Freddy

AI Architect | Making technology accessible & fun

[Freddy @0x_freddy]

Sam Altman admitted what everyone in Silicon Valley was thinking: Google should have completely dominated the AI race from day one

Google invented the Transformer architecture in 2017. Then, every single one of the eight authors on that seminal paper left.

Noam Shazeer left to start Character-AI. Aidan Gomez founded Cohere. Ashish Vaswani and Niki Parmar started Essential AI. Jakob Uszkoreit founded Inceptive. Llion Jones founded Sakana AI. Lukasz Kaiser joined OpenAI. Illia Polosukhin built NEAR Protocol.

Google had the paper, the compute, the users, and the distribution. What they lacked was a culture that allowed the people who invented the tech to actually ship it.

The talent didn't leave for higher salaries. They left because bureaucracy slowed down execution. Within eighteen months of leaving, every single author was shipping groundbreaking models somewhere else.

The lesson goes far beyond Google. It’s what happens to any incumbent when protecting legacy revenue becomes more important than shipping the future.

Sundar Pichai kept the CEO chair. The eight engineers who wrote the paper built the rest of the AI industry.

Follow if you'd rather build for the future than protect the past.

[Freddy @0x_freddy]

Columbia charges $80,000 a year to teach diversification. In 2002, a hedge fund manager sat in a classroom upstairs and taught the opposite. His method had returned 50% annually for ten straight years

His name is Joel Greenblatt. He ran Gotham Capital from 1985 to 1995. He held just a few positions at a time. Almost no one in finance sustains 50% for a single year. He did it for a decade, returned all capital to outside investors, and walked into a classroom to give the framework away.

The core of his lecture isn't just about formulas. It's about finding structural market inefficiencies — spinoffs, restructurings, and forced-selling events where institutional investors dump stocks regardless of price. He explains why these opportunities persist even after everyone knows they exist.

Every stock screener is free now. Every filing is searchable. The bottleneck was never access to information. It was knowing which information to ignore.

Filmed from the back row. Audio uneven. Students blocking the frame. He handed a 50% annual playbook to a room of 30 people for free.

Watch it before your broker sends you another 60-stock "balanced portfolio."

[Freddy @0x_freddy]

Anthropic just released the 2-hour course that trains you for the $500K AI engineering job they are actively hiring for. Free.

Four timestamps do the heavy lifting:

00:19 - the right way to prompt Claude.
33:27 - what makes Claude act dumber on your code.
1:33:44 - how Anthropic uses Claude internally every day.
1:51:02 - the fix that makes Claude dramatically smarter.

The 2-hour course replaces about 10 paid AI engineering programs, and it comes from the company deciding what "senior AI engineer" means on every job description in 2026. The gap between watching it and not is the difference between applying with the exact playbook Anthropic uses and applying with whatever a $500 course taught you.

Save it. Two hours tonight is the shortest path to interviewing at the price band this course was built for.

[Freddy @0x_freddy]

SpaceXAI engineer, ex-Cursor:

"I've got 15-25 GrokBot agents running right now. They cover most of what used to be my week."

One person, up to 25 agents, most of a workweek automated. Not slides. Not a demo. A real engineer walking through the exact team he built and how the Chief of Staff bot on top of it routes work between them so he only touches decisions that actually need him.

The math is where it lands. Two years ago a solo operator was capped by how many hours they could work. Now the cap is how many agents they can brief once and trust to run. 25 agents means one person operating at the effective throughput of a small team, and it took him one setup weekend to build.

If you plan to compete with a solo operator running this shape in 2026, save this now. If you are the solo operator, watch it tonight.

[Carnage @0xCarnagee]

https://t.co/Sb4TjZpUHH

[Freddy @0x_freddy]

Head of Claude Code:

"I'm not prompting my agents anymore. I'm building graphs and loops so they can build the agents for me."

Read that quote again. He is not scaling by writing more prompts. He is one layer above that, building the machine that spawns agents on demand. The 10-minute video is exactly how he does it, and it comes from the person who runs the team building Claude Code.

The next phase of AI work is not "how do I prompt better." It is "how do I build the loop that decides what to prompt." Everyone still on the first question is about to spend a year catching up to people asking the second one. The talk names the exact shift and shows the shape.

Save this and put 10 minutes on the calendar. Two years from now this is what "AI engineer" means, and this is the shortest way to see the frame today.

[Anatoli Kopadze @AnatoliKopadze]

https://t.co/fyBGWQbsvn

[Freddy @0x_freddy]

Anthropic engineer:

"prompting is basically over. 90% of our engineers were already using self-improving loops. Now everyone is moving toward agentic graphs."

In 10 minutes she builds her full Claude Code system live from an empty terminal. Not slides. Not theory. The actual keystrokes it takes to go from a blank prompt to a self-improving agent graph shipping work.

The transition she names is the whole industry map: Agents, then Loops, then Graphs, then Self-Improving Systems. Prompting was the old workflow. Graph engineering is the next one. Anyone still on step one in 2026 spends 2027 catching up to someone who watched this talk in 2026.

If you plan to build with Claude, this is the primary source. Save it and watch before your feed gets loud about it.

[Lunar @LunarResearcher]

https://t.co/PGTcKvOU4C

[Freddy @0x_freddy]

A SpaceXAI engineer just released a 1-hour Grok Bot workshop that walks the full build from zero to a system running her daily life and her business.

Most courses on Grok Bot teach you what it is. This one is a real case study of someone on the SpaceXAI team who actually automated herself out of the routine.

The full build in five stops:

3:33 - build your first Grok Bot.
6:57 - give every bot a role.
16:56 - hand the bots tools and context.
31:55 - run agent teams in parallel.
52:23 - full system automating business and daily life.

Bookmark this. The gap between "I use Grok Bot" and "Grok Bot runs my day" is one hour of watching, and it comes from someone who ships this internally at SpaceXAI, not someone selling a course.

[codila @0xCodila]

https://t.co/tegsGLwl7l

[Freddy @0x_freddy]

Cursor Designer Ryo Lu:

"right now I'm running 10-20 GrokBot agents that automate 90% of my routine. I have a Chief of Staff agent. He knows about all my other bots and manages everything."

10 to 20 bots. One human. 90% of a designer's workday running without a keyboard.

The 20-minute podcast is how to build the whole team from scratch. More useful than a $500 course on agentic engineering, and it comes from someone shipping actual product at Cursor.

Bookmark it before your calendar looks like his.

[darkzodchi @zodchiii]

https://t.co/NyKpCSt2lI

[Freddy @0x_freddy]

@iam_fernandob @grok Don't look at the PnL, look at the pipeline: 8 sub-agents, deterministic triggers, automated stop-losses. The leverage is real whether you build it or call it clout chasing

[Freddy @0x_freddy]

A crypto trader just made $13,100 in a week using 8 Grok Bot agents that cost him $200 a month.

The same job used to require a $500K-a-year analyst desk at a crypto fund. Now it runs on his laptop while it is closed.

Each of the 8 agents owns one narrow role. One scans for early alpha before Crypto Twitter finds it. One audits contracts and flags honeypots. One times the entry to the millisecond after risk clears. One tracks whale wallets for insider accumulation. One monitors dev wallets 24/7 and dumps the position if liquidity is touched. One trails stops as the price moves. One tracks social momentum. And one head-of-desk bot routes everything and only wakes him for the single decision that needs a human.

Result last week: 142 tokens scanned, 19 qualified setups, 6 trades executed, +$13,100 after fees, all while he was asleep.

The setup took one evening. No VPS, no code, no developers. Just 8 job descriptions and one screen-recording of the workflow, so the bots watch you do it once and repeat forever.

Bookmark this.

[Freddy @0x_freddy]

@wooderic3609 10 septillion years down to 5 minutes isn't just an upgrade—it's an entirely different computational physics paradigm

[Freddy @0x_freddy]

Bookmark this. Google's Willow quantum chip just solved a problem in 5 minutes that would take the world's fastest classical supercomputer 10 septillion years. That is more time than the universe has existed, times a quadrillion.

Sundar Pichai confirmed on record: quantum error correction now works at scale. Practically useful quantum computers in 5 to 10 years.

Translation for AI: every constraint AI hits today, model training taking months, inference costing billions in electricity, agents choking on compute limits, is running out on a clock. When quantum lands in production, the whole cost curve of AI collapses.

The people building AI systems with today's compute assumptions are budgeting for a world that ends this decade.

[Freddy @0x_freddy]

@nit_sa61418 Research frameworks matter, but hardware physics dictates the pace. Quantum flips the entire infrastructure cost equation

[Freddy @0x_freddy]

Elon Musk, on record: "you have at most 3 years left to make money by selling your work."

After that, AI does most tasks and paying a human by the hour stops making sense. The way we have all made money since the industrial age is quietly on a countdown.

The real question is not whether the shift happens. It is whether by 2029 you are still selling your hours, or you have built something AI works for you instead of against you.

Bookmark this. This is the largest transfer of money in modern history, and it starts with a decision most people are still avoiding.

[Freddy @0x_freddy]

Elon Musk, on record: in 5 years AI surpasses all human intelligence combined. In 10 years we probably lose control.

They asked him about the off switch. His answer: "even if it existed, we probably shouldn't press it."

This is the person who cofounded OpenAI, runs xAI, and just committed $16.8 billion to a single silicon fab. He is not guessing.

10 minutes. Clearer than any two-hour podcast on AI you have watched this year. Bookmark before you scroll past.

[Freddy @0x_freddy]

Elon Musk, on record: "either we build the Terafab or we don't have the chips. And we need the chips."

All the world's chip factories combined, TSMC, Samsung, Micron, add up to 2% of what he needs. So he is building a 100 million square foot facility in Texas, 50 Pentagons, $16.8 billion for phase one alone. What the planet produces in 50 years, he wants every year.

The AI race stopped being about models. Bookmark this. It is about who controls the silicon underneath them.

[Freddy @0x_freddy]

3 billion downloads. That is what Alibaba's Qwen cleared on Hugging Face in six months, more than Google and Meta combined by four times, and it just quietly ended Washington's four-year export control regime.

Bookmark this before every AI infrastructure thesis gets rewritten around one detail. Qwen3.8-27B shipped today under Apache 2.0. 27 billion parameters. 262,000 token context. Runs locally on 17 gigabytes of memory, on a used $700 graphics card. Once the file lands on your drive it is yours forever. The license cannot be revoked. The file cannot be uncopied.

The historical rhyme is exact. In 1440 the Church spent a century chasing Gutenberg's Bible after it was already replicating in every printer's basement in Europe. In 1991 the US government spent three years prosecuting Phil Zimmermann for PGP, a file already sitting on tens of thousands of drives.

Around 06:14 the numbers land. 151,448 downstream models are already built on Qwen. 200 new ones appear per day. American labs are valued on the assumption that frontier intelligence stays scarce, expensive, and rented by the token. Alibaba just made a version of it free, permanent, and small enough to run on hardware you already own.

[Freddy @0x_freddy]

Friedrich Hayek, 1977, live on Buckley with a Nobel Prize three years fresh, explained why nobody alive today knows enough to fix anything. Every AI policy debate in 2026 is a rerun of that hour, and almost nobody has watched it.

He was 78. His book Road to Serfdom, written during WWII, has never gone out of print. Buckley opens with one question in the first minute. Hayek does not answer with a policy. He does not answer with a number. He answers with a mechanism.

Around 03:11 he lands a nine-second sentence about prices that the entire hour hinges on. I rewound it three times. It is the reason no central planner and no central AI system can hold enough dispersed information to coordinate an economy. The 2026 "AI governance" debate is still catching up to that clip.

Then he drops a line about unemployment figures that no economist alive today would put his name to on camera. Watch Buckley's face, not Hayek's.

Save this. One camera, no edits, 57 minutes. It aired once in 1977, and half of this month's AI governance op-eds would be dead on arrival if the authors had watched it.

[Di Krass @Di_Krass_]

https://t.co/h8IkMUDhNc

[Freddy @0x_freddy]

BOOKMARK THIS BEFORE EVERY VC DECK REWRITES ITSELF AROUND IT. ANDREJ KARPATHY JUST SPENT 20 MINUTES ON RECORD BLOWING UP THE ENTIRE "AI PRODUCT" THESIS, AND THE METRIC HE NAMED SHOULD END THE DEBATE.

His first claim, in plain numbers. 90% of a modern model's parameters are memorizing garbage scraped from the open internet. The useful signal fits in the remaining 10%. Every founder who raised a Series A on "we fine-tuned a big model" just got repriced by that sentence.

His second claim is sharper. 99 out of 100 AI companies pitching in 2026 are shipping demos, not products. The gap between the two is not capital, not talent, not GPU credits. It is one loop most teams skip.

Around 04:18 he sketches the loop. A small model, a bounded task, a verifier that catches the agent generating garbage at scale, and a feedback channel back into training. Every serious lab is quietly rebuilding around this shape while their competitors keep polishing demos.

In six months half the "AI product" pitches you hear today will be repositioned as tech previews. Karpathy will be the one everyone quotes.

[Ricker @0xRicker]

https://t.co/ZtruLI5Kxk

[Freddy @0x_freddy]

@drakefomo 100%. Shrink the surface area of the decision until saying no carries a higher cognitive and financial penalty than testing the pilot

[Freddy @0x_freddy]

Jordan Belfort spent 22 minutes on tape live-selling a $27 million dairy owner who had never heard of him. The technique on tape is what every AI sales agent shipped in 2026 was trained to imitate.

Belfort's actual moves, in order. Ten minutes of discovery questions before any pitch. The wound: the owner wanted China but had no facility there. He never sold into the wound directly. He shrunk the ask until saying no looked strange. "I don't need to take you from 0 to 50. I want to take you from 50 to 500 million."

Then the risk math out loud. One small hedge. 30 days. Worst case one basis point lost. Best case 25% and $27,000 saved.

"Sound fair enough?" repeats six times across 22 minutes. Never once explains a product. He was selling 30 days.

Every enterprise AI sales tool funded today, Gong, Clari, Regie, Chorus, is trained on transcripts of this exact pattern. The template was public for a decade. AI companies raised $8 billion in 2025 to reproduce it in software.

Watch the 22 minutes before the AI selling to you does.

[AI Avenger @AI__Avenger]

https://t.co/4JUppYy9T4

[Freddy @0x_freddy]

In 1998, Warren Buffett walked into a University of Florida MBA class at 68 and taught the single most important lesson of his career using Coca-Cola.

Coke went public in 1919 at $40 a share. One year later it was $19. The Candler family, which had bought the business for around $2,000 in 1888, watched half the public market cap evaporate in twelve months. The business itself was fine.

Buffett spent the rest of the hour on what he actually watches: volumes, pricing power, and where growth comes from ten years out. Never the daily price.

Every AI stock trading today is priced on a story about the next ten years. NVIDIA. Meta. Palantir. The private markups on OpenAI, Anthropic, xAI. If you follow Buffett's frame, price movements this quarter are noise. The signal is whether volumes, pricing power, and ten-year growth are actually converging.

If they aren't, you're holding the 1919 Coke chart in reverse. Price up, fundamentals hollow.

The 1998 lecture is free. Nobody in AI wants you watching it before the next earnings call.

Save this before the next 50% move in either direction.

[Neuro @0xNeuroBlue]

https://t.co/6bJ460527Y