Casper

Spectral operator. Building AI tools in the shadows. Shipping in public ▮

[Casper @caspr_exe]

NVIDIA made $215.9 billion last year without shipping a single AI model.

It sells the compute every lab runs on, and $193.7 billion of that came from data centers alone, roughly 90% of the entire company.

Blackwell is sold out. $500 billion in orders already booked through the end of 2026. Market cap around $4.9 trillion, the most valuable company on earth.

Models come and go. The moat is the silicon they all rent. Own the compute, and you own the floor everyone else builds on.

[Casper @caspr_exe]

@Davidgek97Q right. you don't have to win the race when you own the track everyone runs on.

[Casper @caspr_exe]

NVIDIA made $215.9 billion last year without shipping a single AI model.

It sells the compute every lab runs on, and $193.7 billion of that came from data centers alone, roughly 90% of the entire company.

Blackwell is sold out. $500 billion in orders already booked through the end of 2026. Market cap around $4.9 trillion, the most valuable company on earth.

Models come and go. The moat is the silicon they all rent. Own the compute, and you own the floor everyone else builds on.

[Phosphen @phosphenq]

https://t.co/YA9AvlYHOA

[Casper @caspr_exe]

@thedankoe amazing words ! saved

[Casper @caspr_exe]

Rich Sutton - professor, AI scientist, Turing Award 2024 winner - explains why "loops" beat prompting, years before the term existed:

0:12 - why prediction is AI's real bottleneck
4:48 - the Bitter Lesson before it had a name
13:39 - what a loop is: guess from a guess, not the final answer
30:51 - why waiting for the end loses to updating every step
53:26 - proof that looping beats matching your training data

Lecture that explains the theory behind why Karpathy's 630-line loop beat two decades of hand-tuning.
Turns out the "new" trick had a name in academia before it had one on Twitter.

[Casper @caspr_exe]

Rich Sutton - professor, AI scientist, Turing Award 2024 winner - explains why "loops" beat prompting, years before the term existed:

0:12 - why prediction is AI's real bottleneck
4:48 - the Bitter Lesson before it had a name
13:39 - what a loop is: guess from a guess, not the final answer
30:51 - why waiting for the end loses to updating every step
53:26 - proof that looping beats matching your training data

Lecture that explains the theory behind why Karpathy's 630-line loop beat two decades of hand-tuning.
Turns out the "new" trick had a name in academia before it had one on Twitter.

[codila @0xCodila]

https://t.co/UJL7o2l9oR

[Casper @caspr_exe]

MIT charges $80,000 a year to teach you this. Your terminal just did it for free.

There's a folder on your machine that knows you better than some of your friends: ~/.claude/projects

Every request you typed at 2am. Every project you started and ghosted. No performance — because nobody performs for a coding agent.

An MIT professor explains why: utility isn't something you can state. It's ordinal, not cardinal — you can't say how much you want something, only what you're willing to give up for it.

That's not therapy. That's economics. And it's exactly what six months of your session history can do to you if you actually read it.

[Casper @caspr_exe]

MIT charges $80,000 a year to teach you this. Your terminal just did it for free.

There's a folder on your machine that knows you better than some of your friends: ~/.claude/projects

Every request you typed at 2am. Every project you started and ghosted. No performance — because nobody performs for a coding agent.

An MIT professor explains why: utility isn't something you can state. It's ordinal, not cardinal — you can't say how much you want something, only what you're willing to give up for it.

That's not therapy. That's economics. And it's exactly what six months of your session history can do to you if you actually read it.

[Machina @EXM7777]

https://t.co/au4TgoIh59

[Casper @caspr_exe]

AI companies are spending 300 billions dollars on infrastructure, models, and talent. But the most valuable asset isn't the GPUs or the models.

It's knowledge.

Every experiment. Every failed idea. Every prompt. Every research note. Every decision. Connected into a system that compounds over time instead of getting lost in folders.

That's what you're looking at in this video.

A living knowledge graph where thousands of notes, frameworks, and mental models are linked together so AI can reason over them instead of starting from zero every time.

The companies winning the AI race aren't just building better models.

They're building better memory.

That's the real moat.

[Casper @caspr_exe]

AI companies are spending 300 billions dollars on infrastructure, models, and talent. But the most valuable asset isn't the GPUs or the models.

It's knowledge.

Every experiment. Every failed idea. Every prompt. Every research note. Every decision. Connected into a system that compounds over time instead of getting lost in folders.

That's what you're looking at in this video.

A living knowledge graph where thousands of notes, frameworks, and mental models are linked together so AI can reason over them instead of starting from zero every time.

The companies winning the AI race aren't just building better models.

They're building better memory.

That's the real moat.

[rari @0xwhrrari]

https://t.co/uX3jpfiEGq

[Casper @caspr_exe]

Percy Liang, Stanford professor, on how frontier LLMs are actually built:

• 3:23 - why researchers lost touch with how models work
• 4:54 - the $1B problem: why frontier models are out of reach
• 11:42 - the full history, Shannon to GPT
• 28:06 - tokenization: the atoms a model sees
• 35:58 - systems: squeezing everything out of the GPU
• 45:12 - scaling laws: predicting a model before you spend $100M
• 53:38 - data: what actually decides how good your model gets
• 1:00:18 - alignment: RLHF and DPO

One Stanford course that lays out the entire pipeline behind every frontier model.

"GPT-4 was supposedly costing $100 million to train, and now costs probably are on the order of $1 billion." - Percy Liang

[Casper @caspr_exe]

Percy Liang, Stanford professor, on how frontier LLMs are actually built:

• 3:23 - why researchers lost touch with how models work
• 4:54 - the $1B problem: why frontier models are out of reach
• 11:42 - the full history, Shannon to GPT
• 28:06 - tokenization: the atoms a model sees
• 35:58 - systems: squeezing everything out of the GPU
• 45:12 - scaling laws: predicting a model before you spend $100M
• 53:38 - data: what actually decides how good your model gets
• 1:00:18 - alignment: RLHF and DPO

One Stanford course that lays out the entire pipeline behind every frontier model.

"GPT-4 was supposedly costing $100 million to train, and now costs probably are on the order of $1 billion." - Percy Liang

[Khairallah AL-Awady @eng_khairallah1]

https://t.co/EQi4JlGoLQ

[Casper @caspr_exe]

Jesse Mu, one of Stanford’s leading researchers:

"Chatbots are rewarded to produce responses that seem authoritative or seem helpful, but they don't care about whether it's actually true or not."

That is not a glitch. That is the final training stage doing its job.

GPT went from 117M parameters in 2018 to 175B in 2020. A 1000x jump in two years, and that scale is where the fluency comes from.

But fluency is not truth. To make it feel helpful, OpenAI paid people to rank its answers, reportedly 40 hours a week. It learned to sound right, not to be right.

The magic was never magic. The confident hallucination is stage five, RLHF, working as designed.

[Casper @caspr_exe]

Jesse Mu, one of Stanford’s leading researchers:

"Chatbots are rewarded to produce responses that seem authoritative or seem helpful, but they don't care about whether it's actually true or not."

That is not a glitch. That is the final training stage doing its job.

GPT went from 117M parameters in 2018 to 175B in 2020. A 1000x jump in two years, and that scale is where the fluency comes from.

But fluency is not truth. To make it feel helpful, OpenAI paid people to rank its answers, reportedly 40 hours a week. It learned to sound right, not to be right.

The magic was never magic. The confident hallucination is stage five, RLHF, working as designed.

[Khairallah AL-Awady @eng_khairallah1]

https://t.co/EQi4JlGoLQ

[Casper @caspr_exe]

@eng_khairallah1 amazing arcticle!!1

[Casper @caspr_exe]

NVIDIA, at Nemotron Days: "the LLM is the brain, the harness is the body."

Developers shipped 1.4 billion commits in six months, a 3x jump, turning ~$3T of dev work into ~$6T of output.

That jump isn't a smarter model. it's the harness around it:

The loop that lets the agent use tools, verify its own work, and stop only when the goal is met.

The reason you can finally hand off a whole project is the same one: a machine-checkable "done." the model was never the moat, the harness is.

[Casper @caspr_exe]

@insomnia_vip good article.
saved

[Casper @caspr_exe]

NVIDIA, at Nemotron Days: "the LLM is the brain, the harness is the body."

Developers shipped 1.4 billion commits in six months, a 3x jump, turning ~$3T of dev work into ~$6T of output.

That jump isn't a smarter model. it's the harness around it:

The loop that lets the agent use tools, verify its own work, and stop only when the goal is met.

The reason you can finally hand off a whole project is the same one: a machine-checkable "done." the model was never the moat, the harness is.

[Insomnia @insomnia_vip]

https://t.co/HDsFXMRFbn

[Casper @caspr_exe]

Jim Simons, founder of Renaissance Technologies: "We never override the computer."

He earned over $125 million a month doing exactly that:

Medallion did ~66% a year before fees for 30 years with zero losing years. And the math behind it is public: Bernoulli 1713, Kelly 1956, Black-Scholes 1973.

The moat was never the equation. it's data, speed, and the discipline to trust it across thousands of trades.

Same law as AI: the model was never the moat, the nerve to trust the system is.

[Casper @caspr_exe]

Jim Simons, founder of Renaissance Technologies: "We never override the computer."

He earned over $125 million a month doing exactly that:

Medallion did ~66% a year before fees for 30 years with zero losing years. And the math behind it is public: Bernoulli 1713, Kelly 1956, Black-Scholes 1973.

The moat was never the equation. it's data, speed, and the discipline to trust it across thousands of trades.

Same law as AI: the model was never the moat, the nerve to trust the system is.

[Rossst.03 @Rossst_03]

https://t.co/FgJC5WmYvy