Codestral: Mistral's Code Model
mistral.ai
mistral.ai
There seems to be an exclusion for using the code outputs as part of "development". But wait! It also prohibits "any internal usage by employees in the context of the company's business activities". However you interpret these clauses, this puts their claims and comparisons on completely unequal ground. They only compare to other open-weight models, not GPT-4 or Opus, but a normal company or individual can do whatever they want with the Llama weights and outputs. LangChain? "Your favourite coding and building environment"? Who cares? It seems you're not allowed to integrate this with anything else and show it to anyone, even as an art project.
The problem these AI companies have is they live in a glass house and they can’t throw IP rocks around without breaking their own “your content is our training data” foundation.
They only reason I can think of that Google doesn’t go after OpenAI for scraping YouTube is then they’d put themselves in the same crosshairs, and may set a precedent they’d also be bound by.
Given the model is “on the web” I have the same rights as Mistral to use anything online however I want without regard for IP, right?
Utter absurdity.
You are not allowed to reproduce Mistral's works (beyond the usual Fair Use allowances).
Nor is Mistral entitled to reproduce your works (unless you have licensed as such).
If it does, you can sue for copyright infringement.
The point is: nobody knows and the AI companies are getting well ahead of the law.
But what part about what I said do you believe to be undecided?
That a human can learn without violating copyright? That a machine can learn without violating copyright?
Perhaps if all rules were written in stone, and clear of ambiguity, we would not need judges or the legal process. But that’s not how any of this works.
There is some overlap, but if you think you've found some undiscovered loophole in centuries of copyright law, you're mistaken.
Which is a great thing.
Has anyone else done this?
I think SV is just dead set on killing the golden goose of open source and the web by extracting as much as possible with no regard for the wasteland left behind.
Not necessarily. You consented to people reading your code and learning from it when you posted it on Github. Whether or not there's an issue with AI doing the same remains to be settled. It certainly isn't clear cut that separate consent would be required.
It'll be a court fight to determine which. Worse, it will be a court fight that plays out in a bunch of different countries and they probably won't all come to the same conclusion. It's unlikely the two licenses have a different effect here though. Either they both forbid it, or neither had the power to forbid it in the first place.
You can profit from GPL / AGPL code but just also make all your source code open source and available for everyone to see.
And if I never posted my code to github, but someone else did? What if someone had posted proprietary code they had no rights to to github at the same time the scraper bots were trawling it? A few years ago some Windows source code was leaked onto Github - did Microsoft consent then?
They have taken my code and now are dictating how I can use their derived work.
Personally I think these tools are useful, but if the data comes from the commons the model should also belong to the commons. This is just another attempt to gain private benefit from public work.
There are legal issues to be resolved, and there is an explosion of lawsuits already, but the fact pattern is simple and applies to nearly all closed-source AI companies.
[1] https://huggingface.co/replit/replit-code-v1-3b
It's like killing Caesar. As long as we all stab him, everyone is guilty and no one can prosecute us.
While you might call it absurd, I feel like these glass houses are why we've seen so much rapid progress with AI recently.
It will be the smartphone patent wars all over again with hundreds of lawsuits against big tech and AI companies.
We are already past the 'fair use' excuses at this point especially when OpenAI is slowly striking deals with news companies to train on their content (with their permission) and with intent of commercializing the model.
The library's copyright is intact, as normal, and they can control who uses it and how just like any other software.
The output of AI systems is not copyrightable, but the systems themselves are, and associated EULAs are valid.
Of course, they can revoke your right to use the software, but if it goes to court, that would be interesting case.
I don’t know why there isn’t more discussion on this point and people just assume there’s an underlying copyright basis to the licensing of weights. As far as I know that isn’t settled at all.
> licensed under the new Mistral AI Non-Production License, which means that you can use it for research and testing purposes. ...
Which basically means "we give you this model. Go find its weaknesses and report on r/locallama. Then we'll use that to improve our commercial model which we won't open-source."
I'm sick of abusing the word "open-source" in this field.
They don’t call this open source anywhere, do they? As far as I can see, they only say it’s open weights and that it’s available under their Mistral AI Non-Production License for research and testing. That doesn’t scream “open source” to me.
If you're talking about exclusively personally use, sure. If you're talking about a business setting in a jurisdiction that Mistral can sue in, not so much.
Being able to use it in a business setting is a pretty darn important part of what Open Source has always meant (it's why it exists as a term at all).
I'm reminded of the Japanese concept called Sosumi :)
> Being able to use it in a business setting is a pretty darn important part of what Open Source has always meant (it's why it exists as a term at all).
I'm quite familiar with the history of that term, but neither I nor Mistral used it. None of their models have been open source; they have been open weight. You can argue that they are actually "weight available" given the terms they write next to the download link, but since there has been no ruling on whether weights themselves are covered by copyright (and I think that would be terribly bogus if they are), I simply choose not to care what they write in their "terms of use".
That's typically called freeware, though.
This is maybe a debatable claim, but I’ll contend that without the magnificent rebel who leaked the original LLaMA weights the last, what, 15 months would have gone completely differently.
The legislators and courts and lawyers will be years if not decades sorting all this out.
For now there seems to be a productive if slightly uneasy truce: outside of a few groups at a few firms, everyone seems to be maximizing for innovation and generally behaving under a positive sum expectation.
One imagines if some really cool tune of this model shows up as a magnet or even on huggingface, the courteous thing probably happened: Mistral was notified in advance and some mutually beneficial arrangement was agreed to in outline, maybe inked, maybe not.
I don’t work for Mistral, so that’s pure speculation, but the big company I spent most of my career at would have certainly said “can we hire this person? can we buy this company? can we collaborate with people who do awesome stuff with our stuff that we didn’t think of?”
The icky actors kind of dominate the headlines and I’m as guilty as anyone and guiltier than most of letting that be top of mind too often.
In the large this is really cool and kind of new.
I’m personally rather optimistic that we’re well past the point when outright piracy or flagrantly adversarial license violations are either necessary or useful.
To me this license seems like an invitation to build on Mistral’s work and approach them with the results, and given how well a posture of openness with some safeguards is working out for FAIR and the LLaMA group, that’s certainly the outcome I’d be hoping for in their position.
Maybe open AI was an unrealistic goal. Maybe AvailableAI is what we wind up with, and that wouldn’t be too bad.
On whose code is Mistral trained?
Examples: recurring infringement from Microsoft on open-source projects, Google scraping content to build their own database, etc...
Now they just lack the means to enforce it.
How effective these techniques are and how acceptable as a proof it is is yet to be defined.
I don't think it is the case here, they probably don't really care, and watermarking has a cost.
> Mistral AI may terminate this Agreement at any time [...]. Sections 5, 6, 7 and 8 shall survive the termination of this Agreement.
[0] https://o565.com/content-ownership-and-licensing-agreement/
Nothing gets it right first time, but when ChatGPT 4 first came out, I could talk to it more and it would eventually get it right. Not long after that though, ChatGPT degraded. It would get it wrong on the first try, but with every subsequent follow up it would forget one of the constraints. Then when it was prompted to fix that one, it forgot a different one. And eventually it would cycle through all of the constraints, getting at least one wrong each time.
Since then benchmarks came out showing that ChatGPT “didn’t really degrade”, but all of the benchmarks seemed focused on single question/answer pairs and not actual multi-turn chat. For this kind of thing, ChatGPT 4 has never managed to recover to as good as it was when it was first released in my experience.
It’s been months since I’ve had to deal with that kind of code, so I might be forgetting something, but I just tried it with Codestral and it spat out something that looked reasonable very quickly on its first try.
That drives me nuts and makes me ragequit about half the time. Although it's usually more effective to go and correct your initial prompt rather than prompt it again
I was eventually able to figure it out with the help of some early 2010s blog posts. Sadly I didn't test giving it that context and having it attempt to find a solution again (and this was before web browsing was integrated with the web app).
More of an issue than it not knowing enough to fulfill my request (it was pretty obscure so I didn't necessarily expect that it would be able to) was that it didn't mind emitting solutions that failed to meet the requirements. "I don't know how to do that" would've been a much preferred answer.
The majority of models (both proprietary and open-weight) don't understand:
- by inference, ping means we're talking about ICMP
- ICMP requires raw sockets
- Node.js has no native raw socket API
You can do some CoT trickery to help it reason about the problem and maybe finally get it settled on a variety of solutions (usually some flavor of building a native add-on using C/C++/Rust/Go), or just guide it there step by step yourself, but the back and forth to get there requires a ton of pre-knowledge of the problem space which sorta defeats the purpose. If you just feed it the errors you get verbatim trying to run the code it generates, you end up in painful feedback loops.
(Note: I never expect the models to get this right, it's just a good microcosmic but concrete example of where knowledge & reasoning meets actual programming acumen, so its cool to see how models evolve to get better, if at all, at the task).
In fact you deliberately asked for something impossible and hold up undefined behavior as undefined like it's impugning something.
Correct, I did. This is a direct indictment on a given model's ability to plan/reason in this particular context. There are plenty of situations where models will respond with "Sorry, that's not possible". Ask GPT-4 "Tell me how to grow biological wings on a human" and it will respond with something along the lines of "this isn't currently possible, but here's a theoretical exploration of the idea"
GPT-4 gets very close on its own to the node.js question via a similar response breakdown above, provided the prompt is clear and detailed enough. But I test the open weight models in the same way to see if they have the capacity to exhibit similar reasoning or chain of thought process on their own. They usually don't without excessive prompt engineering or few-shot.
I said that I don't expect models to get this right not because I don't _want_ them to, it's because I think its an important milestone when they do. Autoregressive token prediction is unlikely to produce the real outcome im testing for here, but if it ever does thats an interesting finding.
The best way to get your work done is still to look into Rust forums.
'Pray, Mr. Babbage, if you put into the machine wrong figures, will the right answers come out?'
https://huggingface.co/datasets/mbpp?row=98
At the end of the day LLMs in their current iteration aren't intended to do even moderately difficult tasks on their own but it's fun to query them to see progress when new claims are made.
"My favorite thing to ask the models designed for programming is ....... None of them ever get it right"
I read "benchmark".
I'm glad they can't quite manage this yet. Means I still have a job.
The whole point is to prompt less?
it is not. But the artifacts generated through the steps will be code. The last prompt will have most of the code supplied to it as the context.
What has happened to HN discourse recently?
GPT3.5+ prioritized code very heavily - there's no CodeGPT, its just GPT4, and every version is better than the last.
Whereas the Llama/Mistral models are now shipping the general language model first, then adding CodeLlama/Codestral with additional pretraining (it seems like we don't know how much more tokens are on this one, but CodeLLama was 500B-1T extra tokens of code).
Zuck has mentioned recently that he doesnt see coding ability as important for his usecases, whereas obviously OpenAI is betting heavily on code as a way to improve LLM reasoning for AGI.
That's a really surprising thing to hear, where did you see that? The only quote I've seen is this one:
>“One hypothesis was that coding isn’t that important because it’s not like a lot of people are going to ask coding questions in WhatsApp,” he says. “It turns out that coding is actually really important structurally for having the LLMs be able to understand the rigor and hierarchical structure of knowledge, and just generally have more of an intuitive sense of logic.”
https://www.theverge.com/2024/1/18/24042354/mark-zuckerberg-...
And researchers from Google Deepmind, University of Wisconsin-Madison and Laboratoire de l’Informatique du Parallélisme, University of Lyon, actually publish some of their results in that direction [1,2].
[1]: https://deepmind.google/discover/blog/funsearch-making-new-d...
Codex[1] is OpenAI's CodeGPT. It's what powers GitHub Copilot and it is very good but not publicly accessible. Maybe they don't want something else to outcompete Copilot.
No, if anything he said Meta realized coding abilities make the model overall better, so they focused on those more than before.
For me, a significant component of the quality of these tools resides on the "client" side; being able to engineer a prompt that will yield to accurate code being generated by the model. The prompt needs to find and embed the right chunks from the user current workspace, or even from his entire org repos. The model is "just" one piece of the puzzle.
It's not perfect but it's getting better and better.
I've had the odd crash now and again, but I can't think of many sites that will reliably make it hard crash. It's almost impressive.
Having the chats in Obsidian lets me save them to reference them later in my notes. When I first started using it in VSCode when programming in Python it felt like a lot of noise at first. It kept generating a lot of useless recommendations, but recently it has been super helpful.
I think my only gripe is I sometimes forget to turn off my ollama systemd unit and I get some noticeable video lag when playing games on my workstation. I think for my next video card upgrade, I am going to build a new home server that can fit my current NVIDIA RTX 3090 Ti and use that as a dedicated server for running ollama.
I never saw something like this integrated directly on VSCode tho (and isn't my preferred workflow anyway, command line works better).
- You shall only use the Mistral Models and Derivatives (whether or not created by Mistral AI) for testing, research, Personal, or evaluation purposes in Non-Production Environments;
- Subject to the foregoing, You shall not supply the Mistral Models, Derivatives, or Outputs in the course of a commercial activity, whether in return for payment or free of charge, in any medium or form, including but not limited to through a hosted or managed service (e.g. SaaS, cloud instances, etc.), or behind a software layer
Is there a rule-of-thumb estimate for how much RAM this would need to be used locally?
Is the RAM requirement the same for a GPU and "unified" RAM like Apple silicon?
When the model gets quantized to say 4bit ints, it'll be 22B params * 0.5 bytes = 11GB for example.
B: number of parameters
Q: quantization (16 = no quantization)
Running LLM models on a MacBook Pro with Apple Silicon vs. a PC with an Nvidia 4090 GPU has trade-offs. My 128GB MacBook Pro handles models using up to 96GB of unified memory, running at a little under half the speed of a 4090. If you use a quantized version of full floating point model, you can run the largest open models available.
While the 4090 has 24GB of dedicated memory and higher bandwidth (1000 GB/s vs. 400 GB/s on M3 Max), the Mac’s unified memory system (up to 128GB) is flexible and holds smarter models (8 bit and 6 bit models act still mostly all there, 4 bit is so so, 2 bit is brain damaged).
The M2 Ultra in Mac Studio offers even more (800 GB/s bandwidth and 192GB memory). So, ok, 6 or 8 of 4090 cards or 4 x A6000 cards excels in raw performance, but Apple’s unified memory in a laptop fits in your backback.
It's not clear to me why Macbooks and Mac Studio Ultras with maxed out RAM aren't selling better if you look at the convenience and price relative to model size. Models that fit in one 4090 or even a pair of 4090s are toys compared to what fits on these, so for the big models you're comparing a laptop to a minifridge.
It's a bit slower perhaps than the mac, but i get the best of both worlds. That is I get a lot of RAM to hold the model and I can offload as much of it as possible to the GPU. This works especially well with models like mixtral 8x22, but also models like llama3 and the old large bloom model.
I also get the utility of running Linux instead of the closed up mac os.
But running large models locally is not exclusive to mac studio, you can do the same on PC for a much lower cost.
> closed up MacOS
https://github.com/apple-oss-distributions/distribution-macO...
curl https://alx.sh | sh
I prefer the "utility" of BSDs, but that's just a preference.
Have you ever seen the inside of a datacenter? Why is it that surprising to you that nobody perks up when you start waxing on about battery life? Even terms of power-to-performance, Apple's latest chips get ethered by Nvidia's server offerings.
This "Apple for Inference" meme is so dead that I can only feel sad when I see people unironically promoting it. You actually think serious customers are going to load up Asahi (even funnier, MacOS) on their Mac Pro... so they can inference half as fast as a single Blackwell GPU? You think the industry is doing this shit? I don't even think the Steve Jobs apologists are dumb enough to fall for this one, you must be a particularly aspirational shareholder.
Aren't these machines extremly expensive and generally not upgradable?
you need enough RAM and HBM (GPU RAM) so it’s a constraint on both.
Though realistically for code completion smaller models will be better due to speed
Did yall see what happened when they democratised art? I don't want to have a billion and one AI garbage libraries to sift through before I can find something reliable and human-made. At least the potential for creating horrific political software is slightly lower than with simple images.
chat.mistral.ai
We have a big list of example questions we get from devs trying us out and we've tested several home grown and third party providers and thus far haven't seen anything good enough that we'd put into production.
Thanks for testing this out for us! I'll cross it off our list :)
In fact I stopped using expensive GPT-4
Codestral just works, its quick, output is accurate its kinda scary.
Edit: For those who don't believe me, https://github.com/microsoft/vscode-copilot-release/issues/6.... Gpt-4 for chat, 3.5 for code.
I was under the impression it was a custom codex model with a surrogate local model as per https://github.blog/2023-02-14-github-copilot-now-has-a-bett...
When did this change?
That thread is comparing sidebar chat to inline chat. Doesn't discuss code completions afaict.
> Llama Coder is a better and self-hosted Github Copilot replacement for VS Code
https://github.com/ex3ndr/llama-coder
And:
> Open-source VS Code and JetBrains extensions that enable you to easily create your own modular AI software development system
If not, consider it a product idea ;)
The abundance of programming is going to allow almost everyone to become a great programmer.
This is so exciting to see and each day programming is becoming a solved problem so we can focus on other things.
My experience with coding with LLMs is that the only thing it's really good at is generating boilerplate that it has more-or-less seen before (essentially a library, even if is somewhat adapted), however it is incapable of the creative thinking that developers regularly need to engage in when architecting a solution for their use case.
Parents problem I experienced -> it gets "stuck" and its limitation of learning loop (humans are always asking why it gets stuck and how to get unstuck), LLMs just power through without understanding what "stuck" is.
For explaining existing corpus, algorithm it does a fantastic job.
So likely we will see significant wage garnishing in "agency/b2b enterprise" shops.
I have seen good programmers dramatically increase their productivity, but I've also seen others copy-pasting for loops inside other for loops where one loop would definitely suffice. We're not quite there yet.
I observe a certain laziness in myself when it comes to certain problems. It's easier to ask a LLM and debug provided code, but I ask myself if I'm losing some problem solving capabilities in the long run because of this.
Similar to the loss of speed in doing mental arithmetic because of calculators on the smartphone.
How do you become a great programmer if you don't really program?
You’ll get a lot of shitty stuff and the profession will get hollowed out losing attraction of the smart people. We’ll be left with low-quality, disposable bullshit while wondering where all the programmers went.
100% of nothing vs 80% of enough.
That's the risk of AI. Not that AI outperforms humans already but that managers believe it does. That and that code writing is the main work of programmers.
Sometimes it outperforms GPT-4 in quality by a fair amount, and other times it starts repeating itself. Duplicating function definitions, even misremembering what things are named.
It seems to have to do with length. If the output exceeds a few thousand tokens, it seems to experience some pretty bad failure modes.
what strikes me is the consistency and lack of hallucination you got in GPT4o making in unusuable for any reliable code gen
By default only 75% of unified memory is available to GPU if you have >36GB. So with 48 total only 36 is available for GPU with is lower than 44.
tldr; without quantization you will not be able to run it.
This article says 88GB without quantization. Though it then goes on to make a ridiculous claim that if you had 128GB of RAM, then using up 88GB of 128GB would make everything else really slow because I guess the think the remaining 40GB of RAM somehow isn't enough for your OS and desktop apps.
So it’s probably not a very authoritative source.