A hypothetical GAI remains hypothetical. We’ve been working on self-managing systems for 25+ years and still have very little idea how to make them. IF we can solve hallucinations (which have recently been shown to be equivalent to patients with various stages of dementia) and research can give us new self-* technologies capable of configuring, healing, protecting, and optimizing both logical (software) and physical infrastructure (because what does a GAI do with a cable cut?), then we might see a quick devaluation of human labor. Also, everyone is speaking as if GAI is imminent. It’s not. LLMs are impressive as a nascent technology (primarily because they interact with us using natural language which is one of the largest initial necessary but wholly insufficient preconditions to strong AI) but we are very far off from something that can replace all or even most specialties of human labor.
chiamarc
Karma: 1
- chiamarc 17 Jan 2025 21:36 UTC2 points−1in reply to: Eivind Kjorstad’s comment on: AGI Will Not Make Labor Worthless
I would think that solving problems with swarms vs single agent has a lot in common with parallelization of algorithms (of all classes). Every large problem we want to solve should be categorizable into a complexity class. Then it’s just a question of how small we can make the constants. Running any swarm will still be subject to Amdahl’s law and synchronous vs asynchronous execution. In any cooperating swarm (that is not solving “embarrassingly parallel” or mapreduce workflows), you’re always going to have increased latency and compute costs due to message passing. If the solution requires deep logic loops or complex global constraints, you can get a “telephone game” effect where crucial nuances are lost between handoffs.
All that being said, just as we discovered (and sometimes recategorized) improved parallel algorithms for traditional and distributed computing, e.g., the polylogarithmic-time ones in Nick’s Class, we will eventually hone in on a similar rubric for cooperating (or even adversarial) agents.
One more caveat: language models are intentionally non-deterministic at the software layer and “accidentally” so at the hardware layer. This requires a shift into probabilistic and statistical frameworks, e.g., randomized classes, to truly get a handle on space and time bounds. Interestingly, swarms can turn non-determinism into an asset via error dampening. Whereas a hallucination in the long Markov chain of a single agent can multiply exponentially to result in nonsense, some subset of a swarm can act as critic agents, checking others’ work and arriving at a consensus (at a cost of course).