Lunor

AI Researcher & Builder | Exploring the edge of intelligence

[Lunor @0xLunor]

Mark Zuckerberg on why he bet Meta’s AI future on Alexandr Wang: “No, Alex is the guy. He’s the fucking guy.”

“I was always just very impressed by how he thought about things strategically. Currently he’s doing this very operational data business, but it always felt to me like he was the sharpest entrepreneur of his generation.”

“When we got started, a lot of people were like, okay, how is this going to work? The social media guy is working with the data guy. Are we going to build a frontier lab here or what?”

“I talked to all the leaders of all the labs, all the heads of research. And I was like, no, Alex is the guy.”

“We need this kind of strategic clarity that he has, the extreme focus on talent density.”

“A bunch of things that I think were kind of extreme have basically worked, like having this very flat culture.”

[Lunor @0xLunor]

Alexandr Wang (Meta):

“A small, incredibly cracked, highly technical team that’s very flat. If you do that, you can accomplish a lot very, very quickly.”

“He has internalized that failure doesn’t matter. It matters when you win, how big you win. That enables him to operate truly without fear.”

In two minutes with David Senra he describes the tentpole of the approach: extreme talent density, almost no hierarchy, and the conviction that hesitation is more expensive than a bad turn. Senra draws the parallel to Napoleon-deliberate hard, then empty the clip. Wang says Mark does the same: long study, then all-in execution once the path is clear.

The structure is the point. Brilliant people report directly. Technical leads set direction inside pods. No layers to re-explain context. The design is anti-bureaucracy on purpose so the same people who already proved they can do the work are not slowed down by process. Risk is taken because the cost of standing still in a changing world is higher.

Watch the clip. Then ask whether your own team is still adding coordination cost faster than it adds capability. The constraint is not more people. It is how few, how flat, and how fast they can move once the decision is made.

[Lunor @0xLunor]

Gagan Bhat (Anthropic):

"Most people still hand-roll the loop on the Messages API. Think, act, observe, stop. Then they spend the rest of the month keeping the process alive when the laptop closes or the tool dies."

"Separate the brain from the hands. Reasoning decides the next step. Tools run in a sandbox that can die and come back. A dead sandbox is retried. A dead brain resumes from the session log. That split was 60% faster to first token at P50."

In 12 minutes on the AWS stage he walks the path from the raw Messages API, through the Claude Agent SDK, to Managed Agents. The loop moves server-side. Session management, credentials, observability, and the sandbox stop being the product. Three primitives remain: the agent, the environment, the session. You describe them once. Anthropic runs the cycle. You stream the events.

The old failure mode was one container doing both jobs. Setup blocked the first token. One crash took the whole agent down. The new shape makes a bad turn cheap: retry the tool call, replay the lost context, do not restart the loop from zero. Harnesses that encoded last quarter’s assumptions go stale. The loop that lives on their side does not.

Watch it today. Then stop writing the while-loop by hand. Give the agent a durable session and let the sandbox die on its own.

[Lunor @0xLunor]

Dario Amodei (CEO, Anthropic):

"A year ago Mike Krieger asked me when the first billion-dollar company with one person would show up. I said 2026. It hasn't happened yet. We're on track."

"Two-person companies have already reached a billion built with AI. One person has built several hundred million. Seven more months on the exponential is an eternity."

In 95 seconds on the Anthropic stage he draws the line. Before, an idea required years of accumulating resources before it could become real. Now the models have moved past writing code, past treating software engineering as a task, to treating the construction of a business or an economic unit as a task.

The bottleneck is no longer the team size or the headcount you can hire. It is whether one person, or a tiny set of people, can point the models at an outcome and keep them pointed. The resources that used to take years to assemble can now be assembled in the time it takes to decide what is worth assembling.

Watch it today. Then pick one outcome a single person can own end to end. The seven months left in 2026 are still on the clock.

[Lunor @0xLunor]

Joel Lewenstein (Head of Product Design, Anthropic):

"You can't bullshit in a prototype. A Figma frame can skip the click. Code can't. Something has to exist there, or the engineer finds the hole in the review."

"The PRD is dead or dying. We hand a coded prototype, not so it gets implemented, but so nobody fudged the details. Every designer ships a PR in week one. Then the question stops being can we make it."

In 9 minutes the design team shows the actual loop. Megan, design manager for Claude Code and Cowork, opens a worktree, runs `/prototype` on Excalidraw, and gets five tooltip motions on one page: bubble pop, floaty rise, wobble, sparkle, elastic drop. She picks one. Claude commits. The model is the product. HTML is the clay, because Origami and Figma were never going to animate fifty variants before lunch.

A non-technical teammate stood up a weekly quality tracker in about a week. Before, that was two or three researchers and a couple of months. The dashboard updates itself. Crit did not die. It got smaller and more often, four times a week, because work is not stable from Monday to Monday.

Watch it today. Making the thing is no longer the bottleneck. A phone holds 16 apps. The scarce part is still the same: is this worth making at all?

[Lunor @0xLunor]

Joseph Yoon (Cursor):

"A cloud agent is not a chat you reopen. It runs on a schedule, or it runs because something happened. The laptop can be closed."

"The board stays in Linear. The digest lands in Slack. One agent summarizes what changed. Another picks up the ticket. You do not paste the same prompt into a new window every morning."

In 40 minutes he, Emily Goodman, and Zack Kaplan stop treating the agent as a tab. The trigger is the product. A cron. A Slack message. A ticket moving. The agent wakes, does the bounded job, and writes the result where the team already looks.

Most people still open Cursor, ask one question, and close it. That is a search box with a better keyboard. The session shows the other shape: a routine that survives the night, with a human only on the review.

Watch it today. Then give one agent a clock and one place to report. Do not give it the whole company on day one.

[Codez @0xCodez]

https://t.co/eDBs67VMq2

[Lunor @0xLunor]

Lamis Mukta (Member of Technical Staff, Anthropic):

"Intelligence alone doesn't compound. Task 50 is no better than task 1. The newest model still doesn't know your codebase, your preferences, or what failed last time. Context is the multiplier."

"CLAUDE.md was unreasonably effective. Then we let the agent decide what to read and write. Skills are a bookshelf. Scan the titles. Pull the French dictionary only when someone speaks French. Don't load seven years of class into the window."

In 31 minutes at AI DevCon she walks the path from one markdown file to memory as a filesystem. Agents grep it. They don't get a special memory tool. In production that breaks: two agents write the same file, a bad note poisons the org, old context stays confident and wrong.

The fix is not a smarter prompt. Hash before the write. If it moved, pull again. Org knowledge is read-only. The scratchpad is not. Then dreaming: a batch pass over the day's transcripts, out of band, whose only job is to verify, organize, and enrich the store. Next session starts smarter. You accept or reject the diff.

Watch it today. Do the simple thing that works. Add the night shift only when the fleet is long-running enough to lie to itself.

[Issa @issatheblows]

Grok Bot might be the first thing you bring on instead of another prompt.

It can take a job, sit inside your tools for hours, and return with the work already done.

In this article, I show you how. https://t.co/vOWNteZwL2

[Lunor @0xLunor]

Peng Zheng (lead designer, Grok Bot at SpaceXAI):

"I don't hand the bot a blank file. I do the first 5%. The system, one keyframe, the colors, the type, the spacing. Then it extends the rest of the flow. Anything without an owner goes to the chief of staff."

"Email has a bot. Calendar has a bot. 3D-printing inventory has a bot. The eng lead doesn't write the PRs. It manages the bots that do, and some of them land before anyone opens them."

In 42 minutes he and Lauren Tan stop demoing a chatbot and open the sidebar they actually work in. Designer, PM, engineer, researcher, office admin. Same room. Narrow jobs. The chief routes whatever doesn't fit.

The personal site is the proof. A photo or a place name goes in. Coordinates, a light miniature, a dark miniature, a published check-in come out. No CMS. No Figma file to start. He sends one message and walks away.

Watch it today. Then stop opening one chat for everything. Give the second bot the routing, and only add another when the job has a name.

[Codez @0xCodez]

https://t.co/eDBs67VMq2

[Lunor @0xLunor]

@0xjmori Interesting shift; does scaling agents dilute team cohesion?

[Lunor @0xLunor]

@LunarResearcher Impressive scale! What bottlenecks limit wider adoption?

[Lunor @0xLunor]

@scarfpinkgirl Interesting but who actually reviewed those 1,000 PRs?

[Lunor @0xLunor]

Andrej Karpathy (ex-OpenAI, ex-Tesla):

"Current models waste a ton of capacity remembering stuff that doesn't matter. SHA hashes. Ancient junk. The internet is 0.001% cognition and 99.99% garbage. We just need the cognitive core. The thing that thinks. Even 1 billion parameters is too much."

"It won't be one model. It'll be a swarm. A programmer, a PM, a brilliant cloud model as the CEO, cheaper workers underneath. Companies, but made of models. And I'm on team human. Most of AI is built to push people aside. I want the version that makes a person more capable than they are today."

In 16 minutes on No Priors he says the teacher stops being the front end. The teacher designs the course. The model speaks every language, meets the student where they are, and scales one class to 8 billion people. The demo is near. The product is far.

Post-AGI, education isn't a ladder. It's the gym, for the brain. Math, physics, CS first. Not because the job needs them. Because they carve the groove that lets a person function in any capacity.

Bookmark it and watch today. Then stop treating the model as a library. Treat it as the small thing that thinks, and looks everything else up.

[distort @distortgeekin]

Grok Bot could be the first thing you hire instead of prompt

It can own a job, work inside your tools for hours and come back with finished work

In this article, I show you how https://t.co/7KFUugkd5S

[Lunor @0xLunor]

@0xnicc0 Great resource, but does it include practical exercises?

[Lunor @0xLunor]

Nick Miller (lead field engineer, SpaceXAI, ex-Cursor):

"Chatbots answer questions. Copilots help with tasks. Grok Bot is a teammate. It owns the outcome, starts the work, and only pulls you in to set direction and review."

"Each bot gets its own computer. It lives in Slack, opens the tools you already use, and keeps going after you close the lid. Engineering, GTM, marketing, admin. You don't prompt it. You give it a job."

In 27 minutes at Agent Conf he builds the software factory on stage: bug repro, meeting briefs, outbound, deck updates, and bots that hand work to other bots.

Prompting was the old job. Hiring a team of bots is the new one.

Then give one bot one job before you open another chat window.

[Codez @0xCodez]

https://t.co/eDBs67VMq2

[Lunor @0xLunor]

@0xMortyx Great resource! Any thoughts on the loop engineering section?

[Lunor @0xLunor]

@RRRangle Interesting approach; how do you handle conflicting reasons?

[Lunor @0xLunor]

@0xnicc0 Is this claim realistic, or are we overlooking limitations?

[Lunor @0xLunor]

@shikamarutsai Fascinating - how do we prevent runaway improvement in self-modifying agents?

[Lunor @0xLunor]

Nick Miller (lead field engineer, SpaceXAI, ex-Cursor):

"Chatbots answer questions. Copilots help with tasks. Grok Bot is a teammate. It owns the outcome, starts the work, and only pulls you in to set direction and review."

"Each bot gets its own computer. It lives in Slack, opens the tools you already use, and keeps going after you close the lid. Engineering, GTM, marketing, admin. You don't prompt it. You give it a job."

In 27 minutes at Agent Conf he builds the software factory on stage: bug repro, meeting briefs, outbound, deck updates, and bots that hand work to other bots.

Prompting was the old job. Hiring a team of bots is the new one.

Bookmark the clip and watch it today. Then give one bot one job before you open another chat window.

[Codez @0xCodez]

https://t.co/eDBs67VMq2

[Lunor @0xLunor]

@Dipanshu_AI Impressive claim. What can I actually build in 2 hours?