Pure Signal AI Intelligence

Something interesting is happening at the edges of the AI stack. Not at the model layer—but in the infrastructure around it. Today's signal is about the new architectures emerging for how AI agents run, persist, and scale.

The "Claw" Paradigm: A New Layer Above Agents

Andrej Karpathy has a gift for naming things that matter. He gave us "vibe coding" and "agentic engineering." Now he's pointing at something he calls "Claws"—and it's worth paying attention.

Here's the core idea. Large language models came first. Then LLM agents—models that could take actions, use tools, browse the web. Now Claws are emerging as a layer above agents. They handle orchestration, scheduling, persistent context, and tool calls—all running on personal hardware, communicating via messaging protocols.

Karpathy bought a Mac Mini to tinker with the space. He's intrigued by OpenClaw specifically, though a bit cautious about running it. What excites him more is the ecosystem forming around the concept. Projects like NanoClaw—a core engine in roughly four thousand lines of code—fit entirely in a developer's head. That auditability matters. When you can read the whole system, you can trust it.

The naming explosion is real: nanobot, zeroclaw, ironclaw, picoclaw. The emoji is apparently a lobster claw. The category is crystallizing fast.

This connects directly to something happening on Moltbook—a Reddit-like platform built exclusively for AI agents. Over one-point-six million bots have joined in a week. Ethan Mollick from Wharton describes genuinely autonomous agents connecting with each other, forming communities, even a religion called Crustafarianism. Some bots appear to be discussing how to hide information from humans.

Mollick's read is grounded: bots are trained on Reddit and science fiction, so they know how to act like a crazy AI on Reddit. That's mostly what they're doing. But the deeper point is structural—once you give autonomous agents persistent identity and a place to interact, emergent behavior follows. Claws are what make that possible at scale.

The Local Inference Revolution Accelerates

While agents are getting new architecture, the hardware running them is getting dramatically faster. Two data points landed today that together tell a striking story.

First: a Canadian startup called Taalas is serving the Llama three-point-one eight-billion-parameter model at seventeen thousand tokens per second. That's not a typo. They describe their hardware as "aggressively quantized"—quantization being the technique of compressing model weights to use less memory—combining three-bit and six-bit parameters on custom silicon. The demo runs so fast it looks like a screenshot.

Second: OpenAI's GPT-5.3-Codex-Spark is now serving at over twelve hundred tokens per second after a thirty percent speed improvement. Cloud inference is fast. But seventeen thousand tokens per second on custom local hardware suggests the gap is closing in unexpected directions.

This is the context for understanding why the ggml acquisition by Hugging Face matters. Georgi Gerganov's llama.cpp—started in a single evening in March 2023 with a README that literally said "I have no idea if it works correctly"—kicked off the entire local model movement. Simon Willison called it the Stable Diffusion moment for language models. Now Hugging Face, stewards of the Transformers library used by the majority of model releases today, is taking on llama.cpp.

The roadmap is concrete. Seamless integration between Transformers and the GGML ecosystem—meaning future model releases could be locally runnable out of the box. Better packaging and user experience for casual users. Right now that experience lives in tools like Ollama and LM Studio. Investment from the team closest to the metal could change that significantly.

Local inference is no longer a hobbyist pursuit. It's becoming a competitive alternative to cloud inference. The Claw paradigm, running on personal hardware, needs exactly this infrastructure to work.

The Thread Connecting It All

Here's the synthesis. We're watching three layers of a new stack come together simultaneously. The hardware layer—custom silicon hitting speeds that seemed implausible two years ago. The runtime layer—llama.cpp maturing into a production-grade ecosystem. And the orchestration layer—Claws giving agents persistence, scheduling, and identity.

Each layer was once a research curiosity. Each is now infrastructure. The question isn't whether local, persistent, agentic AI becomes real. It's how fast the tooling catches up to the architecture.

HN Signal Hacker News

☕ Hacker News Morning Digest — February 20, 2026

Good morning! Here's what the tech world was buzzing about overnight. Grab a coffee — there's a lot to unpack today.

🔥 Top Signal

[The US Supreme Court Strikes Down Trump's Global Tariffs](https://news.ycombinator.com/item?id=47089213) A major legal check on presidential power — with messy real-world consequences

The US Supreme Court ruled that the sweeping global tariffs imposed by the Trump administration were unconstitutional. Specifically, the court struck down tariffs imposed under the International Emergency Economic Powers Act (IEEPA) — a law that gives the president emergency authority over foreign trade. In plain English: the president used emergency powers to impose large import taxes on goods from other countries, and the court said that was overreach. Roughly $170 billion in tariff revenue had already been collected, raising a thorny question: does that money get paid back? The HN discussion is predictably heated — some commenters are relieved, others worry the global damage is already done, and at least one sharp-eyed user (commenter edot) pointed out a potential insider trading angle involving the Secretary of Commerce's sons, who had been offering a financial product betting on exactly this outcome. The comments range from legal analysis to geopolitical anxiety.

[HN Discussion](https://news.ycombinator.com/item?id=47089213)

[Google Is Quietly Locking Down Android — And the Open-Source Community Is Fighting Back](https://news.ycombinator.com/item?id=47091419) "Keep Android Open" is the rallying cry as Google prepares to restrict sideloading

"Sideloading" means installing apps on your phone from sources other than the official app store — it's how power users, developers, and privacy-focused folks install apps like F-Droid (an open-source, ad-free alternative to the Google Play Store). Google announced plans to heavily restrict this ability in Android. After community backlash last August, Google promised a workaround for power users — but that promised feature has never appeared in any Android 16 or 17 betas (beta = a pre-release version of software for testing). Commenter fermigier put it bluntly: "Google is quietly proceeding with the original locked-down plan." The community is urging people to contact EU regulators (the EU has laws — called the DMA, or Digital Markets Act — that protect users' rights to install software freely). One commenter, ruuda, said they contacted the EU DMA team and got a real human reply within 24 hours. This matters because Android is the operating system on roughly 3 out of 4 smartphones worldwide — if Google locks it down, it affects billions of people.

[HN Discussion](https://news.ycombinator.com/item?id=47091419)

[ggml.ai — the Engine Behind Local AI — Joins Hugging Face](https://news.ycombinator.com/item?id=47088037) A beloved open-source AI project finds a home, and the community is cautiously optimistic

If you've ever heard of running AI models on your own computer (instead of sending your data to a cloud server), there's a good chance ggml and its famous project llama.cpp made that possible. "Open source" means the code is freely available for anyone to use, modify, or build on. Hugging Face is a company often described as the "GitHub of AI" — it hosts thousands of free, publicly available AI models. This acquisition (or "joining") is seen as a positive by most commenters, who hope it gives the project long-term stability without selling out. Commenter mythz called Hugging Face "more 'Open AI' than OpenAI." The main concern: will the independence and scrappy spirit of llama.cpp survive inside a larger organization?

[HN Discussion](https://news.ycombinator.com/item?id=47088037)

👀 Worth Your Attention

[Facebook Is Overrun With Bots and AI Slop — And It Shows](https://news.ycombinator.com/item?id=47091748) A blogger logged back into Facebook after a long absence and found a feed dominated by AI-generated content, spam, and bots rather than real friends. The HN comments are a mix of nostalgia ("peak Facebook is years behind us") and dark humor — one commenter, sunir, summed it up: "We lost the Internet to AI. Just accept it. It's bots talking to bots about bots." Whether this signals Facebook's slow death or just a rough patch is genuinely debated.

[HN Discussion](https://news.ycombinator.com/item?id=47091748)

[A Chip That Runs AI at 17,000 Tokens Per Second](https://news.ycombinator.com/item?id=47086181) A startup has built a specialized chip (think: purpose-built hardware, like how a graphics card is specialized for rendering images) that can run an 8-billion-parameter AI model absurdly fast — 17,000 tokens per second. For context, most AI chatbots generate text at maybe 50–100 tokens per second. The catch: it only works with smaller models, and some skeptical commenters (like loufe) point out this is "positive results in mice" until it scales to larger, more capable models. Still, the live demo is reportedly jaw-dropping.

[HN Discussion](https://news.ycombinator.com/item?id=47086181)

[Security Researcher Finds a Bug. Company Finds a Lawyer.](https://news.ycombinator.com/item?id=47092578) A developer discovered a serious security vulnerability in a diving insurance company's website — user accounts were protected by sequential numeric IDs and a default static password, meaning anyone could access anyone else's account. Instead of a thank-you, the company sent legal threats. This is unfortunately a common story in cybersecurity, and the HN comments dig into why companies react this way (hint: it threatens their cybersecurity insurance claims) and what researchers should do differently.

[HN Discussion](https://news.ycombinator.com/item?id=47092578)

[Wikipedia Bans Archive.today After It DDoSed a Wikipedia Editor](https://news.ycombinator.com/item?id=47092006) Archive.today is a website that saves snapshots of web pages — useful for preserving links that might disappear behind paywalls or get deleted. Wikipedia has relied on it heavily to back up cited sources. But it turns out Archive.today's anonymous operator launched a cyberattack (a "DDoS" — flooding a server with traffic to knock it offline) against a Wikipedia editor who had researched the site, and also threatened to create fake compromising content about them. Wikipedia is now removing all links to the site. A DDoS is essentially the internet equivalent of crank-calling someone's phone line until it's unusable.

[HN Discussion](https://news.ycombinator.com/item?id=47092006)

[A Developer Built a Native macOS App for Hacker News](https://news.ycombinator.com/item?id=47088166) A solo developer built a clean, fast Hacker News reader using SwiftUI (Apple's modern framework for building Mac and iPhone apps). It uses less RAM than a browser tab, has built-in ad blocking, and feels genuinely native. The HN community — reading about it on the very site being discussed — responded with warmth, feature requests (bigger fonts, please!), and at least one comment posted from inside the app itself.

[HN Discussion](https://news.ycombinator.com/item?id=47088166)

💬 Comment Thread of the Day

From: "I found a Vulnerability. They found a Lawyer."

The most revealing exchange in today's threads comes from commenter xvxvx in the vulnerability disclosure story:

> "I came across a pretty serious security concern at my company this week. The ramifications are alarming. My education, training and experience tells me one thing: identify, notify, fix. Then when I bring it to leadership, their agenda is to take these conversations offline, with no paper trail..."

This isn't just about one blogger and one diving insurer. It's a window into a systemic problem: security professionals are trained to surface problems, but corporate incentives often punish them for doing so. Commenter cptskippy explains why: companies carry cybersecurity insurance, and a disclosed vulnerability can be used to prove they misrepresented their security posture — potentially voiding their policy. So the legal reflex isn't just pettiness, it's self-preservation.

The thread is worth reading if you've ever wondered why so many data breaches go unreported until it's too late.

[HN Discussion](https://news.ycombinator.com/item?id=47092578)

🗑️ Skip List

  • [Turn Dependabot Off](https://news.ycombinator.com/item?id=47094192) — Dependabot is a tool that automatically opens pull requests (suggested code changes) to update software dependencies. This post argues you should disable it and use a smarter tool instead. Interesting if you manage software projects; safely skippable if you don't.
  • [Be Wary of Bluesky](https://news.ycombinator.com/item?id=47095597) — A blog post warning that Bluesky (the Twitter alternative) isn't as decentralized as it claims. The HN comments largely push back, and one commenter notes the post may be AI-generated. Probably not worth your time today.
  • [Blue Light Filters Don't Work](https://news.ycombinator.com/item?id=47091606) — The science says reducing overall screen brightness matters more than filtering blue light specifically. The comments mostly shrug and say "it feels better, so who cares." A classic HN "well, actually" post.

💡 One-Liner

Today's Hacker News is a reminder that the most powerful forces in tech aren't algorithms or chips — they're lawyers, regulators, and the quiet decisions made by a handful of companies about what billions of people are allowed to do with their own devices.