December 1st, 2025
December 1st, 2025
What are some truly super-human capabilities?
What are some truly super-human capabilities?
There is a lot of talk right now about models getting closer or somehow overtaking human abilities...
There is a lot of talk right now about models getting closer or somehow overtaking human abilities...
At the same time, focusing on mimicking (or exceeding) human skills on various tasks misses something that is conceivably more important going forward: there are some truly super-human abilities built in models that can already be leveraged now that no one seems to speak about
At the same time, focusing on mimicking (or exceeding) human skills on various tasks misses something that is conceivably more important going forward: there are some truly super-human abilities built in models that can already be leveraged now that no one seems to speak about
What do you think I am speaking about?
What do you think I am speaking about?
I would tend to guess that the super-human strength that most would ascribe to computers is their raw compute power and memory
I would tend to guess that the super-human strength that most would ascribe to computers is their raw compute power and memory
This is not really what I have in mind in the context of language models or AI agents... they have a truly supernatural power that no human or animal conceivably ever got close to having:
This is not really what I have in mind in the context of language models or AI agents... they have a truly supernatural power that no human or animal conceivably ever got close to having:
Language models can (in principle at least, though this is not really done in practice) spawn other agent copies of themselves (or of other LLMs), and hence make 'experiments' with these copies, providing them with different sets of information and goals, and seeing how those reacts
Language models can (in principle at least, though this is not really done in practice) spawn other agent copies of themselves (or of other LLMs), and hence make 'experiments' with these copies, providing them with different sets of information and goals, and seeing how those reacts
The Superhuman Power of Counterfactual Thinking
The Superhuman Power of Counterfactual Thinking
I personally often find asking myself the question to evaluate situations: "what if I didn't know this, how would I look at the world?", "what if I worked for a day on this idea, would I be better off?"... and this is obviously impossible for living being: we can't make "side effect-free" experiments on copies of ourselves (let alone others)
I personally often find asking myself the question to evaluate situations: "what if I didn't know this, how would I look at the world?", "what if I worked for a day on this idea, would I be better off?"... and this is obviously impossible for living being: we can't make "side effect-free" experiments on copies of ourselves (let alone others)
In some sense, I would surmise that this is what makes part of the beauty of existence: the "yolo" idea that life just happens once, and that most choices are irrevocable, but it is obviously not an asset in terms of decision-making
In some sense, I would surmise that this is what makes part of the beauty of existence: the "yolo" idea that life just happens once, and that most choices are irrevocable, but it is obviously not an asset in terms of decision-making
However, with software agents, this can be done (and is in fact not very difficult technically)
However, with software agents, this can be done (and is in fact not very difficult technically)
This was one of the ideas at the root of the xent project, before we even started thinking about the whole space of games: that is an area where we can leverage an LLM beyond the current generative paradigm (I find myself using
This was one of the ideas at the root of the xent project, before we even started thinking about the whole space of games: that is an area where we can leverage an LLM beyond the current generative paradigm (I find myself using
.