I am the Head of Curriculum Development at Iliad, responsible for the content of the Iliad Intensive. Previously, I did a PhD at the University of Amsterdam working on AI Safety and Alignment, and specifically safety risks of Reinforcement Learning from Human Feedback (RLHF). I also worked on abstract multivariate information theory and equivariant deep learning. https://langleon.github.io/
Leon Lang
Karma: 2,361
Hi,
yes, we received an application from you at 22nd of April, for the August Iliad Intensive. We have not yet begun processing those applications and will do so reasonably shortly after the deadline on 22nd of June!
I think karma is composed of more things than post quality. There is also “how much interest is in the topic at the time of posting”, and “how smart do you need to be to understand the post” and “how controversial is the raised opinion”, and “how much do voters themselves benefit from the post appearing polular”. Probably more.
Anywhere on Earth and End of Day. It’s basically the last time at which the clock changes to the next day anywhere on earth.
Hi! We don’t do interviews.
This post and the application form is all information we provide on what we’re looking for. It’s not very gameable: Simply make sure we can easily assess your mathematial expertise, potential research experience, etc.
There’s a third type of context: Reading up on internal development docs of your company, having access to your chain of thought and all your scaffolding, and generally access to any secret information that is not by default known to other instances of you, including in the form of system prompts and stated user preferences. I think this type of context is relevant and might push everything back into the multi-agent regime that you critique in your post.
You could have the hypothesis that there would be many different AIs for different contexts and skillsets in the past 5 years. But in fact this is not the case. Frontier models are SOTA at (nearly) everything.
I think we’re narrowing down on a crux! I disagree with this statement.
You do in fact need to specialize your AI for the work you want it to perform. That’s why you give it a context in the first place! Without a context it won’t do anything useful.
Frontier models are SOTA at nearly everything, yes, but crucially not simultaneously: For each new task where you evaluate them, you need to provide them a suitable context, usually disjoint from the contexts of all the other tasks.
I think I didn’t swap them! The thing is that
I think I do qualitatively agree with everything you say here, but I still seem to end up with different overall intuitions, but hashing that out would require making our predictions a bit more concrete probably.
I think what I am saying is that arguing for job specific context to matter a lot is a big bet against the bitter lesson. I think that’s unlikely and an epistemic mistake.
I do think if “jaggedness” gets solved, then we will end up with one transformer that can learn any individual job quickly, so that’s the part I do agree with. I do not think that we’ll end up with one big transformer that has the ability/knowledge to do all the jobs at once.
One crucial question is probably whether the amount of new important contexts will grow at least as fast as the capacity of a transformer to learn all those contexts. My guess is yes, this will indefinitely be the case.
It’s very unclear to me whether merging and sharing context is actually easier between claude instances! You can copy-paste context-windows, but eventually the receiving window is full.
You might also be able to share context via weight-updates in continual learning, but continual learning is iirc not solved, and at global scale I also wonder whether you eventually run into catastrophic forgetting issues.
Perhaps things are still qualitatively different from humans but that needs a more careful argument.
I would have written a comment very similar to chasmani’s to your answer to the following snippet you reacted to, if chasmani hadn’t already done so:
Expecting a model to do all the work, solve everything, come up with new innovations etc is probably not right. This was kinda the implicit assumption behind *some* interpretations of capabilities progress. The ‘single genius model’ overlooks the fact that inference costs and context windows are finite.
People overrate individual intelligence: most innovations are the product of social organisations (cooperation) and market dynamics (competition), not a single genius savant.
I think you are not appropriately answering the point made here.
Yes. This perspective is also behind some intuitions for gradual disempowerment: Even if you have an aligned AI, if you specialize it into a billion contexts, each individual AI may try to do good while the collective still destroys the world.
I raised this 3.5 years ago as well.
This is great, thanks for writing it!
As you learn, you may also be interested in the section on singular learning theory in our Iliad Intensive course.
Thanks, we’ll include it in the next version!
My guess is around two months from now! So end of June.
Likely within a week. We will not do interviews.
I think we will not run interviews in this round, and my guess is you should receive an answer within one week.
In the April cohort you would probably have gotten in with this knowledge, but I can’t tell you how our bar will change over time. I definitely recommend you to apply.
We will at the start of May publish the April version of the Intensive course materials, including prerequisites (globally for the program, and more specifically for each module). So look out to when we publish this to know what knowledge is useful to have upfront.
List the things you’ve learned and courses you have attended / are attending in your CV! If you read a technical book, list it too.
In the April edition we have a cohort of 16 people among 40 applicants. If we receive sufficiently many promising applications and our office space allows, we may have cohorts of up to 30 people, or maybe eventually even more down the road.
Part of the reason of writing this post is to encourage more relevant applications, so competitiveness in future rounds may be higher than in the April round, but that’s hard to predict.
If you have uncertainties, please still apply.
It seems like we received a second application from you on June 3rd as well.