Off the radar - Part 1
I went to a belgian AI conference - part 1
On tuesday 23rd of June I went to a conference called "Off the radar", about Belgium and AI.
It started with a fun tid-bits: Ineffable Intelligence was co-founded by a Belgian.
Ineffable Intelligence is also one of the most talked about AI frontier lab talked about during the conference by some participants.
--
AI mini-class and vulgarization:
LLMs as we know them are given shitload amount of text, and they are train by having to guess the next words in a given text.
So they learned in a fixed kinda way. They have their sentences, and they know what word they should guess at training time.
It's a supervised type of learning.
An other type of model is called reinforcement learning (RL), and it's how the ai gets good at chess or go.
They guess the next steps, but the next steps are not judged, they keep on playing until an end goal, and if the goal is reached, then they know it was a good type of steps. While the LLM gets feedback straight at the next word, RL do not know how good they are at steps, but only at the end of the game.
--
Ineffable Intelligence wants to create a kind of LLM but train as a go player. The limits though, that I find, are that it's only possible if they can be trained in a fake world - a world model is the term used in science - which is really difficult to build. So they have 2 challenges - and talk about the second one when the first one is actually important.
That was one of the thing of the conference, we couldn't ask the questions in public. There was a robotic lab that specialised in training robots on modelled world but that were very very limited. So the robots were learning limited actions in a limited world -- they could never stop what they were doing to save someone.
Talking about labs that did robotics -- one of the most impressive ones -- was VerticalCompute. The idea is that they could build a new type of chip component that, if true, would power AI at a better level than the data center level that we have now.
It started with a fun tid-bits: Ineffable Intelligence was co-founded by a Belgian.
Ineffable Intelligence is also one of the most talked about AI frontier lab talked about during the conference by some participants.
--
AI mini-class and vulgarization:
LLMs as we know them are given shitload amount of text, and they are train by having to guess the next words in a given text.
So they learned in a fixed kinda way. They have their sentences, and they know what word they should guess at training time.
It's a supervised type of learning.
An other type of model is called reinforcement learning (RL), and it's how the ai gets good at chess or go.
They guess the next steps, but the next steps are not judged, they keep on playing until an end goal, and if the goal is reached, then they know it was a good type of steps. While the LLM gets feedback straight at the next word, RL do not know how good they are at steps, but only at the end of the game.
--
Ineffable Intelligence wants to create a kind of LLM but train as a go player. The limits though, that I find, are that it's only possible if they can be trained in a fake world - a world model is the term used in science - which is really difficult to build. So they have 2 challenges - and talk about the second one when the first one is actually important.
That was one of the thing of the conference, we couldn't ask the questions in public. There was a robotic lab that specialised in training robots on modelled world but that were very very limited. So the robots were learning limited actions in a limited world -- they could never stop what they were doing to save someone.
Talking about labs that did robotics -- one of the most impressive ones -- was VerticalCompute. The idea is that they could build a new type of chip component that, if true, would power AI at a better level than the data center level that we have now.