xpct

I'd be interested to know whether it's possible to contaminate public data to make trained models do this.

Something like a commonly repeating conditional in the public domain embedding itself in the model's logic.

I think it's the saddest thing that the X-articles thing has somewhat caught on there. Some of it is slop, sure, but people started hosting their longer form opinions on X instead of their own blogs, which is unfortunate.

Interesting. You don't feel like you were building an edge with your understanding of said systems or OS?

I'm very interested in systems as well, but being less sure about the future (on whether this is something I really need to think about, or whether I'll get opportunities to work much at this level), I started reading about such topics a fair bit less.

With how you presented this, it makes it sound like it was either a toy codebase or something outside of your domain already.

I don't think this statement implies they meant 'single files'? They could be talking about complex systems that they could fit into their head, and now they can't do that with new workflows.

Oh, that's interesting. Mind sharing how much was 'too much'?

I think people said that Claude refuses to work on RE sometimes.

I've had no issues with OpenAI.

I don't understand this sentiment. This whole thread is quite positive towards the findings. What we have is a societal verification problem where work has become hard to evaluate and verify, so everyone is being extra careful with their judgements.

I don't know either. My only guess is that this has some helpful context for less powerful models, maybe.

We're kinda far into this LLM thing, maybe it's time to start selling harnesses by leading with how some examples were solved faster with this and how it saved tokens, or similar?

I find it a bit comical. AI tools are so easy to pick up, and it's unlikely you've pushed some highly critical features which couldn't have waited for a few months.

I experimented with something similar with LLMs. They can kinda do it for some stuff.

More interesting, you can ask a vision model to color parts of speech in img2img and it works OK for frontier models.

Nothing fancy. What works for me is trying to lead the conversation: asking for definitions one at a time, asking how they differ from its previous answers and pointing it out when it conflates its own answer. I generally feel like you can't let it dictate the pace, it's not very good at that yet.

I used to edit my prompts and undo messages if it misunderstood something but that's been failing me recently.

I've had the most success with GPT explaining things to me by making it take a few sentences at a time back and forth, instead of reading full write-ups of whatever I asked. It also often poisons the conversation if it misunderstood some part of the question, and I can lead it better by continuously questioning its statements. It's also more engaging that way.

I've been learning music lately and it kept re-pasting the same one chord visualization throughout many conversations, almost randomly and often barely related to the question. So I at least hope this won't be as aggressive so I can prompt it away!

I really don't like having to open Work sessions for one-off questions, only because they're limiting what models they put in the chat. The older models are just too dumb for some things.

> Or do you mean that you're willing to forgive more superficial things like meandering or excessive formality if the post has other redeeming qualities?

This! Well, kind of!

I definitely do some type of screening for pieces that I read: I may look up who the author is or what they've worked on; the piece may have been recommended by someone else I respect on social media; the topic itself may have little other writing on it online which signals that it may contain original thought, it may have been up-voted on HN and had interesting comments, etc.

That is to say, I try to evaluate whether it's worth my time reading the article in full, even as I start reading it. I do have a habit of saving URLs of things I read, and typically jot down a few personal notes in a local .md as I read along.

I do use LLM writing as a negative signal: it could be that the author hasn't spent that much time thinking about the issue, and I can spend that time reading something else from my reading list. But there's definitely been a few cases where I read pieces in full, even though they were clearly heavily LLM assisted, only because the material just seemed worth tanking through for.

Perhaps, rephrased: I seldom drop pieces because of the prose, more often I do so because it lacks substance. And, to add, it could entirely be my selective process that leads me to dropping articles less!

I find that I'm actually not that picky when it comes to reading technical material, at least in blog form. There's very few pieces I dropped because of how they were written.

I didn't mean to oppose OP's point, I just found it interesting as a non-German speaker!

I just did a quick search on this and apparently the misspellings are German surnames as well:

https://en.wikipedia.org/wiki/Reimann

https://en.wikipedia.org/wiki/Reinmann

Presumably they don't because they're training the audience (us) to trust the machine, not its verifiers, even if they were included.

In other words, you've already taken a gambit with the first half of your PhD, now take a second gambit, praying that you have something to publish by the end of your PhD.

It has to feel awful to be in this position.

Just to be very clear: they aren't asking for permission, they are framing it that way because of the bad press.

There's no gatekeeping here!

I don't know why, but I personally find Mistral's marketing strategy much more appealing than that of other companies.

For example, there's something about Anthropic's picked design and their little Claude avatars that's unsettling to me.

At the very least, it will be a binary that hasn't yet been consumed by the LLM, so it's an interesting test!

Well, I meant it in the sense of representing a game, and how NN bots look at embeddings whereas we look at pixels on the screen. We can try to model the game to have similar rules and information to bots as to what is accessible to humans.

For 1., the bot can compare position of its units between frames (this is already done by recurrent memory), and how they were displaced. If a bot's input is constrained to what we see on the screen, this would be very similar to what a vision model would do.

For 2., I imagine this is where bots have the supernatural ability of counting dots on the minimap, unless we limit its accuracy or decrease frame rate of how often they take a snapshot of the minimap.

But I appreciate your points! I had genuinely forgotten about cloaked units as I'm not an active player.

Six months is a long time! I was honestly prepared to see more aggressive timelines when I wrote this comment :)

I think GTA6 will be the real test for how capable these agents are. Some of these mods use well-known emulation methods, and some of these binaries have probably been ingested by the LLM numerous times.

I don't mean to discredit LLM capabilities, it's just that these examples are heavily favored by what's already documented online.

I find it hardest to sympathize with the guy who inserted himself with 'Adapt or Die'..

That's definitely something I'd consider if I had cash to spare for H200's! Unfortunately I think for most of us the price of self-hosting has to be 2-5x lower still.

Well, if your training includes regularization against ruleset changes, the model should simply handle it. (that would be the expensive option, and require vastly more training)

When the Dota 2 bot was made, they retrained the bot only partially when new patches came in, so it was definitely cheaper to adapt.

I love this, and love listening to how much he's connected to the thing he made. This should remind us that not everything is done with an end goal in mind.

If you're unfamiliar with Minecraft, the build he's showcasing isn't actually advanced when it comes to building. Building in Minecraft is almost an art form, and many techniques have been developed over the years. This is because you have limited control over what you want to express: you are operating at a block granularity (more or less) and a limited number of block types that give blocks different textures.

A tech savvy teen could probably write a Java mod that generates something similar to what he's done, even prior to LLMs. An LLM could take a Google Streets feed of NY and replicate it in-game, etc.

And he wouldn't care! The thing he's made is his own, made under his own constraints, shaped by his own mind.

Perhaps a lesson of meaning in the LLM era?