Tech Friend AJ

Your RTX 3090 ships with about half its prefill speed locked away. @0xSero explains how, plus the 81GB model he runs on a base M4 Mac mini with a 3090 over Thunderbolt 5. We also cover his hacked X account, the OpenCode hack, and why chips ship locked. 0:00 Cold open 1:06 Local models are good enough 2:11 Skills that make local models repeatable 4:01 Having models write you a wiki 5:21 AJ's video editor and Opus 5.5 6:29 Growing on X 8:03 Meta's Project Lazarus 9:21 Why X is different 10:30 Binance hack and digital cleanup 12:11 AI slop vs human slop 13:07 Faceless 3D videos vs human content 15:40 Social media funnels 16:47 Why Sero doesn't make short-form 17:31 AI feeds and three-minute learning 20:00 Management work, games and reading 21:45 BB and next-step buttons 22:44 Claude artifacts as a brand system 24:19 Where Opus still falls short 26:02 Claude watermarking its text 27:53 Why Binance prices barely moved 28:59 What to do with $6B 29:21 Living in San Francisco 30:15 Teaching people to use agents 34:17 Getting into public speaking 35:55 DHH's OS built for agents 37:01 How Sero's X account got hacked 38:55 An agent exposed Sero's file system 39:35 How OpenCode got hacked 41:08 What Binance does next 41:37 Anthropic, xAI and Colossus 43:17 Data centers offshore and in space 44:02 Vera Rubin and ASIC chips 45:57 Etched and disaggregated inference 48:53 NVMe offloading and GPU acceleration 51:36 NVMe as cheap RAM 55:07 eGPU: Mac mini + RTX 3090 57:38 Recipes: models as game cartridges 1:00:18 The hidden RTX 3090 speed 1:02:18 Why hardware ships locked 1:03:18 NVLink and multi-GPU interconnect 1:05:58 How chips are manufactured 1:07:54 Adding VRAM to 4090s and 5090s 1:11:09 Have a model crack your GPU 1:11:56 DeepSeek and GLM for reverse engineering

I turned one image of my AI cohost into two sprite sheets. ChatGPT image generated both from a single picture. I'm using them to animate him.

I realised timestamps probably wont work during live premiere but should work after! – at San Francisco, CA

ok first time ive had a major issue with bb > user : what next > opus 5.5 high : 1000+ CLI bb Processes might have to try @aicodeking IDE – at San Francisco, CA

You can watch/listen to our chat with timestamps at the premiere: youtu.be/1LK-k3noEbY or on x at ~9pm SF time – at San Francisco, CA

only been up since the challenge started ! 🚀 but are we on track?? @grok please calculate

i havent opened vscode on purpose in forever

old mate should just be doing first mate. new me 30 seconds later doesnt do multiple threads

hand offs > forks @grok debate 3 points each

first time ive deprecated a plugin

+1 one-prompt @get_bb_app plugin cross harness slash commands + bonus UI

I used 2.2 billion tokens in one day. It was for Swyx's Kill My SaaS competition. I looped agents on conference scheduling software and had adversarial verifier agents check the work until it hit the target state. What is the most tokens you have burned in a day?

After this chat I was inspired to take cyber security more seriously. Sero says he has an agent do a >daily routine check and audit of his home device/network. I am setting one up for myself now.

I built the same landing page 14 times to test 5 anti-slop design skills. 9 of them wrote the same headline. Two skills never ran until I named them in the prompt. Hallmark was the cheapest run and my favourite page. Which design skill should I test next?