kescusay

It's already happening. Software instability is becoming a major problem.

Honest answer? Not much... yet. As the stupidity of "tokenmaxxing" and replacing developers with hallucinations-and-plagiarism engines really comes into focus, I expect developer hiring to ramp back up.

Meanwhile, in the real world, my company has locked down usage of Anthropic's top-tier models, because costs were out of control. Even the low-end models are usage-capped, and developers are starting to write code by hand again.

And we're busily setting up local models we control ourselves on our network edge. We'll probably end up with a modest budget for frontier models (until the model providers collapse), but most of the coding will be done by local models and developers themselves.

And that seems to be the way the entire industry is going, unless Anthropic starts subsidizing tokens with investor money again.

In conclusion: Get fucked, Dario.

As I've mentioned elsewhere, not if by "information" you mean semantic content that a mind can process. What they have are vector fields (essentially just numbers) with statistically more or less likely relationships.

If I say, "take me out to the ballgame" to an LLM, the tokens representing the words in the next verse of the song are statistically "close" in the vector database, so it's likely to generate them. But that doesn't mean it actually knows the lyrics... or even has those lyrics recorded in a regular database anywhere.

That's why they hallucinate. The model determines that the next token is something nonsensical, but it has no way of understanding that it has made a mistake. In a sense, it actually hasn't made a mistake. It's done exactly what it's designed to do. It's just that in the case of hallucinations, its output isn't useful.

They're information, but not the same information that was used to create them.

No, they really don't. That's not how they work. At least, not if the "information" you're talking about is real semantic content that real minds can process.

Every piece of information you think an LLM has access to is actually just converted into a stream of additional tokens that are fed into the model to (hopefully usefully) modify the next tokens it predicts. That's not the same thing as having actual access to information. Tokens are just numbers with statistically more (or less) likely relationships to each other.

I'm not trying to downplay LLMs. They're architecturally interesting and have genuine uses. I'm just trying to head off a bit of technical inaccuracy.

The important thing to remember is that it actually has zero access to information, because that's not how LLMs work.

At their core, they're vector databases, and they're trying to probabilistically come up with the next most likely token in a stream of tokens found in the DB. You can manipulate the stream by injecting text such as the content of existing files (which becomes more tokens) into the stream, but it never actually understands any of it.

That's why hallucinations are inherently unavoidable. It's really all just hallucinations. It's just that you can sometimes get useful text from their hallucinations if they happen to comport with reality.

Jesus Tapdancing Christ...

Isn't this the breach that was widely reported on in March? And they're just now disclosing the scale of the exposure?

Ugh. If this makes my Linux systems less stable, I don't even know what the fuck I'm going to do. I won't go back to Windows, I'm not buying Macs (which have their own software problems anyway), and my hardware probably won't work with any BSD variant.

We're sad because a lot of people have bought into overly-hyped bullshit factories made by assholes who claim they can replace us.

They can't, but our employers are currently too uninformed (or addicted to the bullshit, or desperate to prop up their stock prices) to realize that the bullshit factories aren't capable of replacing us.

That's causing a lot of turmoil in the form of unnecessary layoffs and rehires, worsening software quality, and general job insecurity.

The difference here is that a cheap AI chip won't fix the fundamental software problems with LLMs. We might reach a point where they can produce output faster, but as long as what's actually going on is probabilistic next-token prediction in a static vector database, that just means faster mistakes as well.

There's an odd psychosis going around where people become convinced that actual AGI can be derived from this technology. People who should know better just shut their brains off when it comes to token prediction, because they've had very compelling "conversations" with the predictor. They forget that the actual model is static, has no internal state, and doesn't even "remember" what you've said to it.

What it has is a context window, and your entire conversational history - both what you've said and how it has responded - gets shoved into that window when you interact with it. (Or depending on the chatbot harness, saved in "memory" files that it can retrieve when the context contents indicate that would be useful.)

That's why the bots seem so weirdly forgetful one moment and like they've got photographic memories the next. Stuff that is in the context window and has its "attention" will influence the tokens it produces, but whether or not the right things are in the context window and it's including them in the token prediction is a crapshoot.