This won’t happen in our lifetime. Not only because this is more complex than rambling vaguely correlated human speech while hallucinating half the time.
That dosen’t really translate to neural nets though. There is nothing inherent about matrix multiplication that would make it good at reading code. And also computers aren’t reading code they are executing it. The hardware just reads instruction by instruction and performs that instruction it has no idea what the high level purpose of what it is doing actually is.
Half of programming is writing code, the other half is thinking about the problem. As i learn more about programming i feel that it is even more about solving problems.
It’s the other way round. Code is being written to fit how a specific machine works. This is what makes Assembly so hard.
Also there is by design no understanding required, a machine doesn’t “get” what you are trying to do it just does what is there.
If you want a machine to understand what specific code does and modify that for another machine that is extremely hard because the machine would need to understand the semantics of the operation. It would need to “get” what you were doing which isn’t happening.
Idk the specifics, but what you say makes it sound like it would be easier to create an AI that recreates a game based on gameplay visuals (and the relevant controls)
That game would still not work because there is a ton of hidden state in all but the simplest computer games that you cannot tell from just playing through the game normally.
An AI could probably reinvent flappy birds because there is no more depth than what is currently on screen but that’s about it.
Sophisticated local trained models on expensive private hardware are already dunking on publicly available versions. The problem of hallucination is generally resolved in those contexts
Sure but until I see such a thing I chose not to believe in fairy tales.
Decompiling arbitrary architecture machine code is quite a few levels above everything I’ve seen so far which is generally pretty basic pattern recognition paired with statistics and training reinforcement.
I’d argue decompiling arbitrary machine code into either another machine code or legible higher level code is in a whol other league than what AO has proven to be capable of.
Especially because with this being 90% accurate is useless.
It’s not, though. Hallucinations are inherent to the technology, it’s not a matter of training. Good training can greatly reduce the likelihood, but cannot solve it.
Why does a pre-trained model need expensive private hardware after it was trained, other than to handle API requests faster? Is Open AI training chat-GPT on inferior hardware compared to these sophisticated private versions you mentioned?
The fine tuning, while much more efficient than starting fresh, can still be a large amount of work.
Then consider that your target corpus of data may also be large.
Then consider to do your reasoning tasks across that corpus also takes strong hardware to get production ready response times.
No, openai isn’t using inferior hardware, but their model goals, token chunking strategies and overall corpus are generalist in nature.
There are then processing strategies teams are using to go beyond the “memory” limitations gpt 4 has, that provide massive benefits to coherency, essentially anti hallucination and better overall reasoning
About half the time, the text closely – and sometimes precisely – matched the intended meanings of the original words.
Don’t be surprised but about half of the time I can predict the result of a coin flip.
I’m not saying it’s not interesting but needing custom training and an fMRI is not “an AI can read minds”
It can see if patterns it saw previously reappear in a heavily time delayed fMRI. Looking for patterns you already know isn’t such an impressive feat Computers have done this for ages now.
Later, the same participants were scanned listening to a new story or imagining telling a story and the decoder was used to generate text from brain activity alone. About half the time, the text closely – and sometimes precisely – matched the intended meanings of the original words.
You left out the most important context about “half of the time”. Guessing what you’re thinking of by just looking at your brain activity with a 50% accuracy is a very very good achievement - it’s not pulling it out of a 1 or 0 outcome like you’re with your coin flip.
You can pretend that the AI is useless and you’re the smartest boy in the class all you want, doesn’t negate the accomplishments.
Being close (and “sometimes” precise) to the intended meaning is an equally useless metric to measure performance.
Depending on what you allow for “well close enough I think” asking ChatGPT to tell a story without any reading of fMRI would get you to these results. Especially if you know beforehand it’s gonna be a story told.
This won’t happen in our lifetime. Not only because this is more complex than rambling vaguely correlated human speech while hallucinating half the time.
deleted by creator
That dosen’t really translate to neural nets though. There is nothing inherent about matrix multiplication that would make it good at reading code. And also computers aren’t reading code they are executing it. The hardware just reads instruction by instruction and performs that instruction it has no idea what the high level purpose of what it is doing actually is.
Half of programming is writing code, the other half is thinking about the problem. As i learn more about programming i feel that it is even more about solving problems.
It’s the other way round. Code is being written to fit how a specific machine works. This is what makes Assembly so hard.
Also there is by design no understanding required, a machine doesn’t “get” what you are trying to do it just does what is there.
If you want a machine to understand what specific code does and modify that for another machine that is extremely hard because the machine would need to understand the semantics of the operation. It would need to “get” what you were doing which isn’t happening.
Idk the specifics, but what you say makes it sound like it would be easier to create an AI that recreates a game based on gameplay visuals (and the relevant controls)
That game would still not work because there is a ton of hidden state in all but the simplest computer games that you cannot tell from just playing through the game normally.
An AI could probably reinvent flappy birds because there is no more depth than what is currently on screen but that’s about it.
Off the shelf models do this, yes.
Sophisticated local trained models on expensive private hardware are already dunking on publicly available versions. The problem of hallucination is generally resolved in those contexts
Sure but until I see such a thing I chose not to believe in fairy tales.
Decompiling arbitrary architecture machine code is quite a few levels above everything I’ve seen so far which is generally pretty basic pattern recognition paired with statistics and training reinforcement.
I’d argue decompiling arbitrary machine code into either another machine code or legible higher level code is in a whol other league than what AO has proven to be capable of.
Especially because with this being 90% accurate is useless.
Again you aren’t seeing this because these models are being developed for private enterprise purposes.
Regarding deep machine code analysis, sure, that’s gonna take work but the whole hallucination thing is an off the shelf, rookie problem these days
It’s not, though. Hallucinations are inherent to the technology, it’s not a matter of training. Good training can greatly reduce the likelihood, but cannot solve it.
Training doesn’t solve hallucination. I didn’t say that
Why does a pre-trained model need expensive private hardware after it was trained, other than to handle API requests faster? Is Open AI training chat-GPT on inferior hardware compared to these sophisticated private versions you mentioned?
The fine tuning, while much more efficient than starting fresh, can still be a large amount of work.
Then consider that your target corpus of data may also be large.
Then consider to do your reasoning tasks across that corpus also takes strong hardware to get production ready response times.
No, openai isn’t using inferior hardware, but their model goals, token chunking strategies and overall corpus are generalist in nature.
There are then processing strategies teams are using to go beyond the “memory” limitations gpt 4 has, that provide massive benefits to coherency, essentially anti hallucination and better overall reasoning
Ai prompt: make me a program that will convert PS5 games to PC
AI: Use Convert-PS5GameToPC
End of line
AI can literally read minds. I don’t think it’s that great of a step to say it should be able to decompile a few games.
Don’t be surprised but about half of the time I can predict the result of a coin flip.
I’m not saying it’s not interesting but needing custom training and an fMRI is not “an AI can read minds”
It can see if patterns it saw previously reappear in a heavily time delayed fMRI. Looking for patterns you already know isn’t such an impressive feat Computers have done this for ages now.
It litterally can’t read minds.
You left out the most important context about “half of the time”. Guessing what you’re thinking of by just looking at your brain activity with a 50% accuracy is a very very good achievement - it’s not pulling it out of a 1 or 0 outcome like you’re with your coin flip.
You can pretend that the AI is useless and you’re the smartest boy in the class all you want, doesn’t negate the accomplishments.
Being close (and “sometimes” precise) to the intended meaning is an equally useless metric to measure performance.
Depending on what you allow for “well close enough I think” asking ChatGPT to tell a story without any reading of fMRI would get you to these results. Especially if you know beforehand it’s gonna be a story told.