wasn't waging economic "warefare" on 2/3rds of the world
China is not a threat to you, or anyone in the West.
Not even our leaders say China is an enemy, ita mostly businessmen who are scared of competition and trying to regulate chinese out of their markets so they can make more money milking us.
We hear this about literally every industry the Chinese excel in - that it's only because the government subsidizes them that they succeed. For chip manufacturing, for batteries, for EVs, for solar, for AI. I don't see how the chinese government can afford to subsidize all of these industries and still have them contribute to the GDP.
There is as much to criticize about American hyper scalers and AI labs and the lack of interest in helping humanity or contributing to open source, but that might not be as popular an opinions on this site.
Do you think the world is closer to or more distant from large scale global conflict today than it was 10 years ago, or even 3 years ago?
Source? The last Supreme Leader of Iran issued a fatwa against nukes, way before he and his extended family where bombed in their home in a failed attempt at a decapitation strike.
Any open weights model that Chinese companies are releasing for free are allowing everyone in the world to have access to high quality realizations of these tools without the risk of the US regime arbitrarily cutting you off.
Also, it's funny how all the neoliberal mantras of how free market drives progress through competition stops when it's a US company that's being challenged.
I think the Chinese government is backing their labs by less direct means. For example cheap electricity and investing in chip manufacturers such as Huawei and cxmt.
Deepseek specifically, is known to operate with minimal resources. The entire company has around 160 employees and every model they release must break even within ten months.
It may be that the effect of the backing in both places is essentially equal but this statement is strange. The Chinese government invests directly in Deepseek[1]. Notice the article, in addition to saying the CCP is investing, says Tencent is also a major backer. CCP owns a golden share of Tencent.
[1]https://www.cnbc.com/2026/10/06/deepseek-funding-round.html
Everything a Chinese company does has the explicit backing of the CCP, at least ideologically and usually financially in some fashion. While DeepSeek may not be the CCP, it couldn't exist if it was expressing any kind of ideology that wasn't inline with them and, in this case, is explicitly funded by them.
A Chinese company has no freedom to say Taiwan is a country the way someone in the US could suggest California succeed from the nation.
Any public message you hear coming out of China has the implicit approval of the Chinese government.
If you're not arguing about comparison then don't state this as if it's any different from the US. Because you make it sound that way. Or do you think that if tomorrow OpenAI came out vocally supporting the DSA - that suddenly they wouldn't see lots of barriers rise up out of nowhere?
Come on now. That's fairy tale land.
If OpenAI comes out in support of _any_ political party, they may lose business from the government in the future.
Fairy tale land it is. Consequences for what I said would be far bigger than "lose business from government".
Supporting deepseek is just like supporting the Chinese army, no need for that. Though it goes both ways, OpenAi subscription just lowers the cost of the US army as well.
Good, where do I sign up? At least they aren't exploding little children and generating chaos in the oil market.
https://www.business-humanrights.org/en/latest-news/anthropi...
https://www.ohchr.org/en/press-releases/2025/06/israeli-atta...
https://www.unesco.org/en/articles/unesco-presents-assessmen...
I want you to publicly say that this turtle biologist and this journalist are Hamas: https://www.theguardian.com/world/2026/jun/20/mona-khalil-tu... https://www.theguardian.com/world/2026/apr/23/lebanon-journa...
It's not a subsidy for the Chinese to say they're going to ignore our IP laws. It's a subsidy for us to say we're going to make them and try to push them on the world. It's literally granting a monopoly by legal force. It's very obviously not aligned with the interests of the American people, while China releasing things in the open is.
Ironically your comment will be cloned and used for training.
Same goes for US labs, not everyone are as innovative as google deepmind:
https://www.forbes.com/sites/antoniopequenoiv/2026/04/30/elo...
Yes, just that. Being a huge supporter of the regime - at a time running a department - surely has nothing to do with it.
... You're saying this about _deepseek_?
As a neutral party, this characterization is crazy.
As if the AI companies - Claude and OpenAI are guardians of freedom and humanity and very charitable to the global society without any self interests… “Chinese models are subsidized by the Chinese Government, therefore they’re inherently bad for humanity” is a highly propagandist argument. The politics of US vs China may be whatever it is in reality.. You have one company releasing their models for cheap, actually open sourcing their trained weights, and publishing details of their optimizations and learnings for others to use. The other camp actively “aligning” their models, nerfing their capabilities, hyping their swarm activities from poor sandboxes, and trying their best to lock users into their harnesses and walled platforms. They are subsidized by the capitalist VCs who are essentially waiting for their payouts..
At some point, one has to see things for what they are and evaluate their own reasoning..
I’m happy to stay provider agnostic, try all models and cheer any useful progress as open as possible.
There are many other reasons Chinese companies releasing models open-source or open-weight makes strategic sense.
A really easy-to-understand example is a company who has a near-monopoly on "serving video content" releasing a video model openly.
If you can be relatively certain that video content created by a model (which you have trained, using data from your own platform) will be ultimately served on your own platform, thus generating revenue from watch-hours, it makes sense to make those models as widely-available as possible.
It's also a net-positive if people use your public research to build better video models, because - again - you are reasonably certain that the even-better content those new models produce will be watched on your platform.
The alternative would making models harder to access and learn from (broadly, the current western model). Many would argue that Google, in choosing to not optimise its video generation models for "availability", is directly causing less content to be uploaded to YouTube. This is the trade-off.
I don't know much about DeepSeek's financing specifically, which obviously doesn't release video models - so I don't know how directly this analogy runs, or who directly benefits from the extremely evident rising tide that the public release of DeepSeek's research creates. However, this does not negate the broader rising-tide effect of the scientific method.
It's certainly also true that it's geopolitically beneficial to be able to undercut American labs' models. If I ran a global superpower, I would probably want my country to be technologically competitive too.
But Chinese companies are already serving a huge volume of customers in a complex, existing marketplace, before even thinking about the US market, and it's overly simplistic to assume that their entire strategy revolves around economic warfare directed specifically at the US. It's more nuanced than that.
This is, of course, without even getting into opening the can-of-worms around whether US economic policy also results in the US state functionally subsidising technological innovation, how comparable that is to China's model, etc.
Causing less content to be uploaded, when you are clearly the gorilla in the room, looks to be a wise strategy not a trade off.
Releasing models openly doesn't mean providing the compute. It's beneficial for relatively GPU-poor Chinese orgs to let users run models themselves - by releasing the models openly - as it means the orgs don't have to provide the compute to run their own models.
Why are we assuming a strong US is necessarily good? As a European, I have seen plenty of evidence against that stance lately.
I understand that Americans might prefer a strong US. But conflating them with humanity is a leap that I don't think one can make without any backing.
The government that removed restrictions on how private companies can access capital after a certain scale (the JOBS Act), that removed the need for private companies to report as if they were a public company after a shareholder threshold was crossed, superpowering the access of wealthy private investors to get in earlier in a growing company while at the same blocking the public from participating in funding growing enterprises at an earlier stage (since it required companies to IPO much earlier to access capital) which allowed retail investors to also reap the rewards on funding them early when they grew to become behemoths (like Amazon, Meta/Facebook, Google, etc.).
It's not fair in either place, the USA has its own model of unfairness, China has a completely different one. The difference is that in the USA the government allows private investors to become more powerful than the State (outside of the monopoly of violence) while plunging the rest of society into increasingly more precarious lives while in China the State is the power and its legitimacy only exists while the population feel they have a better life.
As a differentiator from the Chinese system, not so much. For two otherwise equal companies, the one that says things against the party line will experience selective enforcement too.
The propaganda of trying to make US and China government appear the same is making people dumber. and is one of the most heavily used tools in China’s online propaganda arsenal.
The government could do some very easy things to remove this tool from China's toolbelt. Like you know, prosecuting rich people when they've earned it.
If you look at the comment I responded to, they were remarking on selective enforcement in the USA. I merely observed that this isn’t a unique feature of the US system, it happens in China too. I said nothing about the severity of the consequences.
The story is so much more complicated than that, to the point that this economic warfare theory is basically a meme.
Chinese models are open because they don’t have a choice. “When you trail the frontier, openness maximizes reputation per unit of capability. The moment you lead, you close.” [0]
[0] https://earnedintuition.substack.com/p/involution-without-ex...
Does anyone think that likely? I have no clue or bias.
Yes, because everything China does is against the US. That's all they think about day and night. God forbid they want to corner the global market or have a genuine business case. How dare they provide options for those who can't afford a measly $200 a month? How can we let Chinese labs publish research for free for the whole world so that they can benefit? The nerve! To think they can use soft power instead of military might! I mean, Anthropic and OpenAI are the last bastions of human kindness and charity. Right?
Right?
Notwithstanding that Chinese publishing methods actually does help both humanity and open-source.
You think extremely US-centric. China has a different economic model than US and your rules for a specific kind of Capitalism may not apply to them. You assume a country of 1.4 Billion people is obsessed with a couple of foreign AI companies. What if they don't care.
Charity and kindness is not a motivation, it's an outcome of what you do.
US companies are burning colossal piles of cash in ways that makes it unclear if it qualifies as dumping, not to mention their deep ties with the country's regime.
Claiming that companies from a country have ties to the regime and burn through cash is a very miopic accusation.
It is great for everyone except for a few people who want power over everyone else, and the fact it is not charity of kindness makes it more sustainable, because charity and kindness is quick to go when big money and politics is involved.
I want more warfare like this. Building stuff instead of destroying stuff.
So were Amazon and Uber by the US, which have now established monopolies across the globe. To the countries suffering from those, there's zero difference with China doing it to solar. Actually there is, at least solar got them cheap renewable energy in return. This would never have happened in the US because big oil interests would make it take decades. That's the reality.
You need to spend 10 years outside the US, deprogram, and then go back.
Ok, but so what? That's what happening so far. Even a repressive totalitarian government I wouldn't wish on my worst enemy does some good sometimes.
If the Chinese cheap/open models rise up and destroy us, thats on us for giving them access to the tools to do so.
To say that US model providers have a close relationship with their government would be putting it mildly.
You put it like US has a goal of help humanity or open-source
Its them who chose to outsource heavily, not US consumers.
There are no victims here however, both benefited from the arrangement. The consumers and capitalists.
No, it's the result of US leadership letting this happen. This is clear since China themselves would not let this happen. US leadership did nothing because they were best friends with the people who did it, and did not care one bit about their population.
So the most common way to publish manuals?
This will cause the picked winner to get massively ahead with sheer compute alone used both for training and inference dedicated to recursive self improvement.
There are going to still be worthwhile improvements but they are going to be more like not how to make transformers 10x cheaper but how to make next training run cost 9 trillions instead of 10 with a very particular optimization designed at the cost of hundreds of millions for this one specific run.
Are there any use cases that have enabled one customer of a frontier LLM to outperform a competitor using a different frontier LLM? Or is this why we are seeing confected points of comparison like solving challenge problems in mathematics?
I could imagine a belt of data centres around the equator, that hand off their computational loads as the sun sets. Good scifi-esque premise.
There's a law of diminishing returns at play here, and doubling the energy cost of training to wring 2% more performance out of the technology isn't going to be very useful, because most of the problems it is capable of solving will be solvable with the previous-gen 98%-as-good model.
("there's a law of diminishing returns at play here" is an article of faith. But then, so is the belief that these models will keep getting better).
lol You can always tell who has never ran a business before with comments like this
(I don't think it will work).
Btw: with Qwen4 I mean the next large Qwen model that is based on the Qwen4 architecture (Qwen 3.8 flash next was "almost" based on the new arch but obviously was a small model)
What matters more is if firm’s start using a bundle of American and Chinese models and when they find their feet - how large is the market for frontier?
Frontier has to displace labour one for one at some point or it’s over.
In practice, you will be able to run models a bit bigger than 35B.
38GB of vram resident. more tk/s, more prefill.
I wonder if the next DS models will also graduate to 5.x, I think I saw they are training up a 10T model, and just raised $12B too
It's important to realise that Deepseek are a very unconventional model lab, as they were originally funded by the founder's work in quantitative trading.
They mostly don't need that much money, the reason they took funding is to provide employees with equity, as lots of their good people were getting poached by other Chinese model startups.
(I have no inside info, just read the FT articles about them).
The fact that it basically broke open the local model supremacy was just a nice side effect.
I'm running: https://github.com/peonist-ai/halogen-server with a quant4, PLE offloaded, and it's resident VRAM is 36GB at 265k context.
Shave 10 more GB off and the TAM openai and anthropic are targeting is a lost cause. Local models are what 90% of people will need.
If the world governments can get a handle on the memory cartel, then there's no more moat for most normal humans.
From the Mistral site:
> ML4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own datacenters in Europe.
It is pretty capital intensive!
To put that into context, the last wave of capacity SpaceXAI added 400-450 MW.
Chinese companies also managed to put together their models with relatively small clusters.
Perhaps US companies are desperately trying to brute force their way into workable models?
3,800 GPUs is nothing in the frontier side.
This appears to be roughly as good as Sol 6.1 (which is quite good), considerably faster in terms of wall clock for complete tasks, and considerably cheaper (where Sol 6.1 is already good value - just really slow).
That seems too good to be true...
But I really hope it is true...
That seems much more realistic.
If the singularity is defined as an AI sufficiently intelligent to improve itself independently, that AI is still limited by the resources required to do this improvement.
I trust them and their populations to provide a more societal-friendly version of AI, putting pressure on the US tech oligarchy, while also providing democracy-friendly open models that I don't trust to happen with the Chinese labs.
But Opus 5.5/GPT is such a game changer in comparison to sooo many others, its still a moat for now.
The real question is if this model is good enough that it can still accelerate work, and not be a hindrance to real work like older Mistral models often were.
If they can do that, they'll have customers.
If these companies will steal from deep pockets like Disney or Sony (some of the most infamously litigious copyright trolls to ever exist), they won't think twice of stealing every bit of code you upload to them.
If your code passes through an AI company's servers, you can assume you just gave it to them. In turn, when your competitor tries to copy that new feature you just added, the AI is now trained in exactly how to copy you and eliminate your competitive edge. Unlike your employees, the AI isn't bound by the same rules and even if it were and violated them, your company probably doesn't have enough money to prove it in court (and that's if we somehow reverse some of the stupid "AI is the most transformative use of copyright I've ever seen" judges who have drunk the coolaid).
Most companies could build the compute to run GLM or Kimi models for way less than the potential loss due to IP theft from using third-party systems.
> Following an investigation, we have confirmed that Buckmaster’s Codex prompts over the two months preceding this announcement and paper on September 8, 2026, could not have influenced the system in any way, including through training. The OpenAI internal model used for this result was developed through large-scale reinforcement learning on top of a previously pretrained model. Our proofs also differ significantly. In the Euler case, Alpöge and Buckmaster proved a result with external forcing, while OpenAI’s system proved a result without external forcing.
A simple loophole, use the code to create an RLVR environment where the resultant code is the end goal / max reward. Technically the customer data is never trained upon, but effectively you’re using it. Even better, use the code as a seed to generate synthetic data similar to it and use that synthetic data as rewards in an RLVR model.
Unless you can host the ChatGPT model on your own servers, which I know some enterprises are doing, I don’t think there’s any hope of protecting your data / competitive advantage from these frontier companies. Better to be paranoid, than be commodified by these companies.
There is no ceiling on what you can accomplish with more intelligence, so there will always be a market for the best models, and that market is likely to just keep growing. If Opus 13.5 can one-shot a profitable company or discover a new disease treatment or whatever you can think of that a swarm of relentless super-geniuses could accomplish, companies (and governments) will throw money at it.
I also think there will always be a market for many sub-frontier models that will continue to grow rapidly as well, because "good enough" is definitely a thing for a given task.
No company would ever release such a thing
And yes, open weights are still behind, but are catching up.
"Now, here, you see, it takes all the running you can do, to keep in the same place. If you want to get somewhere else, you must run at least twice as fast as that!"
Yeah, I could buy a bigger and bigger vehicle when a new one goes to market, but using a semi-truck to drive myself around would not only be significantly more wasteful but also more inconvenient and limiting compared to a sedan. Very much like advanced closed models vs performant open ones.
People who aren't afraid of rolling their sleeves into any code base? The difference is practically zero.
Maybe it's because people stopped watching what their agents are doing and stopped looking at the quality of the output. But I still see agents being absolutely mindless like a junior dev.
Recent example: it updated an an API to add newly released models to the backend. There's a list of models that require specific configuration for the reasoning effort and temperature or the API call fails. GPT 6.1 Sol misses this and code fails at runtime because the newer models need to be added to the list for special handling of temp and reasoning. Fixes it for one model and tests it for that model using an E2E test. But doesn't test the other models that were added for the same error condition...I had to explicitly ask it to do so and it finds them and adds them to the list and says "that's on me."
Yeah, not that smart.
I keep seeing this kind of thing over and over, and honestly it's not got _that_ much better since the big breakthroughs about a year ago.
For sure I happily vibecode stuff without worrying about it when it's a greenfield project, and if the LLM has written it entirely from scratch then usually it's well structured and sane. But making changes in messy, mostly human-written mature codebases is still a minefield.
This works fine with Opus 5.5. But it also works fine with GPT 6.1 Sol, Kimi K3 and MiMo 2.6 Pro.
It doesn't work equally well with Sonnet 5.5, interestingly.
For people who have some expertise, the models accelerate the grunt work, but you’re the one validating it.
As for Mistral - I got really excited when they said Large 4 was focusing on being #1 in cybersecurity, because that's somewhere that they genuinely could edge out Anthropic & OpenAI. Have it actually solve problems, instead of Anthropic flagging "you tried to find a null pointer exception bug in your own code, we're now reporting you to the US government". But on the Mistral benchmarks I'm seeing, this looks very disappointing, but at least they haven't entirely given up. I genuinely thought Mistral had given up on new general models. They need to learn the bitter lesson all over again.
The capabilities of all models increasing so much all the time means there are simply less and less tasks you need a frontier model for.
Even if Opus 5.5 is 500x better than Deepseek, if deepseek can solve all my problems, why do I need to pay for more?
Obviously any model will do if you use it as a better autocomplete.
I believe that there is a large gap in expectations between different workflows.
Until the AI like reads my mind and produces perfectly production ready apps with minimal intervention from my side, there is still going to be room for improvement.
Uhm, yes?
It's about 12000 lines of code. I could probably understand every line of it if I spent a couple of months on it, although I'd need to learn some things about WebGL, numerical computation, and solid-state physics. But I elicited it in four days. Probably I'd be better off spending the next couple of months doing something else instead.
So, I was faced with the question of how I could keep the artifact thus created from being completely valueless. My solution was to export the heat-equation solution produced by the solver as (gzipped!) CSV, so that I can use my own code (that I do understand every line of!) to check that the solutions found by the solver are, in fact, solutions.
If that check checks out, the backward Euler solver may be of some value even if I don't understand every line of it.
Now my provider serves DeepSeek around 300 tokens a second. The code is shit but I can have a few more rounds of corrections.
DeepSeek was maybe a dollar for the full PR. Opus/Astra over 50 dollars. And double that if you use their fast variants which matches the DeepSeek speeds on certain providers.
Yes, it is so good I rarely test new models anymore. Or think about cost.
Actually I felt since 4.5, it was “good enough” for basic coding (and competing models have reached that level as well). The big improvements since then have been longer-horizon tasks. Giving it a vague description of a problem, then come back hours later and find it did a better job than I would have. And since 5.5 I haven’t encountered any more load bearing seams.
Maybe it’s slow, but I’m not interacting with the models in a live chat anymore for most work. I’m giving them tasks to do async. The limiting factor is actually my bandwidth reviewing their output.
I am eyeing a huge codebase built with Opus. Did multiple passes of refactoring for things like using traits not just free functions. And using the Rust types like PathBuf from the damn standard library not just strings everywhere.
Oh and it is so hard to use a crate. Let's just vibe an inferior version and push another 5k line PR. Nobody reads them anyway right? Logos schlogos, a lexer is easy to write, right?
Maybe I am just getting old, but I remember a time when we were proud on the code we wrote. Now I read some of the old stuff and get a bit sad for what we lost.
Don't get me wrong DeepSeek also does this. But if the model is 100x more expensive I expect better. Of course if you never read the code it's ok. But it does the same mistakes and cannot design great architecture...
This morning I elicited a microkernel operating system from Opus 5.5. Well, mostly. It doesn't implement task switching yet; we'll see if it runs into a wall at some point. But it boots in QEMU, and it's running a user process in ring 3 and serving web pages.
But if on the other hand, I mostly use my human intelligence and just need a dumb model to complement my human intelligence at low cost and high speed (say review every commit to catch obvious bugs), I have a much better chance of building an actual moat than you do.
But outside of coding, it’s even more clear that you don’t need frontier intelligence. My customer service agent is very happy with a 100B param Deepseek flash model, thank you!
I’m using subscription models for exactly that, better models catch more subtle bugs, and they catch them faster. It works out far better in terms of work-hours saved.
Also A/ then OAI slashed token pricing by 2x~5x on their latest models
I have enough real problems in life. I don't need to invent new ones just because a new technology is available.
Many of my problems in life are fully solved far past my satiation point by a 3b model that costs me nothing to run.
Many others are not.
But in either case, when I am acting and living wisely, almost all of my problems exist prior to the existence of technological solutions to those problems.
This is also true for the customers and employers that I care to work with. This has changed in me over time, but I now try my best to avoid inventing new problems. The world has enough big, important problems already.
> This morning I elicited a microkernel operating system from Opus 5.5.
This is cool but also a good example. I don't need a personalized microkernel just because it's possible to have one.
Maybe I need one and I don't know it, but the problem statement definitely isn't "I have inherent desire for a personalized microkernel".
You're talking about definition 1 in https://en.wiktionary.org/wiki/problem, "A difficulty that has to be resolved or dealt with," with the examples given being racism, addictions, and lack of access to health care. Those aren't the kind of problems AI can help with.
The kind of "problem" that AI can help with is definition 2, "A question to be answered, schoolwork exercise." That is the character of engineering "problems", although they are more open-ended than schoolwork exercises, because there are many defensible tradeoffs. "How can I build a bridge here?" or "How can I improve the fuel efficiency of this vehicle?" is a "problem" in the sense of a question to be answered with knowledge and effort, not in the sense of being similar to racism or addiction.
If the questions you're thinking of are so easy to answer that they can be easily answered by a hypothetical AI model 1/500th as good as Opus 5.5 — well, think up some harder questions.
> well, think up some harder questions.
There's nothing wrong with this mindset, but your insistence that other people must have your values is off-putting.
To wit: at the end of my phd I was offered several tenure-track positions where "work on problems that are hard" would've been a big part of my job.
I turned those offers down due to my distaste with "because it's hard" as a justification for working on problems.
I'm much more interested in helping people and solving (def'n 1) problems. Those problems are almost always too large and messy to one-shot, but very often contain many components that can be solved with non-sota models, and make sense to solve with non-sota models for any number of economic/scaling reasons. E.g.: "can I serve this thing that costs $0.25 per run on even c.2024 sota models to 100,000 K12 students who will use it dozens of times a day on a non-profit's budget of ~$0.00."
P.S. The "problem" definition thing is very confusing to your readers because the person you responded to was clearly using the first definition! It would have been helpful for you to sign-post that you are using the same word as the person you are replying to but changing the meaning out from under your reader.
I feel like the "good enough" argument isn't about how big the gap between models is but about how good they are at solving the tasks at hand.
The capabilities of all models increasing so much all the time means there are simply less and less tasks you need a frontier model for.
Even if Opus 5.5 is 500x better than Deepseek, if deepseek can solve all my problems, why do I need to pay for more?
They are clearly talking about "problems" that DeepSeek can solve, and they seem to be using the term "task" as a synonym for "problem", even using the problem verb "solve" with "tasks" as its object. To me, that makes it clear that they are not talking about problem₁s such as racism, addiction, and lack of access to health insurance; while you could imagine "solving" such problems, they are not tasks, and DeepSeek certainly cannot solve them. Rather, I think they are talking about "engineering problems", which are examples of problem₂, more analogous to math questions (to be diligently sought out) than to troubles (to be avoided).
It is certainly the case that the proposition that such problems should be diligently sought out is a value judgment, and therefore it derives from the value system employed by me, the writer. Would you prefer that I refrain from making normative statements in general, or just those with which you might disagree? In either case, is your preference specific to me personally, or do you consider it a duty incumbent upon everyone?
— ⁂ —
The 01913 edition of Webster's dictionary offers:
1. A question proposed for solution; a matter stated for examination or proof; hence, a matter difficult of solution or settlement; a doubtful case; a question involving doubt. --Bacon.
2. (Math.) Anything which is required to be done; as, in geometry, to bisect a line, to draw a perpendicular; or, in algebra, to find an unknown quantity.
Note: Problem differs from theorem in this, that a problem is something to be done, as to bisect a triangle, to describe a circle, etc.; a theorem is something to be proved, as that all the angles of a triangle are equal to two right angles.
The absence of the "trouble" definition suggests that perhaps it is derived from the "question" definition, and, surprisingly, very recently.
You gotta problem wit that?
I had the exact same experience. And unlike Fable, it doesn't gobble up your entire usage limit in a few hours.
I always wonder what the "good enough" people are actually using it for.
If Opus can’t 1-shot it, then it must rely on our human intelligence which can be complemented well enough with a dumb model as a frontier model.
Prices for the same task (ARC-AGI-2) have since dropped by factor 10 if you compare to Opus 5.5 Low.
Look at the context in which I used that term 'good enough'.
What i was saying is that there are tasks for which a dumber model can be good enough, and for organizations with sovereignty/ privacy concerns, those concerns can be strong enough to incentivize the use of a dumber model.
I hear this literally every other week about whatever the newest FoTM model is.
Unless you can provide concrete examples of things you can do with them that you simply couldn't do with last week's model, it's absolutely meaningless.
It's especially the case as more non-Americans look to self hosted models and domestic cloud inference providers using open models that the US providers who are still leading the charge need to drastically drop their prices and find a path to profitability in order to maintain their lead and retain the advantage they had as AI turns into a commodity (which is happening faster than I think even the frontier labs initially predicted).
I use MiMov2.6Pro, DeepSeekv4.1Flash, GLM5.3, Hy4, Qwen3.8 and KimiK3 at home. Opus5.5 is not a game changer.
I have to admit, Sonnet got really good too.
But Opus just uses tools, a broad spectrum of it, etc. it feels like sure if you add some router behind it you could split it up if you need to but if you give me the choice, its opus allll day long.
Right now they don't even get good feedback from local sessions - I can see it make the same mistake two days running, and then months later when a new model comes out, presumably trained on my data, it still makes the same mistake.
When a user says "I'm struggling to undo a bolt on my 1952 Mustang" and the AI responds "try hitting it with a hammer" and the user replies with "that worked thanks" - that is a tiny piece of knowledge which exists nowhere else.
Future AI's can say with more confidence that hitting it with a hammer will probably work.
Across billions of conversations, that can add up to more knowledge than all books.
AI is an idea 60 years old. We are on the 3rd or 4th generation of AI development. Three years into the current iteration of products.
This is not early days by any measure. LLMs are a result of a very, very mature research field.
Disclaimer: I'm not sure how much of an IYKYK factor applies to this joke.
Mistral is also an European company. As we live in a time where the US regime is engaged in pyrrhic geopolitical tactics, it's good to know that it can't threaten to cut access to models during s period where everyone is rushing to incorporate them more and more in our life.
So it makes sense, since all you need is compute, that there's a ceiling and specialization is going to be more valuable then some super AGI.
Especially since the worst people seem to be the ones who think they'll all run away with the bag.
... unless they can legislate it, which is why they are flattering heads of state and scare mongering about dangerous AI.
Obviously, this is only a valid question if you don't believe that open weights are about to eat their lunch and their revenue is about to collapse, or they're running a super unprofitable ponzi scheme propped up by investor money that's about to collapse like a house of cards. I don't find those positions credible at all though.
If you do, then this question isn't really for you, as I'm more interested in thoughts from those who think that OpenAI and Anthropic in particular are about to be the largest companies on earth in a couple years. Could anyone catch them at that point?
I don't know about revenue, but I suspect multiple other labs are already beating OpenAI/Anthropic on profitability. Staying on the frontier is expensive, and it's hard to recoup those R&D costs when you have a bunch of other labs nipping at your heels.
If you concede the previous point, then the only way for OpenAI/Anthropic to keep growing long term is to swallow the whole economy (i.e. mass job replacemnt), and that's a bet I wouldn't take.
The second moat is convenience, which all the big labs make it (comparatively) easy to glide into their models.
If they can carve out a niche of industrial and governmental partners who rely on them for sovereignty reasons, it may be enough.