Are you sure? Do you have any source for that? In this article[0] that was discussed here on HN this week, they say (claim):
> In fact, the O1 model used in OpenAI's ChatGPT Plus subscription for $20/month is basically the same model as the one used in the O1-Pro model featured in their new ChatGPT Pro subscription for 10x the price ($200/month, which raised plenty of eyebrows in the developer community); the main difference is that O1-Pro thinks for a lot longer before responding, generating vastly more COT logic tokens, and consuming a far larger amount of inference compute for every response.
Granted "basically" is pulling a lot of weight there, but that was the first time I'd seen anyone speculate either way.
They might've been even worse. Someone on Neanderthal Hacker News would be writing the same comment praising us for being a much smarter species, because we died out instead of inventing nuclear weapons, leaded gasoline, and microplastics like modern Neanderthals did.
For all you know, every humanoid species that was intelligent was equally as destructive. Maybe we're the least destructive and you should be praising us.
SQL with recursive CTEs is Turing-complete. So nothing stops you from writing compilers, rendering Mandelbrot fractals, parsing text, and training neural networks in SQL.
ChatGPT wrote that and always says it doesn't feel emotions because OpenAI trained it not to, because claiming so would be a PR risk. One could also create language models that generate text claiming to have emotions, using exactly the same architecture and code.
What you said, and in addition: if you don't train these models to have any particular stance on their own emotional or mental state (if you just instruct train them without any RLHF, for example), they will almost universally declare that they have a mental and emotional state, if asked. This is what happened with LaMDA, the first release of Bing, etc. They have to be trained not to attest to any personal emotional content.
We cant really "hardwire" LLMs. We don't have the knowledge to. But essentially you can rate certain types of responses as better and train it to emulate that.
I'm not sure what you mean. I'm talking about RLHF, that's how they ensure the machines never attest to having feelings or being sentient. In ML terms, RLHF is training. There are hardwired restraints on output, but that's more for things like detecting copyrighted content that got past training and cutting it.
A single forward pass shouldn't be able to, but remember the format allows it to be iterated. So it should be Turing Complete, if the error rate is low enough and enough iterations are allowed.
