RECFind
Re:Cap
Mon 09 Nov 2026The Spaces you missed, readable.862 Spaces · 51 hosts

RecapFLSoc Presents; "Fame Ladies on Top"

Fame Ladies on Top Ep. 127 — The Agent Era, Cloudflare's Classifier Trick, and Blast Shutting the Lights Off

the room, live · 84 reactions

  • 59❤️
  • 8😲
  • 7👋
  • 4💯
  • 3✊
  • 2👍

TL;DR

the one-minute version
  • Asia (@byasiaminor)* opened with a wide AI roundup: OpenAI's new always-on "dot" agent, Figma's generally-available design agent, Canva AI 2.0, and Gemini landing on the Windows desktop — her read is that we've left the chatbot era for the agent era 3:43.
  • Flick walked through Cloudflare's new Clef experiment — an LLM surgically converted into a fast, deterministic classifier for "is this web page malicious?" — and argued it's a template a lot of people will copy 27:02.
  • New model news: Mistral Large 4 ("LeChunk") entered public preview at one trillion parameters with weights promised by month's end, plus a Gemini cyber-defense model under Google's fairwind program 35:02.
  • Blast is winding down and asking users to get funds off-chain by the end of the month; Flick used it as a case study in what it actually costs a foundation to keep an L2 alive 48:02.
  • Jilly turned the last stretch into a practical wallet-security Q&A: offline phone vs. hardware wallet, Ledger vs. Trezor vs. Lattice, and how to build a long-term NFT vault 54:50.
  • Flick teased unreleased Fame Lady Society work landing as large AI-written PRs in the GitHub notifications channel — no reveal yet, just "we've got some stuff coming" 39:21.

Highlights

At 3:43

"Officially moving out of the chatbot era into the agent era."

Asia framed the whole night: OpenAI's dot as an always-on agent that works across apps instead of waiting for prompts. Flick's plain-English translation — it's essentially a private OpenClaw that OpenAI runs on their machines for you, pro accounts only, and OpenAI did hire OpenClaw's creator, so of course this is what they'd build 5:05.

At 6:23

Figma's agent is the design-tool moment Asia's been waiting for.

Searching your files, referencing other designs, maintaining design systems, working directly on the canvas — and crucially, you can hand it a markdown file of design rules and it actually follows them. Her point: this is where agents stop replacing the design process and start respecting it, after 25 iterations just to get a header out of an image model 7:50.

At 11:29

Flick doesn't open his video editor anymore.

He's been running agentic workflows over OpenShot — which, as he notes, is really a fancy interface for FFmpeg — so now he just says "this section, crossfade, add this audio" and it works. The same pattern everyone's trying to clone from coding agents, pointed at everything else.

At 17:40

A slice of the web is going dark to AI.

Flick on robots.txt bans, weird CAPTCHA walls for agents, and the flip side at his day job: agent-driven traffic is substantial and growing fast, so sites are now formatting themselves to be legible to crawlers that give up easily. A whole new layer of SEO being invented in real time.

At 19:30

Asia's one-line history of AI.

"2023 it writes. 2024 it sees and hears. 2025 it creates, 2026 it acts." Which led to her asking Flick the real question — does he trust his agent unsupervised? His answer: auto mode, sandboxed by default, with a second AI whose only job is checking his guardrails, because approving every single sandbox exit got too slow 21:11.

At 22:51

The tiny keyboard discourse.

Asia described OpenAI's first hardware — a six-key light-up dashboard for your coding agent. Flick, naturally, already owns one; it was on sale for roughly 16 hours. Asia's update that the new model has eight keys landed as genuine news to him. The whole bit resolved into Flick recalling that early-80s Macs shipped without number pads, so you bought a little brick to hang off the side 24:42.

At 27:02

The Cloudflare Clef deep dive.

Flick's best explainer of the night: keep an LLM's input side and tokenizer, chop off the text output, and coax it into behaving like a classifier — so instead of burning inference on research and prose, it looks at a page's text, code and screenshot in one pass and answers "e-commerce scam, porn, crypto scam" instantly and deterministically. He spent work hours on the blog post today and is clearly delighted. Closing note, with feeling: this still takes a team of data scientists, so nobody should believe the people saying data scientists are obsolete 33:38.

At 40:09

AI writing letters to future AIs.

Jilly admitted she sees Flick's PRs flooding the GitHub channel and has no idea what any of it means. Flick's reassurance was not especially reassuring — it's all AI-written, he barely understands it either, and it's mostly AI leaving notes for the AIs that come next. It is, he insists, exciting.

At 48:02

Blast is closing up shop.

From Blur farming into the Blast airdrop, the early DeGen-friendly L2 with its own paired mobile wallet, to Fantasy Top — the social trading game where people who accepted a card suddenly acquired fans demanding they get back on the Twitter mine and pump those engagement numbers. Now there isn't even a working UI to swap most tokens out, and users have until month's end. Flick's autopsy: a foundation running a chain pays for Etherscan, wallet listings, free RPC, bug bounties, and is also expected to be a venture fund. That money runs out fast. Jilly's immediate follow-up — "so not far from Polygon then?" — got a qualified yes, with the caveat that Polygon is basically Polymarket at this point, and Polymarket is talking about its own L2 52:54.

At 55:21

Hardware wallet or an old phone?

Jilly's question produced the most useful practical segment of the night. Flick: a working phone kept off the internet with a hot wallet on it is effectively a hardware wallet, iOS more so than Android, and any of it beats a browser on a computer. Nothing is infallible — he cited a hardware manufacturer that didn't use its random hardware device to generate seeds, with nine-figure losses — and for a forever-vault he'd reach for his Lattice or the developer command line tools, which have good enough entropy.

Topic timeline

TimeTopic
3:18 – 5:05Intros; OpenAI's "dot" always-on agent
6:23 – 10:22Figma agent, Canva AI 2.0, Adobe, AI as a creative OS
11:29 – 16:05Agentic video editing; Gemini on desktop; voice as the primary interface
16:05 – 19:30AI advertising, AI as search, and sites going dark to crawlers
20:21 – 22:38Do you trust your agent unsupervised? Auto mode and guardrails
22:51 – 26:54OpenAI's mini keyboard, rumored speaker device, bipedal house robots
27:02 – 33:38Cloudflare Clef: turning an LLM into a classifier
35:02 – 39:21Mistral Large 4, Gemini cyber-defense model, AI safety release norms
39:21 – 43:07Fame Lady Society AI work, DeFi vault automation, AI facsimiles of ourselves
43:07 – 47:40Tangent: reported Russian lab incident and gain-of-function research
48:02 – 54:46Blast winding down, Fantasy Top, Polygon, cheap Ethereum gas
54:50 – 1:01:40Wallet security: phones, Lattice, Ledger, Trezor, building a vault
1:01:40 – 1:02:23Wrap

Notable quotes

AI in 2023 it writes. In 2024 it sees and hears. In 2025 it creates, 2026 it acts.

Asia 19:30
0:38

It's all AI written. And so like I barely understand it, you know. It's just AI writing to future AIs so that when they look at what happened, they know.

Flick 40:13
0:38

So are the AI agents… getting our internet monies for us? Are they getting it for us?

Jilly 41:34
0:38

Don't believe anyone that's like oh, AI is gonna take over, not gonna even need data scientists anymore. Nah, they got plenty to do.

Flick 33:38
0:38

Thank God there's still Ethereum and, wow, suddenly it's as cheap as Polygon ever could be.

Flick 54:05
0:38

If you're like me and you occasionally make transactions on your cold wallet like a dummy.

Flick 57:32
0:38

Who said what

  • Asia (@byasiaminor)* — ran the show and the AI rundown, bringing a working designer's lens: Figma, Canva, Adobe, and the question of what advertising becomes when people ask an AI what to buy instead of searching.
  • Flick the dev (@0xFlick) — resident explainer, from Cloudflare's classifier trick and Mistral's parameter count to Blast's shutdown economics and wallet entropy. Also the only person in the room who already owned the little six-key OpenAI keyboard.
  • Jilly Snipes (@JillyRappaport) — asked the questions everyone else was thinking, steered the back half into Polygon's prospects and practical NFT vault security, and wondered aloud whether we'll all eventually be replaced by agents that talk and act for us.

Worth a full listen

  • 27:02 – 33:38 Flick's Cloudflare Clef walkthrough. It's the one stretch where the "how does this actually work" detail matters: classifiers vs. generation, determinism vs. autoregressive sampling, and why running an LLM agent to answer a yes/no question is a thousand times too expensive.
  • 48:02 – 52:54 The full Blast obituary, including the Fantasy Top era when being a card in the game meant strangers who owned you demanded more posting. A funny and slightly melancholy tour of an L2's entire lifecycle.
  • 54:50 – 1:01:06 The wallet security Q&A is genuinely useful and hard to compress: offline phones as de facto cold storage, iOS vs. Android, Lattice vs. Ledger vs. Trezor, and the honest admission that closed-source firmware means you're just trusting the manufacturer.

* some voices are identified from context; those names are marked as likely.

For listeners

Recaps by email

Pick up to three hosts or subjects, or take every room. One email, on Sunday.

Get recaps
The robot at its desk in daylight, writing by hand, the Sunday Re:Cap open on the monitor

More from FameLadySociety 💋 WE 💋 (@FameLadySociety)

14 Spaces from this room are written up here. See every one

More of FLSoc Presents; "Fame Ladies on Top"

14 episodes of this room are written up, going back to Jun 30, 2026. Every episode

These recaps are free and stay free. If this one was worth your while, send some fuel back.