Ask HN: Can we classify AI as: subhuman, quasihuman, human, superhumam?
Gen 2: quasihuman: human assistant that requires human-in-the-loop: human evaluates, gives directions, asks, tries, implements. "you're here": LLMs/chatbots (2020s).
Gen 3: human worker substitute. While not strictly 'human' per se, without will and emotions it can be considered "human job replacement" as it eliminates the need for human-in-the-loop as it verifies successful compilation, rates visual result (videogames it produces/videos/images), validates fiction it writes, marketing strategies/trading strategies and learns from mistakes and improves after many attempts until it finishes task with acceptable result.
Gen 4: Superhuman: like gen 4 but + having full recursive self-improvement (it updates/rewrites its own code and recopmiles while performing a task - full awareness)? Theoretically gen 4 can be ASI if the seed is "make better AI which makes better AI and so on until optimal AI is achieved given the laws of physics" (good learning and ability to update everything should achieve it - though compute limitations may prevent it and well...regulations).
I know the definition is suboptimal and people are extremely opinionated on AI to a point people have different definitions about AGI and ASI but this is at least some definition! I believe after "dry" chatbots in the 30s we may switch to less user-friendly yet smarter and more useful AIs, you will not longer see "you're absolutely right!" and have to spent hours to evaluate and correct errors with the AI, you will just start it and wait for it to finish, kind of like downloading a movie in the old days. Too sci-fi/wrong? Too pessimistic...because Yudkowski/Terminator is coming to get us in 2029? 10x for your comments!
7 comments
[ 0.21 ms ] story [ 20.8 ms ] threadCommercially viable technology is "superhuman" in some sense. That is, tractors and combine harvesters changed everything in agriculture because... they are superhuman. Bulldozers are superhuman, pocket calculators are superhuman, etc. If something isn't superhuman in some narrow sense, humans will just do the job!
Note the gap you talk about between 2 and 3 is a real gap. In the 2010s I was working on (basically) reverse centaur systems that were intended to get radical productivity improvements on specific tasks. I didn't see "Gen 2" systems talking to ordinary people were even possible, at least not until we built a workable world knowledge and language facility via that reverse centaur path: expert input developing the outline of the world model and setting the standards for performance and quite a bit of mass input tagging and evaluating and making judgements that the system couldn't yet been trusted.
As for 4 I don't believe in recursive self-improvement, not like you can't use today's AI to accelerate the progress of tomorrow's AI but it is an evolutionary but not revolutionary change and the limits are fundamental to physics, economics and the nature of intelligence itself. Basically you have heard of this differential equation
as representing exponential growth as seen in ecology and economics, but you never hear of the equation that describes recursive self-improvement not even from true believers in recursive self-improvement because (1) there are no examples of it because it is unphysical, uneconomic and non-sensical, and (2) any serious discussion of why (1) is will lead to numerous refutations of recursive self-improvement, "the singularity" and such.To be specific right now, take a look at the semiconductor industry which has been stuck for almost 10 years. For a long time people talked about Moore's Law and the consequences of being able to make smaller transistors exponentially over time. Looking back though, the real issue is that we've gotten cheaper transistors over time. If they don't get cheaper it doesn't matter if they are smaller or faster because you can't afford them. It used to be a video game console was always the most expensive on Day One but it's not true anymore. It is fashionable to blame this on the pandemic or the AI boom, but the real problem is that those EUV machines from ASML cost too fucking much and are just barely economically viable. The industry is stuck, does not have a clear path forward. Maybe AI can increase our capacity 10x but if the problem is 10x harder you don't get "singularity" growth you get a continuation of the old exponential growth.
Ntheless.. id appreciate your original ideas for ways to forgive those investors hoping for a "phase transition" as an upside (similar to crypto returns). what would be the downsides to acknowledging their feelings of missing out?
Its fun to think of Anthropic/OpenAI as pure entertainment companies and be thrilled when their drawdown looks similar to Netflix's..
I was thinking a more accurate scale would be subintelligent, quasiintelligent, intelligent and superintelligent.
These seem like very logical divisions if you are going to call the whole collection "AI" to begin with.
Even more realistic would probably be based on what capabilities are there, rather than things like "humanity" content or pseudo-IQ.
How about subadequate, quasiadequate, adequate, and superadequate?
Any way you look at it those "super" levels are going to be very challenging to validate or confirm.