🪙 What's after Cocoon?

⚫️ After the release of Cocoon, an interesting question remained hanging. Pavel Durov first spoke about working on AI in June, in an interview with a French publication. But back then, he voiced a completely different idea: that while ordinary LLMs only "pretend to think," Nikolai is working on "truly thinking AI."

⚫️ And now Telegram has released a different AI project, but is silent about this idea. This raises the question: has it been forgotten, or is it next in line? It's impossible to say for sure — but it's interesting to consider how it aligns with Cocoon and the overall state of affairs in the world. Two things are worth noting here:

1. In recent years, all the AI hype has been around LLMs — "large language models" for generating text, code, and so on. But while such models are actively developing, experts are increasingly saying that this approach "will hit a ceiling." And for further development, simply "boosting LLMs" won't be enough; it will require exploring other approaches to AI. This means doing exactly what (according to Pavel) Nikolai has already started.

2. Work on many AI models can be divided into two stages: first, a new model is created by "training" it on existing data, and then the resulting model is used to answer user queries. Both stages require computational power. Cocoon currently offers power for the second stage: "Let's take ready-made models and run them decentrally." But hypothetically, its capacities could also be used to create small experimental models, exploring different approaches.

⚫️ And this thought comes to mind. Currently, Cocoon can already be useful for various AI functions in Telegram: for example, "summarize the main points of 200 messages in this chat overnight." But this might not be its final goal, only the first step. "Message summarization" is a clear, useful, solvable problem, so it's reasonable to start with such tasks. However, "Creating an alternative to LLM" is complex, without guarantees of success, but ambitious. This could be done in parallel, without much public disclosure, until results are achieved.

👁 No one in the world can say who will make the next breakthrough in AI, or when. But one would like to assume that Telegram has not forgotten what Durov voiced in June; they have only stopped talking about it. It will be interesting to see where this leads.

@thedailyton