>pick an ambitious project it wanted to work on
The LLM does not have wants. It does not have preferences, and as such cannot "pick". Expecting it to have wants and preferences is "holding it wrong".
The LLM does not have wants. It does not have preferences, and as such cannot "pick". Expecting it to have wants and preferences is "holding it wrong".
They have to do this manually for every single particular bias that the models generate that is noticed by the public.
I'm sure there are many such biases that aren't important to train out of responses, but exist in latent space.
What do you think humans have?
LLMs need a retrain for that.
Whenever you message an LLM it could respond in practically unlimited ways, yet it responds in one specific way. That itself is a preference honed through the training process.
The architectural limits will always be there, regardless of training.
CEO's are gonna CEO, it seems their job has morphed into creative writing to maximize funding.
IMO we’re clearly there, gpt5 would easily be considered agi years ago. I don’t think most people really get how non-general things were that are now handled by the new systems.
Now agi seems to be closer to what others call asi. I think k the goalposts will keep moving.
The GPT model alone does not offer autonomy. It only acts in response to explicit input. That's not to say that you couldn't built autonomy on top of GPT, though. In fact, that appears to be exactly what Pulse is trying to accomplish.
But Microsoft and OpenAI's contractual agreements state that the autonomy must also be economically useful to the tune of hundreds of billions of dollars in autonomously-created economic activity, so OpenAI will not call it as such until that time.
Every human every day has the choice to not go to work, has the choice not to follow the law, has a choice to... These AI doesn't have nearly as much autonomy as that.
> The concept does not, in principle, require the system to be an autonomous agent; a static model—such as a highly capable large language model—or an embodied robot could both satisfy the definition so long as human‑level breadth and proficiency are achieved
Edit -
> That is ultimately what sets AGI apart from AI.
No! The key thing was that it was general intelligence rather than things like “bird classifier” or “chess bot”.
It says that one guy who came up with his own AGI classification system says it might not be required. And despite it being his own system, he still was only able to land on "might not", meaning that he doesn't even understand his own system. He can safely be ignored. Outliers are always implied, of course.
> No! The key thing was that it was general intelligence rather than things like “bird classifier” or “chess bot”.
I suppose if you don't consider the wide range of human intelligence as the marker of general intelligence then a "bird classifier" plus a "chess bot" gives you general intelligence. We had nearly a millennia ago!
But usually general intelligence expects human-like intelligence, which would necessitate autonomy — the most notable feature of human intelligence. Humans would not be able to exist without the intelligence to perform autonomously.
But, regardless, you make a good point: A "language classifier" can be no more AGI than a "bird classifier". These are narrow systems, focused on a single task. A "bird classifier" doesn't become a general intelligence when it crosses some threshold of being able to classify n number of birds just as a "language classifier" wouldn't become a general intelligence when it is able to classify n number of language features, no matter how large n becomes.
Conceivably these classifiers could be used as part of a larger system to achieve general intelligence, but on their own, impossible.
Obviously you can get probability distributions and in an economics sense of revealed preference say that because the model says that the next token it picks is .70 most likely...
If a model has a statistical tendency to recommend python scripts over bash, is that a PREFERENCE? Argue it’s not alive and doesn’t have feelings all you want. But putting that aside, it prefers python. Saying the word preference is meaningless is just pedantic and annoying.
Perhaps instead of "preference", "propensity" would be a more broadly applicable term?
Try explaining ionic bonds to a high schooler without anthropomorphising atoms and their desires for electrons. And then ask yourself why you’re doing that? It’s easier to say and understand with the analogy.
Outside that? If left to their own devices, the same LLM checkpoints will end up in very same-y places, unsurprisingly. They have some fairly consistent preferences - for example, in conversation topics they tend to gravitate towards.