This shit is simply amazing at so many levels.

They literally feed their slop to each other, producing an convincing cognitive illusion of a meaningful real-time collaboration to a normie.

Yes, the idea to extend the context (adding new tokens to a context ‘so far’ with a hope of better outcome – a “better” stream of generated tokens) is by any means not a new one.

The amazing part is that it basically pollutes the context with barely relevant slop.

So, for example, one “agent” writes about some Python API (and the code) while other agents are “implementing” Rust.

Does the “agents” (the very same GTP-based LLMs) are capable of translating code from one language to another, preserving the semantics? Absolutely not, but they can create a cognitive illusion of being able to do so (while in fact, again, operating merely on the level of “abstract” tokens).

And yet, this is an actual multi-threading – a parallelised brute-force approach.

Just like we all know, the problem is in merging the slop into a coherent whole, not just at the level of source code being merged into version control system without breaking any existing code (which is a major unsolvable problem), but also at the level of different “reading” and subtle differences in “meaning”, leave alone different implementation languages.

Again, merging is a major problem, and they should literally implement all the actual XP, TDD practices to committing into a common “trunk” (non-bullshit CI) and “Shifting To The Left” philosophy and best practices at Google.

I am too lazy to perform an analysis of how to write the prompts efficiently, based on the notions of Neuro-Linguistic “Programming” (NLP) from the 80s, but there is definitely a better way to do it – by planing appropriate phrases which would, hopefully, trigger (jump to) different, “distant” regions of the model, hoping it will cause some “emergent effects” in the slop.

Key concepts and techniques from the clasic CS books are the obvious examples, assuming they trained the models on pirated CS books, which they definitely did, but, perhaps, post-processed (lobotomized) the models to cover this up.

So, the code of 4 agents in parallel was still a low-effort crap, but the show was absolutely amazing. I really said “wow”, as I supposed to do (by the marketing team).