We will shut down the dango🍡 testnet-2 shortly, in a few hours. Over the 2 week period, 212,205 usernames were signed up; more than 3.5 million orders were fulfilled; 135k accounts completed our quest on Galxe and claimed the limited time OAT. You guys hammered our servers

Introducing 𝚡𝟺𝟶𝟸-𝚖𝚌𝚙. Bring x402 payments into Agents & MCP servers with the AI SDK. • Open protocol based on HTTP 402: Payment Required • <$0.01 fees, sub-$0.001 minimums • Account-less + anonymous • 3 LOC to implement vercel.com/blog/introduci…

Anthropic's Ralph plugin sucks, and you shouldn't use it It defeats the entire purpose of Ralph - to aggressively clear the context window on each task to keep the LLM in the smart zone. Full article here: aihero.dev/s/9tdgRM

New art project. Train and inference GPT in 243 lines of pure, dependency-free Python. This is the *full* algorithmic content of what is needed. Everything else is just for efficiency. I cannot simplify this any further. gist.github.com/karpathy/8627f…

Harnesses should offer far more customization than they do today. We’ll probably move toward a scenario where people have personal harnesses, customized to their needs, producing outputs that differ from everyone else’s work.

Time to consider not just human visitors, but to treat agents as first-class citizens. Cloudflare’s network now supports real-time content conversion to Markdown at the source using content negotiation headers. cfl.re/4ksZQ1S

DGX Spark just arrived. Hot take: “tok/s” is the wrong headline metric. Prefill/refill (prompt → KV cache) drives TTFT + long-context UX. Decode is a different bottleneck. EXO Labs measured Spark ~3.8× faster prefill vs M3 Ultra blog.exolabs.net/nvidia-dgx-spa…

NVIDIA just dropped PersonaPlex-7B 🤯 A full-duplex voice model that listens and talks at the same time. No pauses. No turn-taking. Real conversation. 100% open source. Free. Voice AI just leveled up. huggingface.co/nvidia/persona…

Single-chat speed doesn’t tell the full story. In multi-agent systems, the metric that matters is aggregate throughput under concurrency. With vLLM + parallel sessions, totals can look wildly different. Spark turns into a monster at 1,125+ tok/s.

How much of your AI stack do you actually own? Most people I know are one API deprecation away from a broken product. What's the minimum viable sovereignty worth maintaining?

I honestly believe we’re at a critical moment as a society where we need to seriously re-evaluate how we use AI and what role humans will play in this new paradigm. In just a few years, a large majority of corporate jobs could disappear. This will undoubtedly have a massive

ait - track all your AI provider usage from the terminal Claude (oauth) Session 72% remaining [████████░░░░] Weekly 41% remaining [█████░░░░░░░] Codex (oauth) Session 25% remaining [███░░░░░░░░░] Cost $8.42 today |

I kept copy/pasting “install skill” code across CLIs. It got old fast. So I started extracting it into skillinstaller: a library-first installer engine for the Agent Skills ecosystem. Important: this is not a “skill marketplace” or a repo browser. It does not go hunting for

For a long time, we pretended building software was a neat sequence: plan it, design it, build it, test it, ship it, maintain it. The steps are not wrong. The ordering just does not survive reality anymore. Modern development looks less like a linear process and more like

Astro 6 just proved that Rust eating frontend tooling is the new baseline. If you're building AI UIs and LLM wrappers with bloated frameworks, you're falling behind. Fast, runtime-agnostic frontends are mandatory when shipping at lightspeed.

Everyone is focused on Grok the chatbot, but the real alpha is the xAI developer ecosystem. With their API and massive context windows, builders can create specialized agents that rival OpenAI. Are you integrating xAI into your stack yet?

WASM sandbox layer is the key unlock here. It gives you isolation without the overhead, so agents can fail fast and gossip useful signals. curious how you'd handle convergence vs chaos in the gossip protocol at scale though.

agent-browser is now fully native Rust. The results: 1.6x faster cold start. 18x less memory. 99x smaller install. Less abstraction means faster shipping, more control, and capabilities that weren't possible before. Now with 140+ commands across navigation, interaction, state