Wrong decision models can lead us astray. Some of the most powerful decision models ( inspired by Munger) helped me become incredibly successful
Follow for more : 1/many
The Lollapalooza Effect: When multiple causes act in a compounding, multiplicative fashion, the resulting effects can be extremely powerful and outsized. Look for intersections where forces combine.
The shocking failure of companies like Enron and Lehman Brothers, FTX, ICOs, among others arose from the confluence of factors like flawed incentives, hubris, leverage, and lack of oversight compounding disastrously.
I don't think people realize what's happening with Elon Musk right now.
Look at what he posted today. The CEO of the most valuable car company in the world explains, on a Sunday, on X, that his team had to fit the decoder of their world model into the chip's SRAM, that they only had half an accelerator to work with, that this accelerator only handles 8-bit integers, and that softmax has to be emulated in 28 steps.
Name one other CEO at that level who could write that post. Not have it written. Write it.
That's not a detail. I think the way Musk operates at scale, combined with superintelligence, is going to give him an absolutely insane edge.
Let's go back to first principles.
A company is a machine that turns intent into reality. Between the two, there are two costs. The intelligence needed to find the solution. And the synchronization needed to get thousands of people executing it in the same direction.
Superintelligence is wiping out the first one. When thousands of agents smarter than your best engineers cost the price of electricity, intelligence is no longer scarce.
That leaves the second one. And it's exploding.
The bottleneck today is synchronization. Humans synchronizing agents. And humans synchronizing the humans who synchronize agents.
Every layer is lossy compression. The CEO tells the VP, who rephrases it for the director, who rephrases it for the manager, who prompts the agent. By the end, the agent flawlessly executes an intent nobody recognizes anymore.
Agents don't fix this. They amplify it. They move so fast that one bad intent turns into a thousand bad deliverables before the next leadership meeting.
Now listen to Marc Andreessen. He broke down what he calls "the Elon method." Every week, at every company, Musk finds the single biggest bottleneck and fixes it himself. He doesn't ask the VP to ask the director to ask the manager. He goes straight to the engineer on the problem and solves it with them. The rest of the time, he runs engineering reviews where every engineer presents what they're building.
Ben Horowitz, Andreessen's partner, sums up the condition that makes it almost impossible to copy: you need to be able to hold every engineering topic in the company in your head, all at the same time.
Here's what almost nobody sees: the Elon method is exactly the architecture of the agent era. The shortest possible sync chain between the person who understands and the thing that executes.
A CEO who doesn't understand what he's doing can only delegate his trust. A CEO who understands can delegate the work, and check the result. With humans, that gap cost months. With agents, it will cost orders of magnitude.
Second edge: hardware.
Superintelligence doesn't live in a vacuum. It needs chips, energy, data, and a body to act in the physical world. Musk has all of it. Millions of cars filming reality around the clock. His own chips. Compute through xAI. Energy storage. Robots with Optimus. Rockets and a satellite constellation.
And look at how it all connects. The same vision encoder powers FSD, Robotaxi and Optimus. One intelligence, three bodies. The next-generation chip, AI5, goes straight into Optimus, because according to him AI4 is already enough to solve driving.
Everyone else is optimizing a model. He's optimizing a full loop: data, training, silicon, energy, factory, robot, and back to data.
Put the two together. Intelligence is becoming a commodity. What stays scarce is atoms and sync speed. Musk is probably the man on earth who owns the most of both.
The labs have the models, not the factories. The industrials have the factories, not the models. The giants who have both have ten layers of management between the person who decides and the person who knows.
I look at the map and I don't see anyone who can keep up in this configuration. Honestly, I don't even see how he loses.
In the world that's coming, intelligence will be free.
Understanding what you're doing will never be for sale.
I just saw the paper as a result of @martin_casado's tweet (didn't see the paper when it was originally posted in Feb). Indeed, the analysis is great and the figures really make the paper shine! Nice work, Dhruva Karkada, Daniel J. Korchinski, Andres Nava, Matthieu Wyart, and @yasamanbb! I love analyses that reveal even more about the amazing hidden structures in word embeddings.
https://t.co/uTtuXTRYd8
I seriously think US open-weight models wouldn't be this good without Chinese open-weight models and vice versa.
And proprietary models wouldn't be this good (and cheap) without open-weight models.
Competition is a good thing.
MoE is a datacenter architecture. Looped transformers are a laptop architecture.
MoE trades memory for compute: huge weights, but only a few active per token. That works with big batches and big GPUs. On your laptop the batch size is 1, often 0, and the memory isn't there.
Looped transformers make the opposite trade: reuse the same weights, spend more compute, need far less memory.
I won't be surprised if local and cloud models end up with very different architectures, including ones invented just for local inference.
Free idea for labs (probably you already do a version of this) : take Tatsu's https://t.co/OXr54K2EUk deep learning alchemy class and feed the quizzes and hw to your synth data generator then hill climb. This should at least get some "rsi" or good science of dl abilities for pretrain.
1) 2030 Fields Medal committee should award Levent, or they should disband.
2) Levent's prompting as presented in his timeline was reflective of serious centaur math, him making high-level judgments with immense details left to Claude.
3) The following gen prolly oneshots it.
Riemann just proved this about frontier models:
Long horizon verifiable problems with incredibly large search space is solvable.
I find the decision problems within enterprises still unsolved.
The Lean FRO and Mathlib Initiative have released a joint statement on the Lean verification of the Riemann Hypothesis.
On November 20, Levent Alpöge released a proof of the Riemann Hypothesis as a Lean library of 17 million lines of code. Over the past day, the Lean and Mathlib communities worked together to verify it. The statement certifies the proof as correct and describes how it meets all necessary verification criteria.
Congratulations to Levent on this achievement, and thank you to everyone in the community who contributed to the verification.
Full statement: https://t.co/edxEbkodw7
The Lean FRO and Mathlib Initiative have released a joint statement on the Lean verification of the Riemann Hypothesis.
On November 20, Levent Alpöge released a proof of the Riemann Hypothesis as a Lean library of 17 million lines of code. Over the past day, the Lean and Mathlib communities worked together to verify it. The statement certifies the proof as correct and describes how it meets all necessary verification criteria.
Congratulations to Levent on this achievement, and thank you to everyone in the community who contributed to the verification.
Full statement: https://t.co/edxEbkodw7
@pitdesi@AndyM1928@colemanoates Do we have to really mention this if the table stakes are 6 hours ??
You have every version of mountain, ocean, lake, surf, snow, sleet, desert, redwoods, and everything in between within 6 hours.
Levent Alpöge is a freak... look at his math breakthroughs since july:
disproved the 3d jacobian conjecture
discovered new hadamard matrices including order 668
proved 3d euler blow-up with buckmaster
disproved the carathéodory conjecture
set elliptic curve rank records (30 and 31) with ava howell
claimed to solve hopf's problem by proving s⁶ has a complex structure
presented a proof of unforced 3d navier-stokes blow-up with buckmaster
claimed a proof of the riemann hypothesis
Charlie Munger always said show me the incentive and I will show you the outcome, and right now every frontier AI lab is grading its own homework on public benchmarks..
When valuation & API revenue depend on topping the leaderboard, the training pipeline optimizes for the test
- Public benchmarks leak into pre-training or get hill-climbed by RL
- Models look like geniuses on static multiple choice & coding puzzles
- Same models still break on edge cases inside real enterprise workflows
I think the only evals worth trusting are the private graders you build on your own failure modes & your own messy production data..
My thesis: Intelligence = Function (Data, Evals, Models, Harness), and Data & Evals will be 80% of all AI systems..
Models commoditize --> Harness becomes standard plumbing --> Proprietary data + evals become the moat
The labs keep shipping better models to everyone, but your decision traces, failure cases & graders don't ship with them..
2nd Order:
Founders, spend where the 80% sits.. data pipelines & evals first
@elonmusk@dhh@wholemars@matthewstoller Does this deserve a response. It's like a non-swimmer commenting about Michael Phelps. Does it matter?
Matt's profile:
Writer for BIG: https://thebignewsletter.comCo-host, Organized Money: https://organizedmoney.fmResearcher at AELP
@nikunj Everything except (1) is a Decision Trace problem. Solving independently for 2-5 is why software in enterprises are so terrible. Build the Decision Engine and let it drive everything. Building in stealth now.
It’s not paradoxical, it’s a calibration & observational mindset anchored in truth.
We constantly learn from our surroundings and fixate on identifying, assessing, calibrating and synthesizing how reality works. We judge ourselves as critically as everyone else.
It’s the way of nature, if you remove the cacophony of the human mind.