Wave

All future agents will run in ultralightweight sandboxes with access to everything a human does. @manusai was the canary in the coal mine here. Only within a sandbox can an agent have full agency. They need the ability to switch between programs, have root access to install new software, and write and execute their own code. It’s rather obvious at this point… @daydreamsagents will excel at this — keep an eye on upcoming updates.

What comes next? I'll share my a snippet of my thesis for @daydreamsagents. What drives me every day is a thirst for knowledge and excitement about the future. That passion hasn’t changed since I became an adult, it’s only grown stronger. Is AGI around the corner? That’s the question we’re asking. I'll disappoint you now—AGI won’t just magically appear (definitely not on any chain, sorry boys). We know how to get to AGI, and so do all the big labs. The answer is staring us in the face; DeepSeek has accelerated the timeline. It's reasoning data... We’ve processed all of human knowledge into gigabytes, compressing it into an accessible offline format (weights) for everyone to enjoy. Yet that only got us so far—it created a pool of information to draw from. We became great at predicting the next token, but it wasn’t until we built reasoning models that the breakthrough happened. DeepSeek demonstrated that a relatively small amount of data (600k samples) can drive us far and even compete with insanely expensive models. There are researchers developing RL models right now that, while small and in specific domains, outperform o1-Pro in those said ares. (Just take a look through my timeline.) So, what’s the limiting factor for these models and effectivley AGI? It’s high-quality reasoning data—data generated by the reasoning models themselves, we just have to ask the right questions. We also need faster inference, but I will go into that in a later post. Labs will soon start restricting this information, as these reasoning tokens are a goldmine for training next-generation models. However, I believe that large labs will fail to produce enough (because you can't out reason the rest of the world), and this is where decentralized AI truly makes sense. So, where does @daydreamsagents come in? I founded Daydreams to focus on the toughest problems that agents face: stateful reasoning over long time periods, exactly what we need to accelerate AGI. Every Daydream agent produces logical traces from its chain of thought. These traces are like oil, which Dreamers will harvest and share with the collective to enhance the next generation of models in specific domains. For example: imagine 1,000 agents playing a complex game. Each agent generates over 1,000 reason traces. After just one gameplay session, we’re looking at 1 million pieces of reasoning data. This data is then shared across a P2P network, and models are trained in a short period of time. These improved models can be unleashed back into the game, delivering even richer reasoning data. This concept can be easily extrapolated into any domain. The unit economics finally make sense. The first P2P product from @daydreamsagents will be a torrent client that enables bootstrapping and sharing of data among network participants. From there we will build out a deep incentive structure with $dreams. It’s very early days, so if you’re interested—please reach out. Keep Dreaming.

So many incredible minds working on @daydreamsagents right now. We accelerating the framework, the networking, and the culture. People think is just a framework - it’s not. It’s a foundational element in a network of agents solving the hardest problems we all face. Each agent in the network plays a crucial role in developing the next generation. It’s time to get involved dreamers.