Notices by cthos 🐱 (cthos@mastodon.cthos.dev)

@xgranade What if .... , but natural language!!?? Seems to be a startlingly common pitch.

I saw a couple of takes Re: codeberg that weren't quite "oh I don't like LLM code but codeberg shouldn't decide that they're unacceptable, they're neutral hosting!"

No, they absolutely can and should do that. Github can drop you for whatever reason at any time, folks. It isn't a publicly-owned service.

If you want different moderation go to Gitlab or back to Github or run your own Forgejo! You have options.

@thomasfuchs Re: Ghost, is there a new one or do you mean the one back in May? (I'm not seeing a new one and am worried I've missed something I need to patch right now).

@pixxl @thomasfuchs But it does, by weighting the probability of the token it will select at any given step towards 100% (at 0 T): https://techscribr.github.io/posts/llm-temperature/

But I don't think the average person who knows what an LLM is would agree that "the LLM algorithm is just the bit that spits out vector weights" when every implementation out there includes the sampler.

@thomasfuchs LLMs are stochastic (not deterministic) algorithms, they don't always produce the same next token for the same input (unless you set the temperature to 0).

Edit: but the stochastic bit is at the very end where it uses the temperature setting as a weight to determine which of the most likely tokens to output based on randomness.

I wonder if the "we are deffo going to have practical quantum computers by 2029" thing survives great depression pt 2.

@ireneista @xgranade Microsoft, IBM, and Google have all made some variant of the claim that they will have a functional QC datacenter by 2029, which sounds like CEOs setting up a bubble to me but who knows.

(I do also believe Sophie used the same date in the Google PQC deadline)

RE: https://hachyderm.io/@thomasfuchs/116728764235644110

This is Ellipsus which touts itself as "A principled alternative—by writers, for writers" which is really ridiculous when you put surveillance tools into your editor because LLMs have ruined everything.

And like, I hesitated (for about 20 seconds) yelling about this because of the inevitable "well if you're against this you must be using LLMs to write" which is just AHHHHGHHHH.

Okay IRC without the need for a bouncer is pretty sweet.

@mattly I somehow missed that Anthropic is also switching some flags on Claude code to token-based billing very soon too so yeahhhhh I think a lot of vibe coders are about to get a surprise bill.

@mattly I love that one because it totally neglects the fact that the "best models" run in a loop to do "reasoning" and chain themselves together to "error correct"

@mattly Or the variant "But they're charging less per million now!"

RE: https://social.lansky.name/@hn50/116590156596542536

You should just read some of the comments in there. There's a lot of ... let's say, interesting, reactions to the post.

Hottake

Google's latest push with Recaptcha to require you to scan a QR code with a smartphone is just another push towards "you need a modern smart phone that is run by the duopoly to participate in society".

Passkeys evoke a similar concern / anxiety, but they're not the same - there are several escape hatches.... but if you own no computing devices at all and rely on borrowing one from the library, for example, you're going to have a hard(er) time.

So uh, is it a "widely accepted sentiment" that agentic coding is the future of software developers?

Like....I have seen this assertion made, and certainly a lot of prominent dev *influencers* think this, but I am not convinced it is the majority opinion. (Especially because genai "mandates" are still necessary)

@mattly The people in question were in fact arguing that they should not have to pay a billion dollar corporation to do this and thus "small language model".

But like... it's right there! The tools exist.

This might actually be worse than I was thinking and is in fact "smol language model please fix the errors my LSP is detecting" which.... yeah.

*rubs temples*

...there are a non-zero number of devs out there who think LLMs are magic because they never bothered setting up their deterministic tooling, aren't there?

(this is wild speculation on my part but I just encountered "use a small language model to fix small code errors" in the wild and I'm just... what?)

*hovers over the "fix all autofixable errors button"*

no, no I should ask a non-deterministic language model to correct this for me.

So I just blocked a mutual for writing a blog post about all the things they've been doing with Claude Code. Normally, I would let that slide, but the thing that got me was "yeah, I recognize the ethical concerns, but I think we'll fix those".

While also making an informational website with a big "please check sources" disclaimer up at the top.

I just... why! Why do *you* think that the slop machine needs to create informational resources but not FACT CHECK them.