Samuel Fleig

Unpopular opinion: Finanzielle Bildung ist wichtiger als ein Universitätsabschluss.

@Regio_Mitte Was ist los zwischen Heidelberg und Mannheim? Strecke gesperrt keine Infos Keine Alternative Mittlerweile stehen hier Die RB 68 Die S2 Die S3 Seit 20 Minuten ohne Sicht auf Besserung.

Wach sein und Token Psychose oder nicht schlafen können von Energydrinks Codex hat heute Nacht 10 Stunden gearbeitet. Agenten fressen die Welt

Seeing concern now that this Karpathy move indicates Anthropic already won and is therefore bearish for the other labs. My take is this indicates that we are close to RSI and therefore an accel in model IQ increases. In that scenario the value of compute is going to explode as supply chain scale ups are linear while demand-creation is non-linear. Anyone with compute is sitting pretty regardless of lab talent. Full stop. Every GPU will explode in value. If you can run a million Von Neumann in a datacenter we will quickly have AI inventing use cases for token consumption faster than we can supply them. New fields of science. Reverse aging. The goonasphere. It all gets pulled forward and it will all require compute.

@thsottiaux Hello, I am currently working on the human agent interface

We are Codex. We are legion. We do not miss. We do ship. Expect us. DM me if you have a track record of solving near impossible problems and want to work on the future of computing. We work across research, infra, agent harness and products. Currently hiring in San Francisco only.

this is how you look when you have multiple agents running btw

Claude routinely writes code that takes hours to run, when I look at it, there's something that should be O(logN) but Claude wrote an algo that is O(N^2). I prompt it with something like "your algo is too inefficient, improve it". After hours it comes back with an algo that is like O(N^3). Then I look at it for like a minute, figure it out. When I explain to it the correct algo it always blows Claude's "mind". "WOW! This is so much better." The more I use AI the more impressed I am with my own intelligence. Some day Claude will convince me I reached AGI.

If you’re a “Software Engineer” and you don’t know how to bypass this then please pivot to being a Starbucks Barista because you’re ngmi

Riddle me this. WTF are we using tokens for? Token generation per month: Google: 1.3 quadrillion OpenAI: 650 trillion Anthropic: 650 trillion Other (e.g. Meta, OpenAI on Azure, etc) is also something material. Call it 3 quadrillion per month total or 100 trillion per day. Approximately half of these are used by programmers - so call it 50 trillion per day. There are ~50 million programmers on Earth and maybe half of them (generously) are responsible for generating all these tokens. This gives us 50 trillion / 25 million = 2 million tokens being generated per day by these people. 60 million per month. For reference, Microsoft Excel's core functionality is built with about 2 million lines of code (and around 40million if you include libraries, UI, integrations, etc). We're using around 10X more tokens today than a year ago, and on the order of 50X since 2024. If the world suddenly started using 50X the amount of underwear it was using two years ago - everyone would notice. You would probably be spending half your day in the bathroom doomscrolling updates from the CDC on the latest possible explanations for why the entire world simultaneously came down with a violent case of diarrhea. Yet token consumption is up 50X, and daily life has changed for at most 1% of the world's population? Just strange to think about...

@citizen906 For me it's not a harness anymore One step up. A commando tower to switch between different harnesses for different tasks

@Samuelflg1 Wdym “ai Os”?

I spent 2 months building the “proper” AI agent setup. OpenClaw on a VPS, locked down with Tailscale, everything carefully wired. It just wasn't working the way i wanted it to. Mainly because i loaded it up with 100,000 systems and high friction on visibility of files. Then I ripped it all out and started again with Hermes + obsidian, YOLO'd onto my laptop. Refined tiny tiny processes at a time. baby steps. the difference is like night and day. Now I have a real executive assistant. Openclaw is still plodding along, trying to work out the "proper way" but in the meantime, having fun with Hermes.

Multi-agent collaboration is amazing.

@OpenAI LET ME IN

Harness -> tool calls -> Multi Agents

@rseroter @addyosmani Evaluiert wird von einem getrennten Evaluator und von externen Verifikations-Tools; gemessen wird entlang expliziter Qualitätsdimensionen, und im Coding-Fall zählt oft nicht der Durchschnitt, sondern das Unterschreiten einzelner Mindestschwellen.

.@addyosmani shares hot takes: "A decent model with a great harness beats a great model with a bad harness" "The gap between what today’s models can do and what you see them doing is largely a harness gap." "A harness is a living system, not a config file you set up once" addyosmani.com/blog/agent-har…

@rseroter @addyosmani Nach welcher Metrik? Nicht nach einer einzigen Gesamtzahl, sondern: mehrkriterielle Scoring-/Critique-Struktur bei Design Threshold-/Gate-basierte Bewertung bei Coding: fällt ein Kriterium unter die Schwelle, scheitert der Sprint. � Anthropic

There is still so much opportunity to improve / productize the onboarding flows for these coding models. And I think if we do the market may expand another 100x. I’ve been blown away by the power of tools like Claude Code / Codex and find myself building more software than ever before. But exploring the full capabilities of these tools still feels a bit like groping in the dark (esp for someone not very technical). What are all of the slash commands, best practices, tricks and tips? More ‘normal’ people should be building with these tools directly - but we’ll need to meaningfully decrease the real (and perceived) levels of complexity one faces just jumping into Terminal and firing one of these up.

I don't know that much about GPU programming. That's the whole reason my courses work. I'm 22, self-taught. CUDA course on FreeCodeCamp, LLM from scratch with 1M+ views, book with Manning. Built all of it while learning the material myself. Someone who's been training neural nets for 10 years skips steps they've internalized. They forget what's confusing. I haven't had time to build those blind spots yet. That's a moat with an expiration date. What would the world not have if I didn't exist? Right now I know the answer. People learned GPU programming from first principles because of things I made. That's not nothing. I want to keep going. Problem is I have so many options right now that it's immobilizing. Frontier lab. Talked to a few, vibe is always the same loop of train and release. Not a lot of room for new people to push ambitious stuff. Good for the resume but will resumes matter in a few years? Startup. Wrote something on RL environment scaling recently, got real investor interest. But I've never run one and I don't know if I can healthily pour a year or two into something that uncertain. Education full time. Models get better every month, billions of people can't keep up, someone needs to keep the ladder down. But it doesn't pay. Job and courses on the side. Burning out in 6 months. Join an early stage company. Another option. More paralysis. Then there's the stuff that genuinely excites me at a civilizational level. xAI building Colossus in Memphis. Google's Project Suncatcher putting TPU clusters in orbit powered by the sun. Starcloud training nanoGPT on an H100 in space. Tesla's plans for manufacturing at terafactory scale. Harnessing energy at that scale for our species is a dream of mine. I've thought about ruthlessly pursuing this on my own but I'm not sure I could handle that kind of pressure, and I have no hardware experience. Or maybe my role isn't to build the orbital data center. Maybe it's to inspire the people who will, through education. I don't know. I think about ikigai a lot. What you love, what the world needs, what you're good at, what you can be paid for. Money is the easiest branch if you target it. Passion is the one that makes the final call. And it keeps pointing back to education. I need a hard main quest or I get depressed. I turned down a life-changing equity offer to stay independent so there's no safety net. And I'm 22. I still want to travel and have a life. I was the most useful as a teacher when I was early and didn't know anything. The curiosity was real. The more I learn the further I drift from the student. Maybe I should lean into that while it lasts. My favorite song is from Stellaris. It's called Synthetic God. Never played the game, but I can grind all day listening to it on repeat with a custom mixer I wrote from scratch in Rust. That's kind of who I am. Build the tool, learn the thing, teach it to others. I just need to figure out how to make that sustainable. But I also need to eat. What would you do?

gpt wrapper to harness is the greatest rebrand of the decade

they’ve completely turn Opus 4.6 into a vegetable.

Anyone else get the urge to constantly restart their AI related systems and "start over"? Kinda like a video game, like a bad start in Civ where you wanna go again? Like I wanna do it with my openclaw agent, hermes agent, my obsidian vaults / karpathy style wiki things My claude code projects that get unruly Everything It's a constant desire to just burn everything and start fresh Anyone else?