hyperbolic chambers
it's safe to say the lines between sci-fi and reality are blurring. As we move closer and closer to the intelligence explosion, different horizons open up. With the birth of new fields like AI economics, LLM Biology or AI diplomacy, life really does feel like a movie.
Dwarkesh's 3-hour podcast with Scott Alexander and Daniel Kokotajilo dives into a timeline of AI, which, while I do have my reservations, provides a realistic template of what to expect.
While I recommend a complete watch, the podcast covers an interesting scale-up of AI, summarized as:
Solving coding (superhuman-level) -> solving AI research (superhuman-level) -> superintelligent models.
While there are discussions of hives of AI working together, I'd like to take a step back and entertain you with a brainfart.
The Hyperbolic Time Chamber is a dedicated training chamber in Dragon Ball Z which allows any entrant to enjoy borrowed time, i.e., 1 day within the time chamber would be = a year gone experienced by the world outside). Now, while this devolved into essentially a gimmick to allow Goku to get stronger, I think Hyperbolic Time Chambers are now something that is within reach.
As LLMs continue to keep getting better, it does not seem far off when the scaling laws plateau and progress, hitting a hard stop. The solution? LLM Anti-Hyperbolic Time Chambers. The idea is straightforward. Build effective sandboxes that work on accelerated but hard-bounded time to allow LLMs to effectively explore all RL pathways.
Think of it as an escape room. You are provided limited time and a problem to solve. With the fears of a time-limit and a pre-determined goal to achieve, you are effectively forced to co-operate with your peers to achieve said goal. Well, let's try to 100x this with time chambers. Provide LLMs with a pre-determined goal, a hard time-limit and an accelerated perceived time axis.
Note that I said perceived. To trick the LLM into assuming that it is working on a time-axis parallel but accelerated to ours. An hour for them is a minute for us, a month but a mere week. The repercussions are for the time-being, I'd say two -
- modular implementation and development of natural AI agents, self-taught and filled with individual domain knowledge.
- playgrounds to test popular theories of evolution and natural selection.
But how can I continue? I'm just a non-technical dummy.