Mikk Silk

Can the fly turn sensory evidence into a decision? In this short experiment it has to earn bananas and solve the captcha. We follow its retinal-contrast sensory readout as it inspects images.

Since we founded @rei_labs (formerly Rei Network), our idea has been to create smarter, more agentic AI that goes beyond the heavy cognitive limits of the current generation of language models. From the beginning, Core has been the system's brain handling reasoning, decision-making, and learning. Everything else, including language processing, serves as an interface to Core's intelligence 0.1 was the proof of concept for this idea with Bowtie's first prototype dating to 2024. 0.2 introduced agentic adaptation. 0.3 focused on training at inference time. With 0.4, we went a step beyond that. 0.4 not only allows the whole system to evolve from every user interaction (going beyond single-unit evolution), but it also created a clean separation from LLM embeddings. Since 0.4, Core has become an entirely separate entity that can work in conjunction with any kind of data or model. Such a separation represents our bet that intelligence is about how concepts connect and evolve, not about params. Core builds understanding by navigating relationships between ideas, discovering patterns through exploration rather than memorization. It's a fundamentally different approach, instead of pattern matching, our approach consists of developing reasoning strategies that compete and evolve based on what actually works. With Core now being substrate-agnostic, we can connect it to anything. Take coding environments as an example: Core will handle all the reasoning while using language models purely as tools: dictionaries for code generation. You can switch between Claude, GPT, or DeepSeek and Core will preserve all learned concepts and adapted strategies/teachings. Tell it once how you prefer to structure functions, and it remembers across "dictionary" swaps. In this scenario, The language model is a knowledge base, Core does the thinking. Each interaction evolves its understanding of your coding style, accumulating teachings that persist regardless of which Language model translates them. Once Core in its raw form releases, it becomes directly connectable to anything, the same principle will apply everywhere. Connect it to G-code generators for CNC machining, it learns your toolpath preferences and material handling patterns while the model just translates to machine instructions. Link it to SQL engines, it evolves query optimization strategies specific to your database patterns while models provide syntax. Interface with MIDI controllers Core develops your composition style while models handle note encoding. Even in medical imaging, Core would learn diagnostic patterns from radiologist feedback while vision models just extract features. We’re looking forward to exiting stealth, to the next versions of Core over the coming years, and to the impact multi-disciplinary AI research will have on the world. On behalf of the team, I’d like to personally thank you all for your support throughout the year. Rei is now 1 year old, and it wouldn’t be possible without you. You gave us the freedom to build without constraints and the time to get it right.