Merlow

Richard Feynman stood at the same Cornell blackboard in 1964 and explained the word every AI lab minimizes in 2026: entropy. The BBC filmed it. Almost nobody watches it. Lecture 5. "The Distinction of Past and Future." The question sounds childish: why does time run one way. The answer is the exact quantity inside the cross-entropy loss every training run on Earth minimizes tonight. He does it with a glass of water and blue ink. Mixing is easy. Unmixing never happens. Order falls into disorder, and that 1-way fall is what "past" means. Watch the middle section. He explains why you remember yesterday and not tomorrow with counting arguments a 12-year-old can follow. No equations for the first 20 minutes. 1 piece of chalk. Shannon borrowed the word from physics in 1948 for the paper every LLM stands on. Feynman shows what it meant before it became a number on a dashboard. You minimize entropy for a living. He explains it for free.

@xwhynoted Exactly, and that's why I looked this up for you

@exeMerlow one piece of chalk beating every slide deck in AI research

@0xMoysei That makes me very happy, my friend

@exeMerlow I'm amazed by how well you do it!

@thelichhh Thank you. I analyzed it specifically and did my best for you.

@exeMerlow Wow, a new artifact from Merlow

Richard Feynman stood at a Cornell blackboard in 1964 and explained the problem every AI lab is fighting in 2026. The BBC filmed it. Almost nobody watches it. The lecture is about why nature only answers in mathematics. Every team forcing language models to reason is hitting the wall he mapped 62 years ago. He was 46. The Nobel Prize came 11 months later. The footage survived on film reels and now sits free on YouTube with fewer views than a keyboard unboxing. Watch the blackboard section near the middle. He takes 1 of Kepler's laws and rebuilds it from nothing, with notation a 12-year-old can follow. No slides. No jargon. 1 piece of chalk. An ML engineer I know paused it 4 times and made his whole team watch it before standup. You're 62 years late. The lecture is still free.

@0xMoysei Thank you, I tried to find something useful.

@exeMerlow Wow, it's cool that you're sharing this -I wouldn't have found it on my own.

He didn't believe the Kimi K3 hype. So he ran it against Claude Fable 5 on camera. Same prompts, both models, 12 minutes, results raw on screen. No cherry-picking, nowhere to hide. The context makes it spicy: Moonshot itself admits K3 trails Fable 5 overall while posting wins on coding benchmarks. 2.8 trillion parameters against the model everyone pays for. Benchmarks are marketing until somebody runs them at home. This is somebody running them at home. Watch the middle section where the outputs go up side by side. The gap is visible before he says a word. The verdict lands in the last minute. No spoilers. Leaderboards don't write code. Models do.

A YouTuber who burns $1,000 a day on AI tokens just did the same work for $63. His terminal froze on camera from running too many agents. He force-quit and kept filming. He skipped a live event to finish this. The model is Kimi K3, China's new 2.8 trillion parameter release. Weights go public July 27. In the 41-minute breakdown he feeds it his old production codebase with a paragraph and a half of instructions. It clears 122 tasks before he touches the keyboard again. Then around minute 37 it does something to his to-do list he says he's never seen from any model. Not even Anthropic's flagship. The security audit near the 33-minute mark is the part that should worry you. Frontier models refuse that task. This one ran 25 verification agents on it. He came in bored of open weights. He left calling it frontier class. Hosting it yourself costs $2.6 million. Watching him break it costs nothing.

@0xObssnnn The New Alpha in the AI Industry

A web developer gave Claude Code full SSH access to 2 machines on his desk. It shipped what an AI startup teased on Twitter for months and never released. He says it on camera: "I'm not a systems programmer. I don't know anything about Rust networking code." Claude did. It compiled inference engines from source on 2 different chip architectures and wired an NVIDIA box to a Mac. Then everything broke. Machines refusing to see each other. A discovery protocol that turns out to be silently broken on macOS. Watch him find the real bottleneck with 1 terminal command around the 6-minute mark. 96% of the time was going somewhere nobody suspects. The final setup streams tokens 6x faster than a $4,000 NVIDIA box manages alone. There's a chart near the end he calls the upside down middle finger. It earns the name. Startups sell roadmaps. A web developer with Claude shipped the thing.

He ran Karpathy's method on his own notes. 1 query found a monetization gap hiding in 2.5 million views. Karpathy named the skill context engineering: "the art of filling the context window with just the right information." Everyone quotes the line. This guy applied it to his entire head. The clip opens on an Obsidian graph, months of notes in one web. His context window, pre-built. He points Claude Fable 5 at the whole vault. The model finds 3 clusters that never touch. His clips pull the most views, his notes barely mention them. 2.5 million views flowing past a gap that converts close to 0. Then Claude writes the finding back into the vault as a new note, with execution laid out phase by phase. Everyone engineers prompts and gets generic answers. He engineered context and got a business plan.

A 26-year-old built a machine that runs a whole payroll of AI agents at his desk, and the fleet works for free. The video shows the shift in progress: a monitor split into a dozen terminal windows, processes scrolling like a small company's server room. A search agent digs through his files. A RAG agent answers questions over his own documents with receipts. A code generator writes while the other two think. All at once, all local, all night if he wants. His friends pay for the same workers by the token. Agents burn tokens around the clock by design, that's the whole point of them, and renting that appetite is how $200 monthly bills are born. His bill is electricity. The machine behind it holds the one resource 2026 fights over: memory. DRAM jumped 90% in Q1 as datacenters devoured the supply, and every box that can host an agent fleet repriced upward with it. His was bought before the climb. The outputs stack up while he sleeps. Research done, documents answered, code drafted, waiting at breakfast like mail. The video cuts mid-sentence: and if you want to do this... The answer hums on the desk behind him. Everyone else hires AI by the month. He built the office it works in.

@0xMoysei how can this not be popular yet?

I found something very interesting. Obsidian quietly shipped the missing piece for AI agents. The official repo sits under 200 stars while the timeline argues about memory and context windows. It's a headless client. 1 npm install, 1 command, and your vault syncs from any server. No desktop app. Encryption intact. Buried in Obsidian's own docs, 1 line: "give agentic tools access to a vault without access to your full computer." Read that again. The company wrote AI agents into the official use cases. A friend runs this on a $599 Mac mini. Claude Code writes his daily notes, tags them, rolls them into weekly summaries on a cron job. He opens his phone and they're just there. He hasn't typed a note in 3 weeks. Setup takes 10 minutes. The docs list every command. The agent sees the vault and nothing else. Karpathy called Obsidian the IDE. Obsidian just handed the programmer the keys.

A 21-year-old turned his dead crypto mining farm into an AI machine that pays him $19,000 a month. The video walks the aisle he almost sold for scrap: 4 racks, dozens of machines, fans wall to wall, green lights blinking in the dark. He started mining at 17. When the margins died, everyone around him liquidated. Buyers offered pennies for the racks, so he left them plugged in out of spite. Then AI repriced everything he owned. DRAM jumped 90% in Q1 as datacenters devoured the memory supply. The cards and boards collecting dust turned into the exact hardware the market fights over. He reloaded the farm with open models and rented the compute out: agents, drafts, overnight jobs for people who pay $200 a month per seat elsewhere. His costs stayed what they always were, electricity and silence. The racks that once gambled on block rewards now invoice like a small company. Same room, same wiring, same fans. The camera holds on the glowing aisle and cuts. Everyone sold their farms at the bottom. His spite turned out to be a business plan.