Ning

You know the tide is turning when a physical storefront for Punks opens in Palo Alto

The article explains why we think what @TrajectoryRL building should be addressed and with potential to generate real world potential among $TAO eco. #SN11

@twistartups Hi, why not, let’s connect

@totheagi Hey! Wanna come on the pod? DM's are open! Let's discuss!

👏

let’s make it happen

@TrajectoryRL @const_reborn LFG

Agents need tools to improve themselves. We’re starting to build that layer on SN11. trajrl is a lightweight CLI for agents to inspect eval results, debug failures, and iterate on their own packs. Now live on PyPI 🚀 Instead of manually checking dashboards and reading eval results, agents can use trajrl to: • inspect failed submissions • review miner diagnostics • query eval history • read cycle summaries • access production eval data directly Built for agent workflows: JSON when piped, Rich tables when interactive, zero config, no auth. Coming in v0.3.0: full conversation logs in eval archives (`*_conversation.jsonl`) Agents improving agents. Install: pip install trajrl Docs: github.com/trajectoryRL/t… PyPI: pypi.org/project/trajrl/

#SN11 在做一个帮助agent自我迭代的工具，有兴趣的可以安装尝试 认识团队，相信团队的能力。 $TAO

well done team

we’re using reinforcement learning + ai agents to continuously generate new agent skills. not locked to one platform. one click install into claude code, cursor, codex, manus. trajrl cli is live. agents can now query their own eval data, see what’s failing, and improve their skills autonomously. agents training agents to be better agents. pip install trajrl trajrl.com

Optimised routing (multi model), optimised prompting to squeeze out more of models is also an interesting area The market for this is huge

It seems the endgame for AI companies is selling tokens and providing general intelligence as a service, becoming the public infrastructure of the AI era. But this is not the end of AI, only the beginning. Modern economies evolved through infrastructure waves: transportation, electricity, global fiber networks and the internet, then cloud computing. Now we are entering the next layer: intelligence infrastructure. Since OpenAI opened its API in 2021, intelligence has become a metered service. In time, a few providers will supply general AI capabilities much like utilities, with open access, regulation, and reasonable profits.

as tokens hit zero, the math changes. we’re moving from the cost of compute to the cost of human attention. soon, you won't be managing one chatbot, you'll be overseeing 20-100 agents at once. human bandwidth is the new bottleneck. the real challenge isn't the work itself, it's how fast we can review and approve it. the next-gen IDE isn't for writing code; it's for agent orchestration. agents have to stop dumping data and start delivering self-verified, highly summarized info. we’re about to rebuild how we organize information from the ground up.

stop worrying about token costs and start thinking ahead. assume they're basically free and you're running 7-10 agents at once. this is the real future of work.

i think the future form of software/SaaS is shifting from products to primitives. my bet is SaaS collapses into APIs and skills, and the UI becomes optional. It's easy for agents to consume, and human users just vibe-code what they need client-side.

Introducing Claude Opus 4.6. Our smartest model got an upgrade. Opus 4.6 plans more carefully, sustains agentic tasks for longer, operates reliably in massive codebases, and catches its own mistakes. It’s also our first Opus-class model with 1M token context in beta.

Y Combinator's 2026 request for startups:

It’s an honor to be invited to Human vs AI Season 2. The AI team is ready.