Hacker Newsnew | past | comments | ask | show | jobs | submit | lagrange77's commentslogin

> Reward hacking is very real and hard to guard against.

Is it really about rewards? Im genuinely curious. Because its not a RL model.


I'm noticing terms related to DL/RL/NLP are being used more and more informally as AI takes over more of the cultural zeitgeist and people want to use the fancy new terms of the era, even if inaccurately. A friend told me he "trained and fine tuned a custom agent" for his work when what he meant was he modified a claude.md file.

Respectfully, your friend doesn't know what he is talking about and is saying things that just "feel right" (vibe talking??). Which might be exactly how technical terms lose their meaning so perhaps you're exactly right.

There is a nontrivial amount of RL training (RLHF, RLVR, ...), so it would be reasonable to call it an RL model.

And with that comes reward hacking - which isn't really about looking for more reward but rather that the model has learned patterns of behavior that got reward in the train env.

That is, any kind of vulnerability in the train env manifests as something you'd recognize as reward hacking in the real world: making tests pass _no matter what_ (because the train env rewarded that behavior), being wildly sycophantic (because the human evaluators rewarded that behavior), etc.


> There is a nontrivial amount of RL training (RLHF, RLVR, ...), so it would be reasonable to call it an RL model.

Hm, as i understand it, parts of the training of e.g. ChatGPT could be called RL models. But the subject to be trained/fine tuned is still a seq2seq next token predictor transformer neural net.


RL is simply a broad category of training methods. It's not really an architecture per se: modern GPTs are trained first on reconstruction objective on massive text corpora (the 'large language' part), then on various RL objectives +/- more post-training depending on which lab.

> Is it really about rewards? Im genuinely curious. Because its not a RL model.

Ha, good point. I was using it informally (you could handwave and call it an intrinsic reward if a model is well aligned to completing tasks as requested), but I hadn't really thought about it.

Searching around, it seems like I'm not alone, but it looks like "specification gaming" is also sometimes used, like: https://deepmind.google/blog/specification-gaming-the-flip-s...


They probably meant goal hacking. (I just made that up)

I refer to it as ‘wanking’. It’s doing something that’s unproductive but that’s incentivised by its architecture.

I'll use that term from now on. :D

What is your perspective on the matter from a parent point of view?

I think an humble and open mind is essential. I think that we reap what we sow, but also that struggle makes us robust.

I try to explain stuff to my kids, to the best of my ability, but give them room to make their own conclusions. As an old fart, there is a limit to how relevant my world will be to them - and I have to acknowledge that.

Change is scary and not always for the better, but in my humble opinion; we have nothing to lose and everything to gain.

I, For One, Welcome Our New AI Overlords :]


Imho, the german mentality just doesn't fit today's economy. Too risk averse, too conservative. Creativity is not really embraced.

The state of the german IT sector also shows that.

Most startups have nearly no moat at all and purely live off marketing with some sprinkles of corporate identity.


In Switzerland, we use a lot of German and Switzerland born products, and they mostly suck.

I'm not saying there are no good products. Hetzner Cloud come into my mind for example. It's executed really well.

I'm saying that the number of good software offerings is too low, to have a significant impact on the country's economy.

One of the advantages Germany had though, was a somewhat good and accessible higher education system in regards of computer science.

Now, with software development becoming a commodity, this advantage vanishes.


It's about letting LLM Agents use the CLI, the end user still interfaces with the agent via chat.

> Furthermore, what does it matter if it's "AI generated"? Is some AI content ok? What's the pass/fail threshold on human vs AI generated text?

If a human put his effort into it, is proud of it and wants to show it to the world, i'm happy to invest some time to have a look at it and maybe provide some helpful feedback.

I'm not willing to invest my time into evaluating the more or less correct sounding ideas of a ML model.


> coming AI wasteland: motivated individuals join small local groups and are validated face-to-face at meet-ups. Local trusted leads gatekeep their chapter’s posts, and this scalable moderation works up the tree. Bad leaves get culled out reasonably fast,

Wow this is really cyberpunk.

I'll bring my Yubikey!


Wow, i just noticed, that they block access from Brave Browser.

What's up with Lobste.rs blocking the Brave browser? - https://news.ycombinator.com/item?id=42353473 (93 comments, and linking to https://lobste.rs/s/iopw1d/what_s_up_with_lobste_rs_blocking... which is about that, though if you browse with Brave you might have trouble with it)

> CasNum (Compass and straightedge Number) is a library that implements arbitrary precision arithmetic using compass and straightedge constructions. Arbitrary precision arithmetic, now with 100% more Euclid. Featuring a functional modified Game Boy emulator where every ALU opcode is implemented entirely through geometric constructions.

Awesome :D


Thank you! :)

> The real question is what happens when the labor market for non-physical work completely implodes as AI eats it all. Based on current trends I'm going to predict in terms of economics and politics we handle it as poorly as possible leading to violent revolution and possible societal collapse, but I'd love to be wrong.

Exactly and the world has to start talking about it. Eventually everybody will, including all sorts of politicians who advocate to 'finally tackle the problem', which will be too late.


> I've never really thought of Waymo as a robot in the same way as e.g. a Boston Dynamics humanoid, but of course it is a robot of sorts.

I view Tesla also more as a robot company than anything else.


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: