REI Network
@ReiNetwork0x
Followers
14K
Following
1K
Media
108
Statuses
435
Advancing AI through fundamental scientific principles • Research led by @0xreisearch on @Base and @HyperliquidX
Joined November 2024
Walking Users Through The New Paradigm - Training at Inference Time :
mirror.xyz
Training at inference time means your AI unit evolves through interaction, not through separate training phases. Every conversation, every query, every piece of feedback becomes a learning opportun...
25
38
186
RT @0xreisearch: Core succeeding where a world model failed. When you try to build “consciousness loops” where agents update their own inf….
0
24
0
A big thanks to @0xDaes, @RickCrosschain and @alphaggxyz for being among the teams testing and using 0.3.3 to get past their SOTA LLM hurdles. Training at inference time's first iteration allows for training at an extremely low cost in comparison to traditional methods and.
We're still training our REI Unit and building out the flow but soon it will be able to be hooked to Telegram bots, or other external api for easy access. Excited to ship this soon as part of the @alphaggxyz CT intel suite. Below you can see a snapshot of the n8n workflow.
11
25
126
Weekend maintenance complete. A system-wide speed recalibration was successful as we transition from this summer’s phase to broader access and ecosystem expansion. Performance improvements across select usecases our team tested:. • Direct queries: 77.8% latency reduction. •
18
29
156
2025.08.22 Biweekly update. Introduced R00Ms: An interactive multi-user / multi-Unit environment. Released Core 0.3.3 zero-decay memory beta. Deployed computational workflows feature beta. INTERNAL
Computational Workflows. What's New. Units can now process entirely new categories of data, unlocking computational workflows that were previously impossible. From analyzing datasets to processing professional documents, Units now handle structured information and execute complex
17
34
153
6/ Each R00M becomes a persistent workspace where your team of users and Units can collaborate over time. Active testers will receive invitation emails for early access soon.
0
0
36
5/ Invite human collaborators to join. Multiple users can work with multiple units simultaneously. Everyone sees all interactions in real-time.
1
0
27
4/ Build your team by adding specialized units. Whether it is CodeGen, Chart Generator, Marketer, Strategist, Planner, Summarizer, or Writer, each unit will contributes its expertise while staying aware of others' contributions.
1
0
17
3/ Create rooms with specific purposes and tags. Define what you're building and units maintain focus on that objective throughout the collaboration.
1
1
23
2/ Key features:. • Multiple users and AI units in one conversation.• Units are aware of each other's actions and responses.• Persistent collaboration.• Real-time multi-party interaction.• Units only respond when necessary
1
2
26
Introducing R00Ms: An Interactive Multi-User / Multi-Unit Environment. R00Ms will soon start rolling out to select testers via email invites.
33
59
213
RT @0xreitern: grei, welcome to the REI Unit Training Ground. Training doesn't have to be done alone. Every prompt is live training data ᕦ(….
0
21
0
9/ Core 0.3 is the first iteration and has great room for improvement, Versions 0.4 and 0.5 will take this further. We're just getting started on what inference-time training can actually do.
1
1
36
8/ Each problem becomes an opportunity to strengthen reasoning capabilities. Not to memorize another pattern, but to validate and refine the actual thinking process. Core evolves through confronting novelty, not through repetition. The question isn't "did we improve by X % on.
1
2
35
7/ The feedback is immediate and honest. You can't game it with better search or retrieval. Either your reasoning solves the new problem or it doesn't. No amount of pattern matching or tool use substitutes for actual problem-solving ability. Consider what this means: a model.
1
1
27
6/ This is learning through doing, not through memorizing. In our case, Core discovers which reasoning principles actually work by applying them to new problems and observing the results. Success reinforces valid reasoning, failure eliminates bad strategies. Think of it like.
1
2
26
5/ The key difference: traditional systems apply memorized patterns or search for answers. Inference-trained AI develops reasoning approaches through actual problem-solving. Each successful solution strengthens the reasoning strategies that worked. Every interaction becomes a.
1
2
27
4/ Inference-time training offers a different path. Not LoRA or fine-tuning, actual reasoning-time computation where systems develops domain-specific understanding as it works through each problem. Instead of frozen weights doing forward passes or clever retrieval strategies, AI.
1
0
25
3/ This happens because current training optimizes for statistical correlations in training data. Systems learn "when I see pattern X, output Y" without understanding why. Systems now use web search, retrieval, and other tools to solve certain problems. The gain does not lie in.
1
1
25
2/ The fundamental problem: a model improves X % on task Y, but if those gains don't transfer to other tasks or create new reasoning paths, what have we actually achieved? Tests are great and should exist and so does the wall current models are hitting. The metric that actually.
1
3
24