Since cost is so much higher than deepseek though it really seems like the biggest reason to use it would be "in the mistral subscription" for personal use or "because you need to abide by european data privacy laws" for companies.

Mostly I'm happy they're still in competition, as I mentioned already elsewhere.

There is demand for AI that is not US or China controlled. Mistral is serving that demand.

No one prevented other people from running their own AT&T phone company in the US. Except the lack of capital, and endless regulations with running telephone polls across the US.

Google even tried more recently to run their own fiber optics and gave up in the end because of entrenched regulations making it impossible to challenge the modern monopolies internet providers have.

This is a joke right?

Also what about SpaceX, Meta and Google? They are clearly still in the game. They are all ahead of Mistral.

;-)

But you're right, Mistral Large 4 seems a much more expensive but overall weaker model across the board.

The reason to use them would be, well, because they're...Mistral. They're in EU, france is large in nuclear energy so the energy cost of training has much less of an environmental impact, the data protection laws are very good in EU. This can really be a deciding factor for many EU companies, and I think it's good that they now have something which is at least semi-usable. (apparently)

It's also good that other options exist, even if they are worse. Deepseek is kind of like a large specter looming over every model release, because it's just SO cheap and good. That doesn't mean we should abandon every other model.

Still, the model is not really competitive on price and abilities alone. I'm mostly happy to see Mistral still in the fight.

I will see where Chonk is better though, since you have both options in one subscription.

Unproductive company that is 3rd place (assuming we are lumping 4 better US models and 3 Chinese models).

The only thing that gives it value is that it's generated in some region in Europe, like random cheese or wine.

But I still can't imagine a usecase outside patriotism.

Privacy can be solved by self hosting a US or Chinese model.

Seems performative.

Or you have no taste.

I don't use Anthropic products at all, so I was not really sure myself but apart from the models they pretty much only have Claude Code afaik.

And now they're publishing a blog post about writing their own vcs because they have 30000 commits a day. For what? Isn't Claude still based on react?

Do they need this many commits just for claude? Or is that about the model, do they develop that much extra internal tooling?

I admit it is very much possible I am wrong, but without further information this seems a bit like a red flag to me.

On top of that, it also feels very much like a normal CLI tool, which is a big plus for me.

It does not support plugins or MCP, which could both be a big negative for many people, but for now I was able to find a good terminal alternative for everything I needed.

Just a huge recommendation from my side as well, not affiliated in any way myself.

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.

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...

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.

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.

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?

Returning to the substantive matter, I'm happy to try to re-state the topic at hand using your preferred symbols.

Suppose:

1. Someone has a type of task to be done.

2. Finding a systematic way to complete that type of task has immediate value.

3. That type of task can be solved at lower per-unit cost without paying for tokens from the latest Anthropic or OpenAI model release.

Then why would you tell that person that they need to "think of better problems"? Particularly when they've already creatied a solution to the type of task they need to complete with de minimis per unit cost.

There is nothing wrong with consumers who choose to become early adopters and ask what new things can be done with a new class of product. And there is certainly nothing wrong with creating a new technology or making a scientific discovery and spending some time looking around for applications of that discovery/invention. These are important things. The PC and Internet revolutions both relied on this sort of activity, so a culture of doing that type of work is deeply embedded in our field.

But, likewise, there's nothing wrong with driving an ordinary nail through two ordinary piece of lumber to solve a problem for a person who has money and needs two boards fastened together. That framer doesn't "need better problems".

I hope this makes the meaning of my position clear.

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 am always on the lookout to automatate as much boring, everyday tasks. A Smart Fridge like that would allow you to generate shopping lists and reduce food waste by alerting you to possible soon-to-be-spoiled food.

A bit like what roombas are doing for cleaning. You still need to do it yourself a bit, but they help.

Not necessary, not even a gamechanger but another neat home automation. Would be nice, sad to hear they don't work well.

I even store produce (ex onions) on the counter tops, so I guess smart counters as well

It might actually be easier to start with the smart pantry! An in-cabinet kit that can be tested entirely dry and warm and open, and then expanded to the fridge.

The fridge/pantry camera is a useful addition, but most of the value would be a system that just tracks what I put in/take out.

A car is a half solution to going to work. A stove is a half solution to preparing food.

Half solutions are what people need. They’re the solutions we are using every day.

There is this huge problem with advertising that makes consumers think a half solution is a 100% solution. Most smart product makers do that trick. But, apart from that problem, ‘the perfect is the enemy of the good’

Edit: but no way would I buy any smart or connected fridge from Samsung. I don’t anticipate that being a positive experience

> A stove is a half solution to preparing food.

A working grocery tracking system is a half-solution to effortless shopping.

A half-working grocery tracking system is useless, much like a car without an engine.

>A half-working grocery tracking system is useless, much like a car without an engine.

Yes. Something designed to be a half solution can be useful. But a full solution which is broken is probably worse than useless.

Me too. I think I’m not communicating clearly. Sorry.

What I’m talking about is a system that does only one third of the steps in the total end-to-end process for ALL your groceries. You do the rest yourself, or use other help.

I’m not advocating a broken system that claims to do that complete process, but really ignores two thirds of your groceries. That would be damn stupid.

Do you understand what I mean?

Because I feel that the technology and space is - so - hyped and fast moving that a lot of cultish feeling rituals seem to pop up, none of which are backed by evidence. Anthropic openly recommend giving the agents.md file an architectural overview of the code, and the one time this was studied they found the opposite - that the agents.md file is best for concrete commands about how to build stuff and such, and - not - huge overviews. This was, and still is, the official recommendation from Anthropic as far as I can tell.

And then there are the benchmarks, how feel vague and not concrete, and everyone kind of knows they're not the best cuz you can't just assign these tools one fixed number ( for multiple reasons ), but everyone still looks at them and compares them.

People share skills and superpowers and plugins and mcps and very, very few of them have and kind of proof they do much at all.

It all feels a bit weird to me, and I've been on the lookout for exactly these kinds of studies more lately, because I think having this research, even if not done on the exact newest models or not the exact, newest thing, are still - vastly - superior to the alternative.

At least 3 times last month I was asked to review a change in a .md file used by agents. And I am like: "yeah I guess it makes sense?"

It feels we need to write unit tests for this stuff, but even how to do so in reasonable time and complexity seems difficult.

And there might be a point to these arguments, vaguely. However:

There never seems to be - any - kind of counter example or reasoning behind the rationale. You have an in depth and empirical study, done by researchers who, frankly, now their shit (most of the time)

And on the other hand a random internet comment saying "nope" because...the models aren't the latest.

If the latest models really would make a difference, you should at least provide some kind of evidence towards that. As it stands though, every time these comments come up this is missing.

There seems to just be a vaguely defined understanding that "everything changes all the time, and nothing you ever research is transferable to state-of-the-art models"

Which brings me to my second point about these kinds of arguments:

LLM models often - aren't - fundamentally different. Yes, they are vastly more capable. And yes, there are emergent properties. But at their core, they function very much similarly. And for quite a while now, there have not been any of these drastic changes we saw when LLMs first become "good enough" for agentic coding.

I am tired of dismissing empirical evidence and studies every. single. time for reasons without evidence and seemingly a vague sense of "no, but my model is different"

This is a field that changes significantly every few months, so using 1 year old models essentially invalidates the entire report, as they are 2 model generations behind and the newer generation models are heavily RL-d around their harness. For example they put a lot of emphasis on context management but that kind of context management isnt done anymore because many models now have 250k or 1M context and inferencing the models heavily rewards cache hits so you should never touch the context until you decide to compact. They comment on plan mode, which is something that Claude decided to remove from their client because they dont consider it needed at all anymore.

I assume the limit for me is memory bandwith, as the 7900 XTX has the same bandwith as the 3090 from what I can gather and I already reached ~60 t/s with Unsloth. Those would fit with the numbers Byteshape has for their cards.

4090 and 5090 have much higher bandwith apparently, so on those cards you can probably get much more out of the kinds of performance improvements they are doing.

The 5090 destroys both at 1792GB/s.

It's not really one thing with the nvidia cards best I can tell it's that they compounded incremental gains from software drivers, card kernels and optimization from been the primary choice (plus first mover advantage).

I didn't buy the XTX for AI purely gaming but it's a capable enough local card for running Qwen et al.