2,052 karma · joined March 12, 2023
Then it was models like Minerva that could actually solve math problems, and the discovery that LLMs were one-shot learners and could write code.
After that, the improvement felt pretty steady, with IMO gold feeling like a watershed moment.
And recently OpenAI's solution to the planar unit distance problem is starting to actually freak me out a bit.
> The 2026 multicenter placebo-controlled trial extending this work enrolled 240 participants with early Alzheimer’s... The intervention group showed slower decline on standard cognitive scales by about 30% versus placebo.
But there's no such study in the references section. Not sure what's going on there but I want to see the data before I believe this.
> Consider two scenarios, first a firm that used to buy a $10,000 HR service from an outside provider now buys that HR service for $10,000 from an AI HR provider. In that case the output still is captured in national accounts and all that disappeared was the wages and workers. In the second version that $10,000 service is now done internally for $10 of tokens. In that scenario GDP has declined by $9,990 despite the same work being done.
Also sending this article to family members so they're aware of this kind of thing.
"Treated models" - it sounds like they're trying really hard to hide the fact that this was all in mice. From the paper:
> Therefore, using a mouse model, this study investigated whether IN administration of hiPSC-NSC-EVs in late middle age can significantly reduce oxidative stress and curb microglia-mediated neuroinflammaging in the hippocampus.
Cool! But please be honest in your press releases.
This is a good analogy for AI work displacement. Probably would resonate with some of the college students who boo'ed Eric Schmidt.
I'm also afraid of a world where AI completely replaces human mathematicians, but if we remain collaborators, then that's a world I can still feel excited about.
I also wonder if it is to see whether the remaining employees pick up the extra workload using AI...
(I hate ads too, but I think I understand the alternative perspective).
That's a really compelling argument against the Turing Test. But in order to build such a machine, you would need an enormous amount of compute to populate the answers. The interesting question is then whether consciousness emerged while doing all that pre-compute.
I'll even take it a step further; most of an LLM's training is next-token prediction on random internet content. A newly-trained LLM will just continue whatever text appears in its context window, like an extremely capable autocomplete. The illusion of an entity that takes turns in conversation and presents a consistent personality is tacked on at the last minute through RLHF. This was the transition from GPT to ChatGPT.
Any positive evidence of LLM consciousness should probably mostly be taken from the model before post-training, where it displays remarkable capabilities but shows no sign of a consistent personality, and likely no signs of self-awareness or self-understanding.
We may also be overestimating the richness and complexity of an LLM relative to a human when we entertain these possibilities, but who knows.
Some people think that consciousness is related to quantum mechanics, but the laws of quantum mechanics can be simulated with a Turing machine so that doesn't necessarily change the story.
To flesh this out a bit more, I agree that ability to communicate is not enough (ELIZA probably didn't pass the bar, even if it did kinda pass a Turing test). But that's also not what gives me pause with LLMs. It's how much information processing they seem to be doing under the hood.
It's really hard to imagine how next-word prediction could lead to consciousness, but I find it almost as hard to see why evolution did. If we can't even detect whether something has subjective experiences, then how can it have been selected for evolutionarily? The only possibility I see is that consciousness is a byproduct of some kinds of information processing tasks.* And if it's something that emerges naturally, then the line starts to get very blurry.
*This sounds reductive, but I don't at all mean it that way.
I've had a lot of thoughts and conversations over the years that changed my mind on what consciousness likely requires. One was the realization that a purely mechanical computer can, in principle simulate the laws of physics, and with it a human brain. So with a few other mild assumptions, you might conclude that a bunch of gears and pullies can be conscious, which feels profoundly counterintuitive.
I think that was the moment I stopped being sure about anything related to this question.
(Let's suppose the agent did need an API token to e.g. read data).