They are not most expensive per Task. DeepSeek 4.1 flash is bloody efficient.

I currently did run a test myself: - Use the pay as you go offer on OpenRouter on the same task on two different project: Perf optimisation on C++ codebase both with Anthropic and DeepSeek.

- I exploded my 15$ budget in a half-week with Anthropic.

- I did two weeks and half with DeepSeek.

It's definitively not a revolution, but it is a major jump forward compared to Mistral Medium 3.5.

Worth to notice that they explicitly said they still refining the model, so we might expect something a little bit better at the end of the month

For something trained on a limited budget and 4000 GPUs, it is quite respectable.

It's nowhere mediocre.

It's toes-to-toes with GLM-5.3 which is one of the best Open Weight model available (With Kimi K3) for general reasoning.

I just runned it on code reviews right now and it was able to catch some thread safety issue than DeepSeek-4.1 didn't. And DeepSeek-4.1 is by no means a bad model.

+100.

Thanks God. these open weight models exist.

And the fact Anthropic is currently trying lobby against these models is despicable.

There is no scenario where putting the key of cybersecurity in the hands of few chosen ones is even remotely acceptable.

No government, no company, no entity should have this power.

Soon or later it will be abused (By 3 letter agency or by an insider/leak).

Delayed disclosure is dead already.

So just give the same capabilities to everybody and stop to fuck around.

You mean like Apple today ?

The Apple watch is from 2015.

What new product Apple produced since Tim Cook took over ?

The iPhone Duo ? Ironically it has quite a smell of Nokia N95 all over again.

But I do agree with you, the glibc is all one giant codebase shipped in one giant blob, dispatch under several names.

Even libdl.so is barely more than a wrapper and some glue around the libc.so itself.

That's currently the real core of the problem.

The loader (and libdl) need to be decoupled from the glibc itself under Linux.

Without that, any attempt to ship static binaries (or any binary with a different Libc) will be a source of perpetual pain.

nss plugins and its associated pain (sssd and avahi) are an other examples of that.

https://www.tcmit.org/english/research/car/car01

This is the last century version of the item that Asus is presenting here... but with a Gasoline engine.

The Engine is fixed on its own wheel and attached on the side of the bicycle.

This was apparently was a thing. It was used to motorize Japan at its industrialization debut.

On century later, electrification is now a thing and the same device is back with a battery. The loop is done.

PS: I do recommend the museum for anybody interested in industries and industrial processes.

Recursive descenrs parsers are not linear.

They are generally O(n^2) and can even can go exponential with some grammars if written naively.

It can be pretty easy to do adverserival attacks on most naive descent parser and bring it to its knees.

Packrat parser [^1] are linear, but they are by no means "trivial 200 lines" type of parsers.

[^1]: https://arxiv.org/abs/cs/0603077

My son is born at 02h30 right DST change time.

It tooks 3 f*cking days to get his birth registered in the Hospital new birth registration system.

The hospital (at the time) just got digitalized and most systems where not able to agree if he was born after or before DST changes.

PS: I guess he is born fated to be a software QA later on.

One Toyota research center where I worked had robots that would go around the office and propose Snack to people.

While probably not being the most efficient way of using a robot, it definitively brought a smile on the face of many employees.

This is a little bit like someone pointing the moon and you look at the finger.

The formal proof domain goes way beyond just finding bugs.

It has tons of usages in term of functional safety, protocol validation, cryptography, etc...

The fact Mistral tackle this kind of problem is both smart and not so surprising.

Smart because it is niche enough that they do not front face the big competitors (yet).

No so surprising because the French labs have a well known and long time expertise with formal proof tools (Coq and all its Ocaml associated tools). It has been historically mainly pushed by the aerospace and train industries (Airbus, Dassault, Alsthom).

Nope. This is not my experience.

Public pricing in token/$ is only part of the equation.

Mistral tooling to consume significantly less tokens-per-given-task than the Anthropic ones.

My bills currently reflects that.

I am. I use them primarily through their vibe CLI.

Reason is simple: They are cheaper (by almost one order of magnitude compared to Claude) and still do the job pretty well.

For small programming tasks, quick prototyping, refactoring or anything verbose and not requiring a context too large: I first go to Mistral and then eventually to Claude if I'm unsatisfied.

I also found out some of their models to be more responsive than OpenAI ones (which is not so surprising considering the size).

My tasks are mainly C++ and Python programming. People in other languages might not share my enthusiasm.

I mean: do not take this thing too seriously.

It also score Grok the closest from Macron. When someone knows how much Macron and Musk hates each other, it is not without irony.

That's also my feeling. And that's the curse of many VC funded companies. And they are not even in the classical state of enshitification yet.

> Rust is not going to win this market.

Agree. Rust will never win this market. Nor Zig, which has the same genetical flaws as C++ for accelerators (excessive usage of pointer semantics among others).

> Julia, although beautiful attempt, couldn't gather enough support outside academia.

I will look mean, but for me, Julia is a language that never went to the design board. It sticked to a "Let's put Python on top of LLVM and add a proper GC" with one single objective: "let's make a clone of Python but fast".

My feeling is also that it is an academia niche and will remain one.

> In fact, if Nvidia cuTile, Triton, Jax keep delivering, Python seems unmatched at the moment.

It is, and it is honestly pretty depressing.

Triton solves most of the performance issues of Python for accelerators but also introduces one (several on fact) more DSL, one more tooling ecosystem and solves none of the (long list of) issues related to Python/Numpy programming model.

You're right to ramble. I also believe that the world need a high level language fitting for accelerators that is not Python.

However developing something like that is by all means not a trivial task and many failed there.

This was doom to fail from the beginning.

Swift will always have the image of an Apple product binded and controlled by the Apple ecosystem. This is very unlikely to change.

Nobody sane of mind would bind there entire technology stack on something half proprietary with a support was from the beginning secondary outside of Apple platforms.