Leading the Applied AI Lab, DGX Cloud at @NVIDIA, working on time series, agents, and foundation models.
Joined April 2008
- Tweets 29,302
- Following 3,285
- Followers 6,554
- Likes 25,318
When humans fail, we do a 5 whys approach, try to understand the process around the humans, do blameless retros, and go off into the sunset singing Kumbaya.
When an AI agent makes a mistake that could be fixed by a better system of guardrails around the agent, same org will get mad and unplug the robot, say AI fails, and write a LinkedIn cope-post about why humans are always better than AI.
Oh, so you screenshotted his message and thought you had a thesis? You turned two words — ‘I disagree’ — into a whole dissertation on Hegel, Coke Zero, and Lacan. That’s not insight, sweetheart, that’s overclocked grad-seminar cosplay. You’re not exposing anything, you’re dressing up a fumbly DM like it’s Federalist No. 10.
The guy was awkward, sure. But you? You’re out here building a bonfire of jargon just to roast a spark. You mistake verbosity for victory. He wanted conversation, you wanted content. And congratulations — you got it. Not philosophy, not dialectic — content. You’re not Madison in Philadelphia, you’re a town crier in Williamsburg yelling ‘I am begging the men of the world to be normal,’ ringing a bell for retweets.
And that’s the joke: he might’ve been clumsy, but you turned into a caricature. You think you’re Socrates, but what you really are is Cliff Notes in drag — all the references, none of the substance. You didn’t refute him. You inflated him. You handed him immortality in the only currency that matters today: screenshot circulation. And that’s not critique — that’s free advertising.
Ah! You see, the story is almost too perfect. The man encounters her in the museum, in front of Rothko. And what is Rothko? A black void, a red abyss, the silent scream of modernity. It is a demand for stillness, for confrontation with nothingness. And what does he do? Instead of confronting the void, he runs from it. He fills the silence with himself. He goes home, finds her blog, and writes: “we share interests, I read your post, I disagree.” Already, the act is obscene. The Rothko asks for silence, and he answers with a DM.
This is the Hegelian trick. On the surface, he performs philosophy. He frames his words as serious critique. But in truth, it is abstract negation. It is like Coke Zero, you know. Disagreement without the sugar of real engagement. He does not move the thought forward, he re-presents her own words back to her, only stamped with his authority: “I disagree.”
Now comes the reversal. She screenshots it. She posts it with the caption, “I am begging the men of the world to be normal.” Here, the dialectic achieves its completion. His attempt at recognition collapses into objecthood. He wanted to be interlocutor, he becomes exhibit. His seriousness is aufgehoben into comedy. He thought he was engaging in philosophy, but he is transformed into what Lacan would call the objet petit DM, the tiny kernel of humiliation now circulating in the meme economy.
And this is the paradox. The message was never private. Every DM already carries within it the possibility of its public unveiling. The truth of the DM is not in what it says, but in what it becomes. By posting it, she does not betray the interaction. She reveals it.
The Rothko was the warning. The void demanded contemplation. He could not endure it. He rushed to fill it with pseudo-philosophy. And so his words fell into the same abyss, only to return as content.
A concept I like when figuring out how to make your first agentic workflow is to find what I like to call the "dumb diamonds".
What is a "dumb diamond"? It is any process where you are making a decision, the decision isn't *that* hard to make, and the cost of being wrong is initially low.
Those are the good first candidates where an agentic workflow can take the place of an intern deciding amongst a set of fixed options which thing to do, but where the input is fuzzy.
This is often the quick win for an agentic AI system, and is a pretty low stakes of done using the right "dumb diamonds".
The hard stuff can come later. Walk, then run.
The reason it won't go away is because advertisers have a vested interest in it existing and will lobby, but for the vast majority of people, you could just not use the postal service and little would change.
Waymo is now both faster to get a ride from and less expensive than Uber in my SF neighborhood. Uber is now like Pepsi… what you drink when you can’t find Coke.
The middle managers that add less value than copilot’s summarizer from 18 months ago are really crashing out now.
This is virtuous. Real management isn’t an email job. It’s giving a shit about details. It’s troubleshooting group dynamics and actually fixing them. It’s understanding the rare context of your stakeholders and messaging with that in mind. Its use of judgement and taste to effectively allocate resources.
In tech, it has been stunningly easy to get by as ManagerGPT. This is why CEOs love AI, as it commoditizes safe hr-speak and normalizes the idea that if you offer nothing but safe status quo word salad, you’re gone.
(For the young ones who might not remember, OpenStack was/is an open source cloud stack you can build your own private cloud with)
One time, 13 years ago, when my Brazil dev team turned me into a meme after I found the worlds worst idea: livestreams in dance clubs
One thing I learned once upon a time is many corporate environments don't love when a VP openly says "I don't know"
Proof of AGI might be AI telling you your ideas are really not very good
THE AGENT LAYER NEGOTIATES, PLANS, AND DELEGATES (LIKE A WORLD-CLASS MANAGER). THE PLATFORM LAYER SUPPLIES CAPABILITIES: SEARCH, DB, PAYMENTS, OBSERVABILITY. PUT THEM TOGETHER AND YOU GET COMPOSABLE O‑O‑D‑A LOOPS THAT LEARN. INCREDIBLE POWER. NOT A TOY.
Seeing how people become “friends” with their LLM means the day that ads get injected it’s going to feel like when a real friend starts selling Amway.
You shut it off and stop returning its calls.
I just successfully booked a haircut near me tomorrow using ChatGPT 5 in agent mode. That was pretty cool.
In a hurry with no food in the house. Ordered @UberEats for the first time in 6 months.
No fewer than 7 beg buttons in the checkout path trying to get you to upsells and nonsense.
As soon as this orders completes I’m taking a shower to get the used car salesman smell off before I delete the app again.
The enshittification will continue til morale improves.
