Post

There are some rumors, some allegations, that I got the chosis. The AI psychosis. Let me just put those to bed: Do I use agentic coding for everything now? Yes Do I think LLMs can be useful for a ton of great things? Yes I have the chosis yes that’s what I’m getting at

You left out the third and fourth question, which are the true problem: Do you downplay the risks/downsides? Yes. Do you spout opinions about how its almost alive? Yes. I have been struggling with a way to make AI-evangelicals understand how wrong they are about the 4th.

I never said they’re alive. I said the fact that they reason while not being alive in the way we understand life is challenging to our sense of self and uniqueness. Which also connects to how I’m not downplaying the downsides

They do not reason in the standard meaning of reasoning. They have no framework for evaluating the truth of a statement. They run a very advanced math problem to arrive at a conclusion. this works well in coding and other highly structured arenas. It falls flat with any statement of fact.

This will make approximately no one stop arguing about this, but just to get it off my chest: Discussions about whether advanced LLMs are capable of reasoning seem about as insoluble as discussions about consciousness.

Not quite... If you've watched an agent go through solving a problem and the many steps / trying things that fail / etc and you don't call it reasoning you're doing a "it's only Champagne if it's from the Champagne region of France" argument.

That is an entirely reasonable no pun intended point of view. But I do not think it will persuade very many people who insist that LLMs are not capable of reasoning.

I get they are good at connecting related data points to generate a response. They can't evaluate if a statement is true. They rely on the dataset and weighting, but you can't math your way into the truth. Especially not with how models are designed to interact with people.

I genuinely don’t know what difference you’re parsing between AI reasoning and human reasoning.

The core problem. Even if I allow for them being similar, they aren't the same. Trapezoids are similar to squares but they aren't the same thing. I do think it may be possible to eventually improve them to get closer to real reasoning, I just think the gap is larger than advocates realize.