ASTRA USES A NEW REASONING TECHNIQUE CALLED "RECURRENT DEPTH". TL;DR: > instead of stacking more Transformer layers, it loops the model's hidden state through the same layer multiple times. some of the reasoning happens internally rather than writing it out as chain-of-thought tokens. why this may help: > more internal reasoning compute without generating a huge CoT > it doesn't keep adding tokens to the KV cache > as @eliebakouch pointed out in his post, recurrent depth isn't necessarily faster. each loop still costs compute and must happen sequentially. the real question is if spending that compute to loop is actually better than investing it on a deeper model. > and there’s a safety concern. CoT reasonings are easily readable. they helped retrace July's Hugging Face attack. so if more reasoning moves into latent spaces, that visibility goes away. > an AI-safety paper co-authored by researchers from OpenAI, Anthropic, Google DeepMind, etc, warned about exactly this last year: “Such latent reasoning models might not need to verbalize any of their thoughts and would thus lose the safety advantages that CoT confers.” > this is exciting because internal reasoning is more similar to how humans think, as we don't speak out everything we think about. > I think it's a matter of time before we see this adopted by more open models, just like how OpenAI’s o1-style reasoning quickly spread across the industry.

Prime Agent, our new general CLI agent, is out. We show a huge narrative violation where models have the potential to outperform their native harnesses across a variety of evals Huge shoutout to @a1zhang @omouamoua @kevinjosethomas for working on this together

🚨BREAKING: Non-coders are shipping AI apps right now HowOneAI just silently dropped the 10-minute app builder of 2026. I just shipped a real app in 10 minutes with it: Upload image → 3D miniature transform → cinematic animated video - with no coding Here’s the breakdown: 🧵

Google’s new AI marketing tool Pomelli is on fire lately 🔥 …but here’s the twist: you can build the same kind of platform on @HowOneAI in just 10 minutes ⚡️💡 Try it yourself 👉 apn3vfy-6ff205.howone.app/ #Google #HowOneAI #NoCode #AIagents #MarketingTech #Pomelli

HowOne Beta is officially live. A new way to build real Agentic Apps — no code, no flowcharts, no glue tools. But let me ask you something: Still spending hours learning “vibe coding”? Still stitching together pieces generated by different AI tools just to make one product? Still wiring model keys and configuring APIs for every single workflow? You don’t need to do that anymore. With one prompt, HowOne generates: frontends backends databases multiple agentic workflows…and even self-evolving agents — all deployable in minutes. We’re opening our first wave of beta access. If you want to become an Agentic App Builder, join early and create tools that think, act, and improve — this is your moment. Beta users get: • Free credits • Workflow evolution features • Visual workflow editing • Support to find real use cases & customers • Priority in our upcoming builder marketplace 👇 Join the beta (invite link in comments) The next generation of apps won’t be coded — they will be evolved. #AgenticApps #AIApps #IndieHackers #BuildInPublic #AIStartup

HowOne AI Alpha Test Begins NOW! You don’t need a team to build your AI startup anymore. With HowOne AI, you can turn your idea into a fully functional AI app—in minutes. Build → Launch → Earn—all in one platform. 🎥 Here’s how it works 👇 #Solopreneur #AIapps #HowOneAI #NoCode #AgenticAI

New EvoAgentX Talk coming this Sunday! 🎙️ Session Overview: Large language models are evolving fast—but how do we move beyond “text generators” to create true autonomous agents? Join our upcoming EvoAgentX Talk and follow Guibin Zhang, PhD candidate at the National University of Singapore (NUS), to dive into: ✨ How Agentic RL Transforms LLMs into Automated Agents He’ll explore: 1. Why traditional supervised learning can’t unlock the full potential of LLMs on complex tasks 2. How Reinforcement Learning (RL) enables long-term goal optimization and dynamic interaction 3. Key techniques and challenges of Agentic RL 4. How multi-agent systems, memory mechanisms, and self-evolving paradigms build truly autonomous LLM-based agents 📅 Date & Time: Sep 28 2025, Sunday, 9:30 AM (London Time) 📍 Location: Zoom Online 🔗 Join Here: Zoom short-url.org/1fZsg or Meeting ID: 862 8273 5663, Passcode: 396325 👤 Speaker Bio: Guibin Zhang 🎓 PhD Candidate, School of Computing, NUS (Advisor: Prof. Shuicheng Yan) 🔬 Research: Multi-Agent Systems, Agent Memory, Self-Evolving Agents 📚 Publications: 10+ papers in NeurIPS, ICML, ICLR, KDD, TKDE 🌟 Representative Work: Graph for Agent series G-Designer (ICML 2025 Oral) G-Safeguard (ACL 2025) G-Memory (NeurIPS 2025 Spotlight) 🤝 About EvoAgentX Talk EvoAgentX Talk is a recurring online series from the EvoAgentX open-source community, bringing together global researchers and developers to discuss cutting-edge topics such as Agentic AI, RAG, workflow automation, agent evaluation & evolution, and Agentic Memory. Community members can: Stay updated on the latest research and technology Connect with leading researchers and developers Gain inspiration and contribute to the open-source ecosystem 🌟 Join Us & Get Inspired Click the Zoom link above to participate, and follow EvoAgentX to catch future talks. Let’s witness together how Agentic RL can drive LLMs to become self-directed intelligent agents. 💡 Interested in sharing your own research or project? Reach out to the EvoAgentX team to be our next speaker! By joining our Discord channel: discord.gg/XWBZUJFwKe or following us on Linkedin: evoagentx Github: github.com/EvoAgentX/EvoAgen… #EvoAgentX #AgenticRL #LLM #AutomatedAgents #ReinforcementLearning #MultiAgent #OpenSource #AIResearch #AICommunity #NUS