clhodapp

No, but I might have someone who is subject to the restrictions host it or its distillates for me. That entity is going to end up having some kind of non-public deal with Mistral. And that can affect me.

Since the analogy is to source, no, that's not fair and accurate. It's misleading. The fair analogous term would be weights-available.

Try searching for it hyphenated (open-source, open-weight)

It depends how you do it. Everything is a tradeoff. Do you want everything to get assembled on your laptop and pushed to your machines? That means your laptop ends up holding a copy of every package that's going to any machine. Do you want each machine to do its own config build, merely being "puppeted" by your laptop? That could be slow if your machines are small embedded devices. Or maybe you want to run the builds on a beefy remote server because you are doing custom builds from source instead of using the cache... All of this is possible with Nix, but there are real design considerations.

The only thing that's really wrong with the nix language is that it's not statically typed, which can let errors pop up far from the mistake that created them

Depending on your specific requirements, yes. And quite elegantly.

There are several ways to do this. For one, see: https://github.com/nix-community/colmena

> plus I can easily manage multiple machines

> I prefer to bootstrap them from I laptop with a single command

Nix is great at that kind of stuff

In many cases they are not deterministic even with a deterministic seed because many runtimes embrace races that interact with floating point rounding to improve latency.

This talk serves as sort on update on where they are with it: https://youtu.be/r305-aQTaU0?si=j3IqG9BAT0uumPeA

At a high level: It doesn't benefit from batching to reduce memory bandwidth in the same way that a "standard" autoregressive model does, which ends up meaning that it's great at reducing latency for individual/low-concurrency users and delivers better outcomes on certain tasks but much more expensive per token at scale.

It looks like this release will likely unlock the ability to develop relocatable Nix packages on Linux, thanks to some great work by Farid Zakaria and Christian Brauner! I'm hopeful that it also eventually improves the story for mixing Nix and non-Nix software.

I guess what I'm really saying is just that being impressed with a model's size and being impressed with a model's speed are roughly the same thing for similar architectures and hardware.

At risk of being blunt: Why the surprise that a model that activates 12B parameters is faster than one that activates 16B parameters which is faster than one that activates 42B parameters?

It feels like that would be the expected result, all else equal.

I'll avoid naming specific names here so that the reader can (probably rightly) project this onto whoever they want instead of rejecting it for naming their idol but... This is the problem I have with listening to the public "opinions" of major CEOs in general:

Their statements are usually made in service of a goal, not earnestly-held good-faith opinions. Even if someone is "very intelligent", what makes you think they are giving you access to the fruit of that intelligence? How do you know they aren't using their intelligence to try and manifest an outcome that would be beneficial to themselves, completely divorced from what anyone with their level of access to information could possibly believe (much less someone intelligent)? Sure, they'll tell you some truth if they happen to believe that doing that will reward them. But they'll also tell you that they're deeply committed to your local city when they have a meeting to discuss tax incentives for moving their HQ later that same day or that their plant has a sterling safety record while deliberately burying records of payouts for injuries. The only real difference between this and someone like Holmes is that they have a better level of skill in forecasting consequences.

What this ultimately means is: There is often little value in listening because you end up having to estimate the truth from less-manipulative sources yourself. And if you're going to spend the time and energy to do that anyway, you can skip straight to this step.

This is really cool but... At some point, don't we have to start simulating things like internal interference in the conductors, induction, and sending/receiving radio signals?

Hearing a traditional Odersky drone about how the latest language feature his lab has built into Scala just happens to be exactly the tool everyone has been missing from their AI agents somehow establishes a bridge across these crazy last few years.

I think it's less that Odersky sold out but rather that the original company vision wasn't economically viable and that business leadership has pivoted it away from him and his work. He's not even on the leadership page anymore.

I've been thinking about different techniques for making my NAS backup-able to cold storage and... the thing that has given me pause with tape is how closed-down the ecosystem for the LTO drives themselves seems to be. What's your strategy for knowing you'll be able to access your tapes to transfer them forward to the next generation in 10-15 years when the tapes start reaching the end of their dependable life?

It's not supposed to be about superior and inferior, it's supposed to be about bigger and smaller.

Fable kind of screwed that up, though.

Needing fine-tuning for the use-case completely changes the product category

Personally, I think it's more that text diffusion is not the ideal driver of an agentic work loop than that text diffusion is a total dead end. I am still hoping to see how it does on authoring and editing with further scaling and optimization. I think the push for AGI has put a bit too much focus on the idea of one general model doing everything.

Honestly, it's an addressable market versus development cost question... How many clients will you lose if you support OIDC but not SAML? Does the delta justify carrying a SAML implementation? If so, do it. But the post is still correct that SAML is a fractal of bad design either way. And it's good to say this openly, and to run this calculus each time you are considering a new SAML implementation.

I can see how it might require an explicit state machine (keep a mutable state number and switch over the inlined bodies of what could be functions), but I'm not seeing how it could require `goto`.

Are there more-complex relationships that might require it?

No algorithm requires tail-call elimination in a general-purpose language with imperative mutability. It's just another way to express iteration.

There has stopped being a formal procedural consequence for OpenAI leaders to declaring AGI, there is a clear (small) business benefit to doing so, and the capabilities of all the frontier models are impressive. So why not declare AGI? It's not like anyone can prove it's not...

Don't be surprised to see other (or even the same) people declaring AGI again and again, as it becomes the best time to do so for different parties.

It seems to support USB4 with what they call "stock firmware". They haven't clarified what that means, but my assumption has been that it's a binary blob made by the manufacturer of their microcontroller that doesn't support all of their fancy debugging features (thus actually turning my into an Aliexpress special eGPU dock with a high quality buck converter).

It feels apparent to me that LLMs don't do what is colloquially thought of as thinking.

What is less apparent is that humans do.

Yup. You basically just need something for probability to push off of

Seems like they are closer to scratch than reasoning... Generating some scratch to draw from helps make it easier to compute the real answer.

Hrm, but the insistence is more "NAT should not be a practical requirement in a pure IPv6 network", not "No NAT ever for any reason".

NAT-based compatibility strategies are widely deployed in the real world (most cell carriers and Google's IPv6-mostly corporate network use NAT to allow communication with the IPv4 world from IPv6 only devices).