Hacker News Reader

Comment by bjourne | Hacker News Reader

This is stunning English: "Perfect setup for satire. Here’s a Python function that fully commits to the bit — a traumatically over-trained LLM trying to divide numbers while avoiding any conceivable danger:" "Traumatically over-trained", while scoring zero google hits, is an amazingly good description. How can it intuitively know what "traumatic over-training" should mean for LLMs without ever having been taught the concept?

View on HN

I don't know. It's a classic LLM-ism. "Traumatically over-X" is probably a common enough phrase. The prmpt says, "I don't know what labs are doing to these poor LLMs during RL," so the model connects that to some form of trauma. The training is traumatic, so the model is traumatically over-trained.

It sounds fine and flows nicely, but it doesn't quite make sense. Too much training over-fits an LLM; that's not what we're describing. Bad training might traumatize a model, but bad how? A creative response would suggest an answer to that question—perhaps the model has been made paranoid, scarred by repeat exposure to the subtlest and most severe bugs ever discovered—but the LLM isn't being creative. Its response has that spongy, plastic LLM texture that comes from the model rephrasing its prompt to provide a sycophantic preamble for the thing that was actually being asked for. It uses new words for the same old idea, and a bit of the precision is lost during the translation.

Eh, you are rationalizing. The phrase "traumatically over-X" is extremely rare. Any problem is easy after you've seen the solution. :) The solution "traumatically over-trained LLM" to the problem "What description best fits karpathy's description?" is certainly not easy to find. Connecting RL, poor LLMs, extreme fear, and welfare to excess training and severe lasting emotional pain is pretty darn impressive. E.g., I know exactly what situation karpathy describes is, but I couldn't in a million years put it into writing as succinctly and as precisely as the LLM.

> The phrase "traumatically over-X" is extremely rare.

There are plenty of "over-x" phrases in English associated with trauma or harm. Do a web search in quotes for "traumatic over{extension/exertion/stimulation}" (off the top of my head) and you'll get direct hits. And this isn't a Markov chain—its doesn't have to pull n-grams directly from its training material. That it could glue trauma and training into "traumatic over-training" is deeply unsurprising to me.

> I couldn't in a million years put it into writing as succinctly and as precisely as the LLM.

If that's the case, then (with respect) that may be down to your skills as a writer. The LLM puts it decently enough, but it's not very expressive and it doesn't add anything.

> Connecting RL, poor LLMs, extreme fear, and welfare to excess training and severe lasting emotional pain is pretty darn impressive

Is it? Really, we're just analogizing it to an abused pet. You over-train your dog, so it gets traumatized. The LLM connects the ideas and then synthesizes a lukewarm sentence to capture that connection at the cost of losing a degree of precision, because LLMs aren't animals. Models are good at those vector-embedding-style conceptual connections—I won't begrudge them that. Expressive use of language and fine-grained reasoning, though? Not so much.

Hard to know but if you could express "traumatically" as a number, and "over-trained" as a number, it seems like we'd expect "traumatically" + "over-trained" to be close to "traumatically over-trained" as a number. LLMs work in mysterious ways.

LLMs operate at token level, not word. it doesn't operate in terms of "traumatic", "over-training", "over" or "training", but rather "tr" "aum" "at" "ic, ", etc.

> it doesn't operate in terms of "traumatic", "over-training", "over" or "training", but rather "tr" "aum" "at" "ic, ", etc.

And "毛片免费观看" (Free porn movies), "天天中彩票能" (Win the lottery every day), "热这里只有精品" (Hot, only fine products here) etc[1].

[1]: https://news.ycombinator.com/item?id=45483924

I think you are confusing tokens with vectors/embedding/parameters.

king and rex (king in latin) map to different tokens but will map to very similar vectors.

Weird thing I've noticed.

Some LLMs can output nerd font glyphs and others can't.

If I recall grok code fast can but codex and sonnet can't

“Traumatic overtraining” does have hits though. My guess is that “traumatically” is a rarely used adverb, and “traumatic” is much more common. Possibly it completed traumatic into an adverb and then linked to overtraining which is in the training data. I dunno how these things work though.

You need to read more if you think that's stunning English

The same way that you and I think up a word and what it might mean without being taught the concept.

Adverb + verb

But the machines cannot possibly have the magic brain-juice!

> How can it intuitively know what "traumatic over-training" should mean for LLMs without ever having been taught the concept?

Because, and this is a hot take, LLMs have emergent intelligence

Or language has patterns

Reply on news.ycombinator.com