super-alignment co-lead with Ilya (who resigned yesterday)
what is super alignment? [2]
> We need scientific and technical breakthroughs to steer and control AI systems much smarter than us. Our goal is to solve the core technical challenges of superintelligence alignment by 2027.
[1] https://jan.leike.name/ [2] https://openai.com/superalignment/
I'm guessing, but OpenAI probably wants to start monetizing, and doesn't feel like they are going to hit a superintelligence, not really. That may have been the goal originally.
Super intelligent AI seems contrary to the goals of consumerist Capitalism, but maybe I'm just not smart enough to see the play there.
There's zero chance LLMs lead to AGI or superintelligence, so if that's all OpenAI is going to focus on for the next ~5 years, a group related to superintelligence alignment is unnecessary.
You mean like PALM-E? https://palm-e.github.io/
Embodiment is the easy part.
My intuition leads me to believe that these are arising properties/characteristics of complex and large prediction engines. A sufficiently good prediction/optimization engine can act in an agentic way, while never had that explicit goal.
I recently read this very interesting piece that dives into this: https://www.lesswrong.com/posts/kpPnReyBC54KESiSn/optimality...
I'm tuning my probabilities back to 99%, I still don't believe just feeding more data to the LLM will do it. But I'll give the chance a possibility.
Next steps would be in totally different fields, like implementing actual reasoning, global outline planning and the capacity to evolve after training is done.
>A language model is a mathematical construct
That is like telling someone from the Middle Ages that a gun is merely an assemblage of metal parts not too different from the horseshoes and cast-iron nails produced by your village blacksmith and consequently it is safe to give a child a loaded gun.
ADDED. Actually a better response (because it does not rely on an analogy) is to point out that none of the people who are upset over the possibility that most of the benefits of AI might accrue to a few tech titans and billionaires would be in the least bit re-assured by being told that an AI model is just a mathematical construct.
Do you have a better analogy? I'd like to hear more about how ML models can't be intelligent, if you don't mind.
I'm pretty skeptical of the idea that we know enough at this point to make that claim definitively.
Books (and writing) are a big force in cultural evolution.
Books aren't data/info-processing machines, by themselves. LLMs are.
To some of us, that sounds like, "Fire all the climate scientists because they are needless cost center distracting us from the noble goal of burning as much fossil fuel as possible."
This is a tortured analogy, but what I'm getting at is, if OpenAI is no longer pursuing AGI/superintelligence, it doesn't need an expensive superintelligence alignment team.
Also, even if they never produce a superintelligence, they are likely to produce insights that would make it easier for other teams to produce a superintelligence. (Since employees are free to leave OpenAI and join some other team, there is no practical way to prevent the flow of insights out of OpenAI.)
Certainly, the researchers want the model to be as useful as possible, so there we have what I would call a 'directional bet', but since usefulness is correlated with capability to potentially do harm (i.e., dangerousness) that bet is probably not what you are referring to.
An interactive mathematical model is not going to run away on its own without some very deliberate steps to take it in that direction.
Maybe the guys who point out tar in tobacco is dangerous and nicotine is addictive maybe we shouldn’t add more for profit and such things would be useful just in case we get there.
But even if we don’t - an increasingly capable multimodal AI has a lot of utility for good and bad. Are we creating power tools with no safety? Or safety written by a bunch of engineers whose life experience extends to their PhD program at an exclusive school studying advanced mathematics? When their limited world collides with complex moral and ethical domains they don’t always have enough context to know why things are the way they are and our forefathers aren’t idiots. They often blunder into a mistake out of hubris.
Put it another way the chance they succeed is non zero. The possibility they succeed and they create a powerful tool that’s incredibly dangerous is non zero too. Maybe we should try to hedge that risk ?
Basically- the LLM won't run away on its own.
I do agree with a safety focus and guardrails. I dont agree with chicken little sky is falling claims.
What leads you to believe that's true?
Can somebody translate this to human?