CamperBob2
CamperBob2 on Hacker News
OpenAI is widely understood to be next in line to go public.
It doesn't get much simpler than "Follow the money."
Some are suggesting that this specific technique only works with recurrent transformers:
https://nitter.netbub.com/AlexKrentsel/status/21173766837246...If so, it is something different from what you've been working on, since the models that support it are limited to the very newest releases from the big-two labs. Thoughts?
Claude, build me a copy of Oracle RDBMS. Make no mistakes
Elon Musk gave Trump $290M.
Greg Brockman, OpenAI's president, gave Trump $25M.
Dario Amodei gave Trump a hard time.
Do the math...
"Government doesn't work. Elect us and we'll fix^H^H^H prove it." - Republicans
Then again I should probably have checked the front page before saying this.
The frontier labs will make damned sure their models can't do this, if they haven't already.
I want to be able to play back the tape and pinpoint the source dataset + training method responsible for the failure.Not even remotely possible. Not with our current understanding of how these things work.
I'm not sure the comparison with Hilbert's dictum holds up in our field. Yes, Godel made a mockery of it, and mathematics is now known to be limited by incompleteness, but only at the theoretical horizon. Working mathematicians will not run into it as a problem that blocks their work in other areas.
In computing, though, the halting problem hoses us every time we have to hit ctrl-c to break an accidental infinite loop. Our tools can't always flag those situations ahead of time, so we often have to find them the hard way.
It seems clear that in software development, the specification is the program. To the extent there's something that the spec doesn't cover, that's a bug in the spec, not in the AI that turns the spec into code. Just as you wouldn't blame a compiler for your infinite loop, we will eventually reach the point where blaming a clanker for generating imperfect high-level code sounds just as bogus.
"Build me a Super Mario Bros. clone, make no mistakes" is the new hello_world.c. For anything innovative -- anything with commercial value -- the spec will have to be a lot more elaborate. There are plenty of specs that are already too complex for any one person to hold in their head, except perhaps for a few godlike prodigies. Look at the MPEG-4 document collection, or for that matter our tax and legal systems. I think the lesson is that language alone can't save us from complexity, any more than a better set of axioms can save us from incompleteness.
You know the LLMs that are finding all these CVEs lately are finding them because they are better than the people who wrote the original code and better than every one of the people who ever laid eyes on it.
An LLM doesn't need a memory-safe language. It needs one it knows like the back of its metal hand.
Sounds like somebody else's problem to me. When I want to see the chain of thought, I use a local model. It's seldom enlightening.
Awesome. Next stop Kardashev Type III, I guess!
Who cares if the thinking traces are bullshit? They are not the work product.
This is inherently my whole problem with the "AI utopianists", who somehow think we'll all be happy once we're sitting around with no disease just eating bonbons all day or something.I don't know about Gauss, but when I am given a more powerful tool, I use it to tackle harder problems, the kind that were out of reach before.
Call me when you run out of those sorts of problems. I'll propose some more.
i don't know how more people don't experience this, but staring at an llms output be confidently wrong eventually to get there in the process, or looking at the commands its running and see some really stupid things its doing drives me crazy to the point i will just stop it before it answers, i cannot handle something being that confidently wrong (even though it'll probably still get there in the end)Let's talk about being "confidently wrong." I don't know if the original author is reading comments here, but if so: how long has it been since you had the experience above?
Have you noticed it happens less and less often as time goes on? And that the models take less time and fewer trials to "get there in the end?"
Because the rate of change is the important part, not the state of the art at the particular moment in time when you formed your opinion.
An interesting question is just how much high-level programming language development we'll see going forward. What is the incentive to create new languages? They're all just object code to an LLM, so when starting a new project, you might as well go with an old-school language like C++ or Java or Python or JavaScript that is well-represented in the training distribution.
Why do we need Rust, for instance, when LLMs know how to use C++ without shooting themselves in the foot? If that's not entirely true today, it certainly will be tomorrow.
See also the lack of new instruction-set architectures in the CPU business. Why do we need more, in the absence of fundamental and revolutionary hardware innovations? Even RISC-V is just a rehash of MIPS.
How the heck did the model figure out the generation recipe? That's impressive in itself.
My problem is, "unbridled capitalism" is not the same thing as "Infiltrate the government and turn it into your hatchet man."
The Reddit mindset is to pile all ills related to labor and trade into one anticapitalist bucket. I don't want to see that kind of stupidity around here. Giving government more power is not going to fix problems caused by inappropriate leverage of government power. In many cases, the problem is the fact that such power exists in the first place.
Just gotta make sure that pentagram is rock-solid!
No, people using AI are doing illegal things with it.
4x 96 GB runs the NVFP4 quants, which are indistinguishable from Q8 in real-world performance for this particular model. That's not always true but DeepSeek is unusually quantization-friendly. Also remember that a big chunk of the model consists of engrams that can live in system RAM or even on an SSD if necessary.
I imagine the various GGUFs are still pretty impressive, but haven't tried them myself.
More like, the ability to chain a demon in your basement that does what you tell it and can't be rug-pulled by his bosses in San Francisco.
The main barrier to wider spread adoption of this right now is that very few people have the legit on-prem hardware resources/budget to run something like an abliterated Q8 of deepseek v4.1 flash 1017Not as bad as you might think. This costs about as much as a decent new car nowadays.
The abliterated 1017 release I've seen is crazy good.
They are meant to block abusive mass onslaughts from armies of bots, not agents of individual users acting in good faith.
Nothing great was ever built by someone who was sweating the reviews. Fuck 'em, do what feels right to you.
Anything he doesn't like is capitalism. And we're all Redditors now, don't you know?
Failure to take advantage of available tools does not suggest whether you are a good guy or a bad guy, just whether you are a competent guy or an incompetent guy.
Have fun keeping this up to date.Why? These rules are written down somewhere. Looks like the spec was created by pointing the model at irs.gov and/or the Federal Register itself and telling it something like, "Create a specification for a verifiable tax model and the tools to verify it."
Next year, all the maintainer has to do is issue the same prompt, no?
The statistics aren't backing you up on this.