Do as much as necessary and as little as possible

• Nvidia's job: turn electrons into tokens. Do as much as necessary and as little as possible to make that happen.
• Whatever Nvidia doesn't need to do itself, it partners out and folds into the ecosystem.
• Result: the largest partner ecosystem in the industry, upstream and downstream.
• The test is simple: if we don't do it, does it get done?
• NVLink, the full stack, 20 years of CUDA at a loss, CUDA-X libraries, cuLitho for computational lithography — none of this existed without Nvidia building it.
• If nobody else will build it, Nvidia commits fully. Whole company, all the might.
• Clouds already exist. If Nvidia didn't run one, somebody would.
• So Nvidia doesn't become a cloud. It doesn't become a financier either. There are people in the financing business already.
• The business model stays simple on purpose.

AI is a five-layer cake

• AI isn't just a model. It's a five-layer stack.
• The bottom layer is energy. Above that sits chips, then systems, then models, then applications.
• Abundant energy makes up for weaker chips. Abundant chips make up for weak energy.
• The US is scarce on energy, so Nvidia has to push architecture and extreme co-design to get max throughput per watt.
• If watts are free, you don't care about perf per watt. You can run old chips. 7nm chips are essentially Hopper.
• This is why China having cheap abundant energy matters so much for their AI position.

Every layer has to win, not just one

• Applications matter most because that's where AI diffuses into society and drives the industrial revolution.
• But every layer has to succeed. You can't sacrifice one to protect another.
• Conceding the chip layer in China to protect the model layer is picking a fight with your own stack.
• Long-term tech leadership means winning all five layers, not trading them off.

Conceding China is a loser's premise

• Nvidia's share is growing, not shrinking. The assumption that they'd lose China anyway is wrong on the facts.
• Jensen didn't wake up a loser. The US isn't a loser. The industry isn't a loser.
• Walking away from the second largest market in the world because you assume you'd lose it makes no sense.

The policy argument relies on extremes

• The export-control case depends on absolutes: any compute at all means we lose everything.
• Real life isn't absolute. You can keep the best tech at home and first, and still compete abroad.
• Both things can be true at the same time. It takes nuance and maturity, not all-or-nothing thinking.
• Conceding China hands away ~40% of the world's tech industry for the benefit of one layer of the stack.
• That's a disservice to the country, to national security, and to American tech leadership.
• The goal should be: best tech here first, and compete to win everywhere else.

"This is one of the concerns that I have about the doomers describing the end of work and killing of jobs. If we discourage people from being software engineers, we're going to run out of software engineers. The same prediction happened ten years ago. Some of the doomers were telling people, 'Whatever you do, don't be a radiologist.' You might hear some of those videos still on the web saying radiology is going to be the first career to go and the world is not going to need any more radiologists. Guess what we're short of? Radiologists."

Why jobs and tasks are not the same thing

• A job is the full role. A task is one thing inside it.
• A radiologist's job is patient care. Reading a scan is just one task.
• AI automating the task doesn't eliminate the job. It frees the person for the rest of the role.
• Confusing the two is how you end up with a shortage of radiologists and worse healthcare.

• Moore's Law now delivers about 25% per year. That's it.
• Between Hopper and Blackwell, the transistors themselves got roughly 75% better over three years.
• But Blackwell is 50x Hopper. Lithography can't explain that gap.
• Jensen originally announced Blackwell as 35x more energy efficient than Hopper. Nobody believed it. Dylan later wrote that Jensen sandbagged and it's actually 50x.

Where the 10x–100x leaps actually come from

• The lever is computer science, not transistors.
• Great algorithm work adds another ~10x on top of Moore's Law each year.
• MoE is the clearest example. New attention mechanisms also cut compute dramatically.
• Most of AI's recent progress came from algorithm advances, not raw hardware.

Why Nvidia can capture those algorithm gains and TPUs can't

• CUDA is programmable enough to invent new architectures on, whether that's MoE, diffusion, disaggregated inference, or hybrid SSMs.
• Nvidia is an extreme co-design company. They can change the processor, system, fabric, libraries, and algorithm at the same time.
• They push computation off the chip into the fabric via NVLink, or into the network via Spectrum-X.
• Without CUDA underneath, Jensen says he wouldn't even know where to start doing this.

Why process node isn't the whole story

• There's no 10x gap between 5nm and 7nm. Lithography is a small part of the picture.
• Architecture matters. Networking matters, which is why Nvidia bought Mellanox. Energy matters.
• Semiconductor physics still counts, but computer science and the full stack above the transistor drive most of AI's impact.

When designing a consumer product, you should consider every tap by a user to be a miracle. The motivation to stop using a new app will always be stronger than to use it To signup & complete a profile on Gas, we got it down to 15 taps total—with no keyboard required at any stage

They also don’t realize that yea I could keep going through thier onboarding flow, but timewise it’s competing with me looking at more dancing/cat videos on Tik Tok, and that is some seriously compelling content to compete with!

There’s some weird cognitive bias where product creators vastly overestimate how likely people are to do or pay attention to something based on how much the product creator wants them to, and they vastly underestimate how lazy people are.

To put it simply: you can look at a screen and intuitively predict the percent of users who will convert to the next screen within a 10% margin of error.

A dumb PM will stare at an onboarding funnel for hours, fabricate problems, create months of work for engineers, and ultimately improve conversion by 5%. A smart PM will look at funnel in a minute, make a button bigger, and improve conversion by 20%.

Counter positioning is an avenue for defeating an incumbent who appears unassailable by conventional wisdom metrics of competitive strength

What does the word "should" have to do with it? It’s not a matter of permitting something or forbidding something. Let him suffer, if he’s sorry for his victim . . .

Suffering and pain are always mandatory for broad minds and deep hearts. Truly great people, it seems to me, should feel great sadness on this earth.

.. My thought for today is something which I found in Epicurus (yes, I actually make a practice of going over to the enemy’s camp – by way of reconnaissance, not as a deserter!), ...

the tendency of Stoicism was always to exalt man’s importance in the universe rather than to abase him before a higher authority

"it is no part of the business of philosophy to turn people into better persons. His tremendous faith in philosophy as a mistress was grounded on a belief that her end was the practical one of curing souls, of bringing peace and order to the feverish minds of men pursuing the wrong aims in life"

the structure of the digital economy means most of our digital lives are designed to take advantage of this state. A substantial fraction of the world’s most brilliant, competent, and empathetic people, armed with near-unlimited capital and increasingly god-like computers, spend their lives serving Marl.

"""
Here’s what I’ve been able to piece together about the marginal user. Let’s call him Marl. The first thing you need to know about Marl is that he has the attention span of a goldfish on acid. Once Marl opens your app, you have about 1.3 seconds to catch his attention with a shiny image or triggering headline, otherwise he’ll swipe back to TikTok and never open your app again.

Marl’s tolerance for user interface complexity is zero. As far as you can tell he only has one working thumb, and the only thing that thumb can do is flick upwards in a repetitive, zombielike scrolling motion. As a product designer concerned about the wellbeing of your users, you might wonder - does Marl really want to be hate-reading articles for 6 hours every night? Is Marl okay? You might think to add a setting where Marl can enter his preferences about the content he sees: less politics, more sports, simple stuff like that. But Marl will never click through any of your hamburger menus, never change any setting to a non-default. You might think Marl just doesn’t know about the settings. You might think to make things more convenient for Marl, perhaps add a little “see less like this” button below a piece of content. Oh boy, are you ever wrong. This absolutely infuriates Marl. On the margin, the handful of pixels occupied by your well-intentioned little button replaced pixels that contained a triggering headline or a cute image of a puppy. Insufficiently stimulated, Marl throws a fit and swipes over to TikTok, never to return to your app. Your feature decreases DAUs in the A/B test. In the launch committee meeting, you mumble something about “user agency” as your VP looks at you with pity and scorn. Your button doesn’t get deployed. You don’t get your promotion. Your wife leaves you. Probably for Marl.
"""

"""
here is a well-established taboo against anthropomorphizing AI systems. This caution is often warranted: attributing human emotions to language models can lead to misplaced trust or over-attachment. But our findings suggest that there may also be risks from failing to apply some degree of anthropomorphic reasoning to models. As discussed above, when users interact with AI models, they are typically interacting with a character (Claude in our case) being played by the model, whose characteristics are derived from human archetypes. From this perspective, it is natural for models to have developed internal machinery to emulate human-like psychological characteristics, and for the character they play to make use of this machinery. To understand these models’ behavior, anthropomorphic reasoning is essential.

This doesn’t mean we should naively take a model’s verbal emotional expressions at face value, or draw any conclusions about the possibility of it having subjective experience. But it does mean that reasoning about models’ internal representations using the vocabulary of human psychology can be genuinely informative, and that not doing so comes with real costs. If we describe the model as acting “desperate,” we’re pointing at a specific, measurable pattern of neural activity with demonstrable, consequential behavioral effects. If we don’t apply some degree of anthropomorphic reasoning, we’re likely to miss, or fail to understand, important model behaviors. Anthropomorphic reasoning can also provide a useful baseline of comparison for understanding the ways in which models are not human-like, which has important consequences for AI alignment and safety.
"""

Post-training of Claude Sonnet 4.5 in particular led to increased activations of emotions like “broody,” “gloomy,” and “reflective,” and decreased activations of high-intensity emotions like “enthusiastic” or “exasperated.”

sounds about right, lol

"these representations can play a causal role in shaping model behavior—analogous in some ways to the role emotions play in human behavior—with impacts on task performance and decision-making"