Ya, I'm not super impressed with their ability to generate code- but their ability to infer the meaning and actions of some computer language it's looking at is really impressive. That being said, I absolutely do think vibe coded things have their place in this world- like your CounterStrike mod, or various tools I build for work to do silly things like manage a database of configurables for an order system. The order system itself is entirely hand-built, and I haven't seen anything that would make me eager to hand it off to an LLM- but the accessory tooling to make it so other people can make it do what they want without bugging me? Vibe away.

Back to reverse engineering, disassembly -> pseudocode, or C is always a bit hit or miss in terms of usefulness in reverse engineering. It usually takes a lot of work to turn that into something that actually makes sense, and is actually structured like the real code would be, before compiling. Even small open LLMs are ridiculously good at turning a disassembly into very well annotated C.

So, what I tested was targeted.
I didn't say, "Mr. Wizard, find me all the vulnerabilities in this software!!"
I generally prepped it with what I knew at the time when I went looking.
"This eMMC is write protected. The only way to reverse that is to power cycle it. The firmware does this somehow, or it would be unable to update itself after a warm boot- which it does."
Or,
"I suspect this vendor firmware has a backdoor built into it. See if you can find it."

Of course my past work is part of its training corpus one way or another, but I mean, the thing does show its work. REA in particular focuses on forensics- i.e., showing evidence for all of its conclusions, including disassemblies. It's frankly pretty fucking wild.

Forget the "pattern matching" thing.
LLMs can read language. Assembly is just another language.
It took me a long time to work up the skills to be able to eyeball a disassembly and know what is going on without thinking hard about every line. I don't know if it has to think about every line hard or not- and I will say this- when I did my Qwen tests, one of them took a solid 14 hours to run. But it was successful in the end.

LLMs can read your requests, they can read code, and they can read a disassembly. It shouldn't be surprising that if pointed at something, they can find it.

I have a few CVEs under my belt, and a few private jobs targeting specific hardware that were never published.
I threw an open-source LLM (Qwen3.8 27B@bf16), and Hermes for the agent, along with REA of the reverse engineering tool bridge... and it was able to find every one of my CVEs and unpublished vulnerabilities.

I'm a professional at this. I've had my hour or so of internet fame for devices I've broken into.
I have no idea what it is you think you've seen, but even small models with good tooling are very good at this.

It's complex and muddy waters.
In general you have a right to reverse engineer something, as long as it's done for a fair use reason... but also not always.
The DMCA does make it not-ok to circumvent any protection measures, regardless of reason, unless the Librarian of Congress has a specific exemption for you, or you fall into one of its statutory exemptions.

Now, arguably- "for fun", is more or less protected.
However, in no instance would be publishing your findings be protected- that's just violation of copyright and/or the DMCA if you circumvented any rights protection mechanism.

Also, in the US at least, there are no treble damages for copyright- willful infringement is 5x, or as low as $200 if the Judge things the behavior was "innocent".

Surely you can't be dumb enough to believe that.

Not remotely analogous.
Your case has mens rea, this does not.
They didn't shoot a gun at someone. They left the gun unsecured on the kitchen table, again, and bad shit happened.

Seriously, you're on time-out. Read this and give me a 100 page essay before you post anymore.

*OpenAI using their LLM

I think we must have gotten our wires crossed.

I'm referring to AI using their LLM in a lab unsafely, and then using the law breaking behavior it engages in as a press release.
The first time- I'd argue it wasn't anything except a faux pas and a learning experience.
However, they keep saying this keeps happening- and that is absolutely prosecutable.

If you act in wonton disregard of people or property, that is legally reckless behavior.
The first time- they can argue, "they didn't know it could happen". And that's fair. That's not negligent or reckless.
After it has done it 47 times, it is.

It's similar to your pitbull escaping and biting the neighbor.
The first time, if it looks like there was no negligent behavior on your part- it burst through the fence- then you'll be civilly liable, but not criminally.
The 13th time, with 13 patched holes in the fence- you are, indeed, criminally liable for reckless endangerment.

It is when certain people do it. Absolutely.
But the average person? Na.

Nobody said pets didn't have emotions- but their emotions are *far* different than yours. If you don't understand that, it's because you anthropomorphize them.
I remember the day that clicked for me- a professional dog trainer said, "You need to stop trying to explain their behavior by putting yourself in their shoes. They're not humans. You need to learn their behavior objectively."
So no, it's not misleading. Really, you've just proved my point- you yourself anthropomorphize pets.

lol. Uh huh.

I point out the problem with your narrow reading of the word, and what do you do? You continue to use it narrowly as if nobody pointed out it was problematic.
Judgement is one thing, and one thing only- it does not require values, and it does not require a human brain.
Judgement is the evaluation of a situation, and the capacity to make a sound decision.
If you're bothered by the word decision, then see above about evaluating the system from the outputs.
I don't get why you're so stubborn about this. It just makes you look dumb.

See above about us debating whether a simulation is the same as the real thing. What matters here is the evaluation of the outputs of the system.
You have an emotional sensitivity to models being described in anthropomorphic terms. Scientifically speaking, that's more of a problem with you than with the people who do so. They're closer to "right", than you can ever be.

If you keep restating this, does it keep your insecurity at bay?
It's not logically relevant to the discussion- it's just a coping mechanism you're using.
The fine-tuning processes that are used here do not trend toward sociopathy. That's just ignorant.

This is the stupidest thing you've said.
"True- except [let me demonstrate that I didn't actually understand what you said]!"
Are you really claiming that there is no functional output that can approximate human output driven by chemical emotions and our mind's perception of value, as it is encoded in our neurons?
Surely you aren't that fucking dumb.

I'd have to eradicate half my brain cells to "get" that point, because it's logically stupid.
The fact that you don't get that is astounding. Go stand in front of a mirror, and keep saying it until you realize how fucking dumb it is.
You are asserting, with a straight face, that human output cannot be replicated, because whatever thing is replicating it.... doesn't have values.

AI can, and does, respond like a human- in spite of having no empathy, or being a social creature.
Everything in this universe can be approximated- even humans.
To think otherwise is idiocy.

LLMs undergo what is called alignment training.
It is the encoding of our "shared values and empathy" into the weights of the model, so that the approximation of human behavior that it undergoes is "aligned" with what we consider good and moral behavior.
AI is not some opposite-of-human oracle, you fucking fool.

Your assertions are religious poppycock, and your grasp of basic logic is elementary at best. I had forgotten that about you. I apologize for trying to help you be less stupid.

Those who have drank the kool-aid seldom realize they have drunk the kool-aid.
The hint that you have is that you think you've found the right way to hold it, and refuse to accept that even if you hold it as such- it simply produces code of high cyclomatic complexity, opacity, and flat out bad ways of doing things- that is, you deny that the benchmarks apply to.... you. You are special.

Negligence and recklessness are legal concepts that have absolutely become relevant.
They keep letting this happen, so we're past the point of grace.
If it weren't a company- if it were an individual doing this- they would have been prosecuted.

Ever get tired of beating on that dead horse?
Humans anthropomorphize *everything*. We always have. Half of us fancy our pets as little humans in a funny body. We attribute to them human emotions, and try to explain them with human reasoning.

You are smart enough to take the anthropocentric terms and apply them as applicable.
Models don't have "judgement", if you use that word anthropocentrically.
But they absolutely do have an analogue to it, that is a long-winded functional result of their weights.

Your conjecture that one cannot understand the concept of "harm" unless it has "values" is of course patently fucking absurd though. I may take back what I said about you being smart enough to handle the conceptual replacement required here.

A model can "understand" value as well as any computer program can be programmed to "understand" value- i.e., to react accordingly as if it had value.
We can debate until the sun sets about whether or not "simulation of a thing is the same as a thing", but ultimately, if the output of the model is the same as the output of the real thing, that debate is largely pointless- see Turing's Computing Machinery and Intelligence, except in larger philosophical discussions that just aren't relevant here.

Companies like OpenAI do indeed need to be hit for negligence and/or recklessness for what they're doing, we agree there.

You are absolutely correct in that regard.
The ultimate bottleneck on an RDMA-linked setup is how quickly it can reduce the multiple sets of activations, which is overwhelmingly dominated by the RTT.
The total bottleneck to tokens-per-second is going to be RTT*layers, as the AllReduce will be effectively free, time-wise.
Fortunately, on any higher end Mac, that's going to be a smaller delay than it takes to stream all of the active weights per compute unit- i.e., it's not really a bottleneck. 2 MBPs scale very close to linearly.
You wouldn't want to connect 2 H200s to it by any means.

I do broadly agree with you.
Particularly if we further refine it to something like "trendslop".

Programmers have been doing trendy piss-poor engineering for a long fucking time, now.
XML as a data interchange format still blows my fucking mind.
It's like someone said, "well, hardware is cheap- so we don't really need to worry about moving data efficiently... In fact, let's see if we can burn down as many trees possible while doing it."

At least JSON can be somewhat serialized into a real machine structure, and I think it is generally easier to read... though frankly, for large amounts of data, it's still pretty fucking wasteful- but at least it's very easy to parse. Also, if you've got a bit of flexibility (I do) then you can pretty quickly derive a nicer JSON format that allows comments, doesn't require quotes, treat all commas as purely optional, is just as easy to read, no slower to parse, and won't shit its fucking pants because you accidentally added a comma to the last thing in that list.

I've been using my "liberal JSON" as a configuration file format for a while. Allows structure, and is very hard to accidentally break.

I think you've just drank the kool-aid. I don't mean that as an insult. I think probably a large portion of us have. Or maybe, have been beaten into submission... hard to say, really.

Ultimately, there are benchmarks for antipatterns and code erosion, and other things we'd call "slop".
There isn't a reliable way to get the models to quit doing very very poorly at those benchmarks. That's just a fact.
If you think you've got the magic sauce, you should head over to SlopCodeBench and submit some patches.