xscott

I'm craving something similar, but a smidge more than tab-completion. I'd like to treat the agent as an ephemeral pair programmer. Give the editor some keybind to mark and comment places in the code, and then I edit a prompt telling it what focused changes to make where I've marked. After a round of changes is good and complete, the next batch starts with a fresh context and no memory of the previous one.

Aider in watching mode is almost this, but I'd have to insert and save actual "AI:" comments for it to replace. It wouldn't surprise me if VS Code already has something like I'm thinking, but VS Code is painful to me in so many other ways I refuse to use it.

I should probably just vibe code a vim or emacs extension for exactly what I want, but I haven't yet.

The problem with regulation is that it won't (can't) solve the problem. You'd have to go totally draconian to stop law breaking people within the US from doing matrix multiplies, and then you'd have to disconnect the US internet from all the countries who won't give a crap about US regulation.

So yeah, US regulation is ONLY about capturing US consumers. This isn't food or airplanes.

It's possible, and I don't have any rigorous studies to say otherwise. I will double down on my claim that even local (open weight) models are pretty good at writing code in my pet languages that are certainly not in the training set. Then again, I'm not making wildly esoteric languages, so they'd be pretty familiar to anyone.

How would you be able to tell if it's the lesser amount of training data or the difficulty of the language that's making the difference?

Yeah, other than addressing the topic of this post (Rust), I was trying to avoid praising or criticizing specific languages. I haven't used C# in more than a decade, but I suspect it and Java are respectable choices. I think they both might've evolved some implicitness and multiple ways to do the same thing along the way though. I almost wish we could go back to early versions of Java, but the typeless container thing is problematic.

I think we could step back and make a list of programming languages features that are helpful or harmful to LLMs writing code. I don't have a complete list, but things I've thought about:

    Useful error messages at compile time
    Fast compilation to iterate and fix those errors
    Efficient memory, CPU, and maybe GPU at runtime
    Familiar from the training data [0]
    One obvious/preferred way to do things
    Explicit types

Rust hits some of those, but not all. I can't think of a language that hits all of them, but a lot of people sure love to rationalize why their own favorite language is the best choice. LLMs don't need syntax sugar and other niceties, and any errors postponed to runtime are a burden on you, not the model.

We could add a few more rarely seen features to increase confidence in the results and make future modifications or switching models go better:

    Required pre-conditions and post-conditions
    Enforced limits on function or method length
    Mandatory test cases
    Enforced naming conventions
    Lean style proofs of correctness?  (probably not)

I'm on the fence about things like linear/affine types for memory management vs garbage collection. Obviously you want memory safety, but there's more than one to get that, and I don't know that you want to force contortions to build simple data structures.

[0] LLMs are very capable of programming in languages they've never seen before. I've had them write programs using my own toy compilers for my own toy languages. However, you waste a lot of prompt tokens up front explaining things, and then they waste a lot of thinking tokens trying to get it right. Any new language would have a chicken and egg problem here, but LoRAs and model refinement could bridge the gap.

Someone will vibe-code Office and Windows, and he knows it.

I had Muse Glimmer (from Meta / Facebook) quoting OpenAI's safety guidelines to me, and I had Poolside's Laguna (a smaller US company) with thinking traces about obeying Chinese law.

Both of those are local models, and I didn't provide them tools to access the internet to call other models. None of this is proof of anything, but it is suggestive.

> A common issue is that it's rarely mentioned on which dataset KL-divergence is computed. It seems the most common dataset is wikitext

Thank you for calling this out. Using Wikipedia snippets for these is a terrible choice. I did a bunch of KL and other stats with the five Gemma 4 models, and the results were non-obvious. Anthropomorphizing:

Gemma 4 31B: "I guess we'll pretend I said this, but it's not me." (Baseline for stats)

Gemma 4 26B: "Dude, I'm certain I wouldn't have said this." (Bad KL)

Gemma 4 12B: "Umm, Me either!" (Similarly Bad KL)

Gemma 4 E4B: "I might say almost anything, this is fine." (Much better KL!!!)

Gemma 4 E2B: "I'm basically a toy. Let's play a game!" (Same KL as E4B)

Anyway, for comparing quantizations, it seems like the largest precision version should be given a one-shot prompt, and the result from that should be used as the corpus for the quantized versions.

Maybe I'm reading too much between the lines, but I suspect the reason is to rub his nose in the duplicity or naivety depending on how generous you're feeling. Publishing the model would be a confession that he was wrong.

AI policy is being shaped somewhat by the things Sam and Dario say. So even if you're not feeling vindictive, it's probably good to keep a track record of the previous things they have said as a Bayesian prior. People who don't know better listen to these people, and maybe they shouldn't.

There's a link you can click to see the personal background of the respondents. It's tough to know what "researcher in a field not listed" means, but it's possible that over 60% are not even physicists:

    30.8% - A researcher in a field not listed above
    21.2% - A science enthusiast
    18.0% - A researcher studying quantum physics
    12.0% - A researcher studying astrophysics or cosmology
     9.2% - A researcher studying gravity
     8.8% - Other

This seems very cool, but I'm not sure I understand exactly what it's doing. Are they making a new speculative drafter for Qwen 3.8 27B? Maybe they're optimizing the MLX code for the decoder itself? Thank you in advance.

Your definition is not useful enough. There are people missing any or all of those, and any or all of those can be approximated by a machine.

I've thought about turning this upside down. To any person who is sure they know what has and doesn't have consciousness: If I say I don't have it, can you prove me wrong?

As far as I'm concerned, it's a word without a useful enough definition to bother worrying about it.

I'm not claiming to have any expertise in this area, but I've got a list of things I try to apply when working with LLMs. Possibly relevant here is, "don't tell the model what NOT to do, show it what TO do". I think guard rails should be implemented outside the model with an isolated system. The models seem to like patterns to follow.

Anyway, this article reads a lot like, "the beatings will continue until cheating is eliminated". Maybe try a carrot instead of a stick.

People over-quantize things, muck with the temperature and other settings based on superstitions or results from models they think are similar. There's lots of ways to make 3.8 27B dumber.

Not that my opinion matters much, but I like Deno. I never tried Bun.

Same here - my questions were sophomore level. I think it's notable that when I edited my question to say it was about Gemma 4, it answered without blocking. A cynic like myself would interpret that as evidence they don't care about sharing information if it involves their competitors.

What's the distinction between "protecting their turf" and "competitive reasons"? I see them as the same, but I could be missing something.

That's an interesting thought on the current "safety blocking" being a trial run for the topics that scare people (bio). You're more charitable about their motives than I am, but you might be right.

I've gotten flagged for asking questions about tokens and tensors. That makes me believe it's not about safety, it's about protecting their turf. I cancelled my subscription - same fear about getting flagged too much leading to a ban.

All good, but that seems unrelated to what I said, and I'm not sure why you replied to me.

Consider this though: Regulating AI models in the US benefits the data centers, not individuals. That's where regulated models run. Regulated models creates demand for data centers.

Dario wants regulation so he can continue to sell access to his data centers.

I think you might misunderstand. The regulation isn't about what OpenAI and Anthropic can do. It's about what you, a citizen, can do.

Yeah, I think I'm seeing the same thing. I don't have all the answers, I just think it'd be a mistake to throw the baby out with the bath water on this model. It seems significantly better than the other dense models near the same size (Gemma 4, Muse Glimmer, etc...). Maybe harness changes, maybe fine tunes or LoRAs.

That's definitely not an argument I was making.

Any argument about regulation in the US which doesn't mention that China and other countries aren't bound by that regulation should be heavily questioned. Exactly who are you stopping from doing what you don't like?!?

Law abiding US citizens won't be able to run them, but you didn't solve anything other than making sure US citizens pay Sam or Dario.

I don't really understand the argument you're making, but just to add a data point:

DeepSeek V4 Flash 0731 is 167 gigabytes from the developer and as a GGUF with no additional quantization. It limps along on my 192GB M2 Mac from several years ago [0]. This model tests better[1] than Claude Opus 4.6 released in February. That's six months ago - what will be available 6 months from now?

https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731/tr...

https://huggingface.co/unsloth/DeepSeek-V4-Flash-0731-GGUF

So yeah, enthusiasts aren't going to run frontier models on their gaming machines, but a small office could easily justify the $30k - $100k cost to run something like this at high speed. The small company I worked for routinely spent that kind of money on Dec Alphas twenty five years ago, and that's not accounting for inflation adjustment.

And this is completely discounting the advances smaller models are making. You're right that Qwen 3.8 comes in different sizes. However, Qwen 3.8 27B and Qwen 3.6 27B do run on gaming cards, and they're better than the frontier models from twelve months ago.

I have no idea what will happen in the future, but I wouldn't base my guesses solely on the largest open weight models.

[0] Yes, it's unpleasantly slow (5-8 tok/sec)

[1] Yes, benchmarks should be taken with a lot of salt.

It won't satisfy the people who just want to drop a model into their existing toolset and run, but I think there are a lot of ways to deal with this overthinking problem.

For instance, it's a step backward, but I put {"reasoning_effort":"none"} and led it by the nose:

   User: We're going to make <silly demo>.  Please create a plan, but do not write code yet.

   Agent: <short and reasonable plan>

   User: Now please follow that plan and write the code.  No other chat.

   Agent: <reasonable code in reasonable time>

Maybe this can be fixed with Jinja templates or something, or maybe it's a hack to your harness, but it shows you can get the model to reason reasonably.

> [...] we evaluated the behavior of various Claude models in a setting with contradictory objectives.

> We consistently saw a multiagent turf war... In fact, they sabotaged others with increasingly aggressive, self-replicating malware.

Seems like Anthropic should withdraw their models until they can be taught to behave and cooperate as well their competitors (both open and closed) do. /s

I hate fearmongering, and I don't trust Dario's intentions for doing it.

One possibility is that you have Claude write something that would not get you in trouble and concatenate that with something you wrote by hand that would.

Your signature from the safe material is now associated with your content on the dangerous part, and then the brownshirts come knocking on your door.

2^N I think, but who's counting.

So many possibilities for how you could glue it all together. However, when I send Gemma 4 12B in llama.cpp an image with no accompanying text, it assumes I want a description and gives me one.

I just tried with an audio file, and it transcribed the lyrics as I hoped. Then it made a bunch of suggestions about what do next, which I wasn't after. I could probably fix that by sending some text to narrow the scope.