I don’t need an LLM that is smarter than me
*No LLMs were used to write this text*
Recently, working on hard problems with LLMs became quite frustrating for me. It happened somewhere around Astra / Fable launch. Initially, I wasn’t able to put my finger exactly on the main source of this frustration, because there is a always a background noise of the unevitable slop, “gates”, “D1/D2/D3” decisions and other forms of reasoning that AI tends to generate at every research problem. But, some quite new form of frustration was added at this generation of models. After some introspection I found it — those models have became much smarter and knowledgable then I am.
Sorry if comes as offence, but surely you folks know this feeling — when you are explaining something to someone who is not that smart and you can clearly see that your explanation doesn’t land at all and the person can’t even recognise that they are absolutely not getting it.
I clearly remember that case, when I was like 8 year old and my dad was trying to explain to me what is a mathematical function. I was honestly trying to grasp but my best interpretation was “function of sun is to rise, right dad?”. A miss by a very long distance and not even close to the concept.
So, I guess this is happening again. During yet another discussion with frontier LLM around work related ML problems an LLM in one paragraph LLM dropped names of a few theorems I never heard of, connected them into a new one, drove conclusion, used statistical methods I have never seen and just assumed I should know that and able to follow. I was left with a few choices — either trust it or spend significant amount of time trying to tie all that in my head, learn new concepts and draw the connections. And that was quite over my head even in terms of mathematics I am familiar with.
So, the invariant I am arriving to is this:
Intellectual parity is required for productive and engaging work with AI
Surely clean intellectual parity setup is useless from POV of the predicted singularity. We need models that will push the frontier of human knowledge. But, as many great researchers have already noted — it seems that the future knowledge will be incomprehensible for humans.
Get Oleksandr Savsunenko’s stories in your inbox
Join Medium for free to get updates from this writer.
But, practically speaking, I see a clear need for a translation layer. I still want frontier-level models to drive research further and beyond of what I can produce even theoretically. But I need to:
- be able to understand results
- be able to steer results
- be sure that I am not missing some critical edge case and/or fundamental misalignment
It is absolutely not clear to me how to solve this. When I started to write this note, my idea was to propose a naive “translation” or “dumb-it-down” layer. Think of a normal agent-subagent relationship, but inverted. Usually we have smarter main agent delegating piecewise work to “stupider” systems. In this case I wanted to have “stupid” a.k.a. intellectually-aligned model to talk to me and explain to me, while delegating real research and real hard problems to frontier model.
But is this the right idea? From pre-AI era we have examples of how it worked in human organizations. Ideally, you want smartest and brightest in the top of the structure, but surely this is not always the case — and some of such non-meritocratic structures are still extremely successful.
The skill set of solving hardest problems and the skillset of monetizing/implementing and driving the changes are clearly different in human organizations. Think of archetypes of scientists, engineers and leaders — all are needed to make something big a reality.
And current generation of LLM seems to be more aligned with science and engineering, much less with real-world organizational skills, decision making skills, etc.
Question: do we need a special class of CEO-aligned, orchestrating models, that care about crisp executive summaries and clean decision making alignment?
