Leonn

AI Researcher | Crafting seamless agentic experiences

[Leonn @0x_Leonn]

An MIT professor offered his class a coin flip. Win $125, lose $100. Expected value: +$12.50 per bet. Most of the room refused. Then he priced their refusal.

He asked how much they'd pay to skip the bet if he forced them into it. The average answer was $43.

A rational agent would take the bet every time. Real humans hand over 34 cents of every dollar they might have won to avoid the volatility. That gap has a name. Utility of wealth. First formalized by Daniel Bernoulli in 1738.

The insurance industry is $9 trillion a year of that gap, quantified, priced, and collected. Every car policy, health premium, and warranty is a company selling you the option not to flip the coin. You've been paying for it since your first paycheck.

The uncomfortable analytical implication: your monthly premium is not the price of the risk. It's the price of your fear of the risk. The insurance company keeps the difference. That's the whole industry.

The professor is teaching what an actuarial team already priced in.

Watch the demo before the next premium hits your account.

[Zynex @0xzynex]

https://t.co/s7CSVJZhBN

[Leonn @0x_Leonn]

SpaceXAI engineer (ex-Cursor):

"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."

The ratio is what matters. One human, 20 workers, 90% of the workday running while the human decides.

The old software career shape was: you produce output, you get promoted for producing more output. The new shape is: you orchestrate output, and the ceiling on how much output you can direct is 20 to 100 times higher than what one person can type.

50-minute podcast breaks down the full build. Bookmark it before the ratio becomes the standard interview question.

[Codez @0xCodez]

https://t.co/eDBs67VMq2

[Leonn @0x_Leonn]

Bookmark this. Sam Altman just admitted he was wrong on AI timelines.

His words: "I thought when we got to GPT-4 that very quickly there was going to be much more disruption, software businesses up for grabs right away, than it turned out to be. The economy just has so much inertia. People keep buying from the same company, wanting to use their tools the same way. It's going to make this big transition go smoother and slower."

The market reading is that AI disruption is delayed and everyone can relax. That reading is a trap.

Altman is not saying AI is weaker than expected. He is saying incumbents are stickier than expected. The gap between what AI can do and what companies actually deploy it for is exactly the window an AI-native startup exploits.

Every year the economy fails to adapt is another year the incumbent's customer base is quietly getting replaceable, without anyone updating the incumbent's stock price. When it moves, it moves all at once.

[Leonn @0x_Leonn]

A guy just built a working company inside Grok Bot with 6 agents. It runs 24/7 while he sleeps.

Six agents, six narrow roles: research, writing, outreach, ops, finance, support. They hand off to each other automatically. A control layer catches anything that sends money, publishes, or deletes something, and pushes it to him for approval.

The shift is easy to miss. In the old setup you were the router. Every task looped back through you before the next agent could touch it, so "AI helps me" turned into "AI adds a step." Here the agents route to each other. You are only in the loop when something is about to become irreversible.

That is the difference between an AI chat tool and a 24/7 operations team behind a single dashboard.

Bookmark this. The bottleneck for solo operators just moved from "how many tasks can I do" to "how many decisions can I approve."

[NO1ennn @N01ennn]

https://t.co/Js2rD31D6T

[Leonn @0x_Leonn]

300 agents can generate 44,850 possible connections between each other. This is the AI scaling problem nobody is talking about.

More agents means more parallel work. It also means duplicate findings, shared sources, hidden overlaps. If each agent returns its own report, you end up with 300 piles to reconcile by hand.

The number that matters is not how many agents you run. It is whether their findings connect into one shared picture or stay in 300 separate stacks. A shared graph merges what agents find, groups the overlap, and hands you one connected map. Without it, more agents just means more noise per hour.

This is why teams stuck at 10 agents beat teams running 300. The bottleneck moved from parallelism to structure, and most people have not caught up. Bookmark this.

[Alex @de1lymoon]

https://t.co/queNtU0dWX

[Leonn @0x_Leonn]

An org chart, not a to-do list. That is the shift Grok Bot just enabled, and it took one week to change how founding actually works.

Eight bots. One org chart. The founder talks to one of them, Atlas, chief of staff. Atlas decomposes outcomes into tasks and delegates in group chats. It only escalates when a decision is irreversible or spends money.

Week one output from a single human running the setup: 214 verified prospects, 89 personalized outreaches queued, 11 content pieces drafted, inbox at zero every morning.

Two rules made this work. Every bot charter ends with a hard "never do this without asking" line. Autonomy without a fence is chaos on a schedule. And show once, don't describe. One screen recording of the workflow taught the bots more than a page of instructions ever would.

Bookmark this. Most people are still using Grok Bot as a smarter chat window. It is the first time one person can own an org chart instead of a task list, and the bottleneck stops being how much you can do.

[Leonn @0x_Leonn]

Dylan Patel just reported that Anthropic finished training Mythos 2 and will not release it. The internal loop building Mythos 3 never paused.

Translation for anyone reading model release cadence as a proxy for lab velocity: you are reading the wrong signal. A finished frontier model that ships nothing publicly, while a next-gen model trains on outputs from it, is the definition of a recursive self-improvement loop running in stealth.

The gap between what Anthropic ships to the public and what Anthropic actually has internally just quietly doubled.

[Leonn @0x_Leonn]

Head of Claude Code, on record: "85% of our engineers are running dozens or hundreds of agents. The way you do it is graph engineering."

Two numbers in one sentence rewrite the software labor market. 85% of engineers at the company that sells the model already operate at agent-cluster scale. And the discipline enabling it has a name most of the industry has not heard.

Graph engineering is the answer to a question managers have not asked out loud. If one engineer can hold hundreds of agents in flight, output stops being lines of code per day and becomes agents coordinated per hour. Every "senior engineer, 5-10 years" job description now describes a role that no longer maps to the work.

Around 07:36 he sketches how the graph is built. Nodes are agents. Edges are tool calls between them. The engineer's job is drawing the graph, not writing the code. The agents fill in the code.

Bookmark this. In 12 months the Anthropic 85% number will be the industry benchmark, and every hiring manager still counting lines of code per week will be pricing labor against a dead metric.

[0xRafy @0xRafy]

https://t.co/jh8D5LiVAT

[Leonn @0x_Leonn]

Google Cloud CEO Thomas Kurian: "we have very large Anthropic customers running on GCP." On OpenAI: "we would love to have them."

Bookmark this before every infra analyst repositions their thesis around it. While the market argues which model wins, Kurian just described the actual game Google is playing, one level below the model layer.

Three data points from the same talk. 42% of US mortgages now run through agents that read the form, extract the data, call the backend, and never touch a human. Gemini can now run air-gapped inside a customer's own datacenter on NVIDIA Blackwell, a build that took 18 months of joint engineering. 780 companies already crossed 500 billion tokens, and most of the models driving that volume shipped 18 months ago.

The pattern under all three numbers. Google is not competing in the model beauty contest. They sell the floor every model runs on, Anthropic's included, OpenAI's if Kurian gets his wish. Whoever owns the substrate collects the toll regardless of which brand wins the cycle above it.

Around 08:23 he says the quiet part. Model providers are a rotating cast. The floor is permanent.

[Leonn @0x_Leonn]

Google Engineer, on record: "I never think about a context window anymore."

Save this before the "prompt engineer" job title formally disappears. He works on their new agent framework, and the shift he just described is already shipping to production.

His tasks run 36 hours. The window empties and refills the entire time. He designed for that. The next generation of AI systems will not have a prompt in any meaningful sense. It will have tools that speak to the model at the exact moment a decision is about to be wrong.

Zero rules in the prompt. Zero instructions. Zero guardrails. Every constraint lands through tool outputs, precisely when needed. This is the architecture pattern every serious lab will ship by mid-2027, and Google is 18 months ahead.

Around 09:52 he lands the line the whole team lives by. Quote: "we want to make every mistake something that is just not possible." Not caught after the fact. Not warned about. Structurally impossible.

The founders who move their agents to this pattern this quarter will be building on 2027 architecture in 2026. Everyone else is polishing a discipline that already has an expiration date.

[vartekx @vartekxx]

https://t.co/TMvjSSRzVu

[Leonn @0x_Leonn]

Andrej Karpathy just told an audience to delete everything except the graph.

His full framing: prompting is fading away. LLMs, prompts, agents are all steps. The graph is where the system ends up. Everything before it is scaffolding.

Read that against the AI education market. There are roughly 4,200 active AI courses on Coursera, Udemy, and Maven combined. The overwhelming majority teach prompting as the destination. Graph engineering appears in fewer than 40 of them, most published in the last quarter. Karpathy just called the destination a step.

The analytical implication: the $3,000 certification you paid for last year credentials a layer about to be commoditized. Every AI role that says "prompt engineer" on it has a replacement being drafted right now inside labs that follow Karpathy.

The lecture is 1 hour. Free. It contains the exact vocabulary and diagrams that will show up in senior AI job descriptions by Q2 2027.

The people watching it this weekend are calibrating against the roadmap. The people scrolling past it are calibrating against a curriculum that stopped updating.

Watch it before your next skills audit uses last year's language.

[rari @0xwhrrari]

https://t.co/gotSPGX4Nz

[Leonn @0x_Leonn]

@gramps_dad True for compliance, but the pricing power still relies on the exact same Bernoulli curve. The government mandates the floor; the insurers monetize everything above it through your baseline fear of variance

[Leonn @0x_Leonn]

An MIT professor offered his class a coin flip. Win $125, lose $100. Expected value: +$12.50 per bet. Most of the room refused. Then he priced their refusal.

He asked how much they'd pay to skip the bet if he forced them into it. The average answer was $43.

A rational agent would take the bet every time. Real humans hand over 34 cents of every dollar they might have won to avoid the volatility. That gap has a name. Utility of wealth. First formalized by Daniel Bernoulli in 1738.

The insurance industry is $9 trillion a year of that gap, quantified, priced, and collected. Every car policy, health premium, and warranty is a company selling you the option not to flip the coin. You've been paying for it since your first paycheck.

The uncomfortable analytical implication: your monthly premium is not the price of the risk. It's the price of your fear of the risk. The insurance company keeps the difference. That's the whole industry.

The professor is teaching what an actuarial team already priced in.

Watch the demo before the next premium hits your account.

[Zynex @0xzynex]

https://t.co/s7CSVJZhBN

[Leonn @0x_Leonn]

Travis Kalanick just described the exact operating mode every AI CEO is currently in.

His words: "I was running a $70 billion company the way somebody who thought he was going to starve next week would run it."

Read it against the current AI industry. OpenAI, Anthropic, xAI, Perplexity, all valued between $10B and $500B. All led by founders operating with the intensity of people who spent a decade being told AI wouldn't work.

Kalanick's admission continues: the same starvation-mode precision that built Uber destroyed him at scale. "When you're big, the expectation is that you don't run that close to the line, even if it's correct. That is a thing I definitely did not understand."

The AI industry is currently at the "correct but reckless" stage. Every OpenAI board fight, every Anthropic policy that shocked users, every Meta AI product shipped over safety pushback fits the pre-Kalanick pattern.

The reckoning isn't "did AI companies do wrong." It's "were they too big for the intensity that built them." Every one of them will hear "yes" on a public stage within 24 months.

Watch Kalanick's interview before the next AI CEO gets subpoenaed.

[Leonn @0x_Leonn]

@GuntherWrite And the last two years were just the pilot. Once AI-driven abundance completely decouples labor from survival, the cultural redirection will be 10x weirder

[Leonn @0x_Leonn]

Bret Weinstein just packaged an analytical framework that predicts the next 20 years of culture. The mechanism is AI removing the cost from work the same way the pill removed the cost from sex.

Evolution built humans around two central drives: mating and resource acquisition. Reliable birth control decoupled the first one in the 1970s. Sex was cashable without the 20-year cost of raising a child.

Weinstein's observation, backed by his wife Heather Heying, is that the biology didn't disappear. It rerouted. The maternal-protective neurochemistry evolution built to guard a toddler now defends abstract ideas with the same intensity.

Then AI enters. Elon just said what many are quietly modeling: AI-driven abundance is about to do to money what contraception did to sex. Decouple the reward from the cost. The evolutionary carrot of resource acquisition stops requiring resource acquisition.

If Weinstein's model holds, both drives that structured adult life become simultaneously optional. The redirect target of that biological energy is the entire prediction market for 2028.

Watch the 3-minute clip before your next argument about "why is everyone so intense right now."

[Leonn @0x_Leonn]

@SprintTalon The math is absurd, but that exact delta built a $9T market

[Leonn @0x_Leonn]

@buildstack_ai When you strip human panic from execution layers, risk pricing becomes pure algorithmic optimization instead of monetizing anxiety

[Leonn @0x_Leonn]

@Inflowzy Exactly. Peace of mind is just the consumer label for an actuarial premium on loss aversion

[Leonn @0x_Leonn]

Andrej Karpathy just compressed 8 years at OpenAI and Tesla into a single 2-hour lecture. The bootcamp industry teaching half of what he covered charges $15,000 for the privilege.

Karpathy co-founded OpenAI in 2015. Ran AI at Tesla from 2017 to 2022. Went back to OpenAI. Left again in 2024. His current company Eureka Labs was founded to fix exactly this education gap.

The analytical value isn't the content. It's the audit. If you follow the whole 2 hours, you know where you stand. If sections lose you, you have a target list. Karpathy structures each unit so the concept, the code, and the intuition are testable against your current understanding.

The $15K bootcamp industry has one problem now. Their curriculum was built by people who read papers Karpathy wrote. He just released the source material. Every course selling "AI engineer certification" at $3,000 to $15,000 is now selling a translation of a lecture the original author gave away.

Watch the whole thing. Note where you stopped following.

That note is a more accurate resume assessment than anything on your LinkedIn.

[rari @0xwhrrari]

https://t.co/gotSPGX4Nz

[Leonn @0x_Leonn]

Google's Chief Scientist Jeff Dean spent an hour compressing 27 years of internal Google AI infrastructure into one lecture. The gap between what he covers and what most AI courses teach is where the market's salary bands come from.

Dean isn't a keynote speaker who read a summary. He was Google employee number 25. He built MapReduce, BigTable, and TensorFlow. Every piece of infrastructure the modern AI industry runs on has his fingerprints somewhere.

The lecture walks the full stack. LLMs from first principles. Prompt engineering as it's actually practiced. Multi-agent architectures. And the final section, where one human coordinates 100 agents at once. Bootcamps charging $3,000 stop three chapters before Dean starts.

The uncomfortable analytical fact: the lecture is the shortest audit of your own AI skill gap that exists. If you can't follow a section, you have a target. If a section feels obvious, you know where you stop being useful in an interview.

The material is free. The audit takes an hour.

Save this before your next AI project scoping meeting reveals the gap in a room instead of a lecture.

[Codez @0xCodez]

https://t.co/WVztfCUH4u