Comments URL: https://news.ycombinator.com/item?id=48431827

Points: 9

# Comments: 2

Then the next question becomes "HOW do they predict the next token?" There are many ways that can be done, why is this particular algorithm so GOOD?"

When people say "We don't understand how LLM works" isn't it really saying we don't understand how this specific algorithm used to predict the next token works? No, it is not, because "we" do understand how all those algorithms work there are many descriptions of them available.

So the question then really is "Why is the prediction this algorithm makes, so good, as compared to some other statistical algorithms?"

It's not about "Why does AI work so well?". It should be "Why does this particular XYZ algorithm work so well?"

So what is it that we don't understand about why theyr work? The algorithm? We have the code. Why the specific algorithm makes such good predictions? I see it as a generalization of trying to predict who wins Kentucky Derby.

It is the combination of LLM and agent-harnesses that make it look really smart. Agent-harness is a programmatic device that lets us tap into the vast knowledge in the LLM.

It is probabaly true that many TV-commentators fail to appreciate this fact and therefore think LLMs are super-intelligent. No, it is the combination of LLM and the programmatic agent-haness that is the breakthrough.

An interesting thought is that the LLM could in theory code the agent-harrness, start it running every time we interact with it. Currently the agent-harrness I think is pretty static I think. In theory it could be dynamically created for every task. Would that make it better don't know.

https://en.wikipedia.org/wiki/Molten-salt_battery

I also gave some presentations in IT-related conferences. People in conferences often get tired. So once as I entered the lectern I turned on a piece of music I had on my thumb-drive. It was the loudest hardest Rock'n'Roll I had. The audtorium had good loud-speakers. I think it woke the audience up. But soon I saw them nodding again. :-)

Comments URL: https://news.ycombinator.com/item?id=48417450

Points: 4

# Comments: 0

But I don't want to souond too pessimistic, there are lots of good people and progress is inevitable. AI gets put down a lot but I think it will make people much smarter.

Maybe briefly show how that adavanced topic will be taught and let them realize they can not possible even start to understand advanced topic because they are missing the more elementary pieces.

Similarly why they can't got further without doing their homework. How mastering the homework exercises let's you solve more problems.

I know that is not easy, the teacher may not quite understand how topics relate, why each of them is needed in a specific order, if they have not thought about that much.

But LLM has similar capability, it can look at its previous outputs, and think what further thoughts the it should generate next based on those. It can keep some of its outputs to itself, thinking in its own head,and can examine it previous private thoughts easily.

Not sure how much current LLMs do that but clearly they can at least in theory use their own outputs as their inputs too.

Current LLMs however are not conscious of pleasure and pain which really is the root of goal-oriented behavior. But maybe something like that could be programmed into them.

How would you cause an LLM pain? Or pleasure?

Solution might be legislation that puts limits on how much money each person can spend on elections. But it may be too late, there are so many rich people in the congress that such laws can not pass.

The rich not only want to get richer they also want the lower classes to get poorer so they will work for less and will have to work longer hours so they will have less time and money to educate themselves, and thus will remain clueless about what is going on.

It Catches: Mismatched function arguments, missing object properties, and typos in variable names.

It Misses: Invalid JSON from an API, unexpected database outputs, and bad user input.

That is a very good thing to help us reason about the program, we have invariants we know must hold true if the program does not stop in a type-error.