Lima

Content creator | Building AI agents | 1% better today

[Lima @limalemonnn]

https://t.co/YxQbOs7iZg

[Lima @limalemonnn]

@0xOrionVega One should listen more carefully to such people.

[Lima @limalemonnn]

@crimson_moti The wet floor call is where the post earns credibility

[Lima @limalemonnn]

@0xKorens Thanks for sharing such profiles it's very valuable

[Lima @limalemonnn]

@ridark_eth that's incredible result

[Lima @limalemonnn]

@0xGenAi The Coala framework sounds interesting😁

[Lima @limalemonnn]

AI labs are buying old libraries and scanning ancient books. Not for culture. For tokens.

The internet ran out. Ex-OpenAI researcher, on a Stanford stage: pre-training has hit the frontier of available data.

Same talk, second number. DeepSeek V3 pre-training: about 2.4 million H800-hours. The RL that turned it into R1: about 150 thousand.

Five percent of the compute bought the entire reasoning layer.

So the expensive half of AI hit a ceiling, and the cheap half still compounds. Labs already act on it: RL runs now span multiple datacenters, and evals get guarded like source code.

The token race is ending. The environment race is starting. Keep the five percent in mind next time a release gets called a breakthrough.

Bookmark this. Follow @limalemonnn

[Lima @limalemonnn]

@misat0x People love magic pills

[Lima @limalemonnn]

@0xVibeDino If you're talking about DeepSeek, then I guess so.

[Lima @limalemonnn]

AI labs are buying old libraries and scanning ancient books. Not for culture. For tokens.

The internet ran out. Ex-OpenAI researcher, on a Stanford stage: pre-training has hit the frontier of available data.

Same talk, second number. DeepSeek V3 pre-training: about 2.4 million H800-hours. The RL that turned it into R1: about 150 thousand.

Five percent of the compute bought the entire reasoning layer.

So the expensive half of AI hit a ceiling, and the cheap half still compounds. Labs already act on it: RL runs now span multiple datacenters, and evals get guarded like source code.

The token race is ending. The environment race is starting. Keep the five percent in mind next time a release gets called a breakthrough.

Bookmark this. Follow @limalemonnn

[Lima @limalemonnn]

https://t.co/YxQbOs7iZg

[Lima @limalemonnn]

@nguoideptegieng Libraries and books are among the best sources. There is still hope that the internet won't die out after all.

[Lima @limalemonnn]

@ZentrixHQ that's crazy !

[Lima @limalemonnn]

@velesxbt Samuelson wasn't a hypocrite. He was calibrated. The theory holds for 99%, he identified the 1%, kept quiet because saying it publicly would break the frame. Rare mix of academic honesty and private discipline.

[Lima @limalemonnn]

@crimson_moti hands on the table!

[Lima @limalemonnn]

@0xKorens thanks🫡

[Lima @limalemonnn]

@GuntherWrite also compresses information. no capitalization overhead, no formal punctuation ritual. reader spends attention on the idea, not on the shape of the sentence.

[Lima @limalemonnn]

My Claude bill funds Nvidia more than it funds Anthropic. Stanford's new AI-economics course showed why in one chart: of the last $350B in revenue this industry added, about 75% landed at the chip layer.

The mechanism, straight from the lecture: the incremental AI user isn't free, you burn GPUs. Software ran 80-90% margins because the next user cost nothing; AI apps run 0-30. That's why companies with billions in revenue still lose money, and why nobody above the chips can subsidize your wasted tokens for long.

You don't negotiate with that triangle. The one lever left on your side of the bill is sending fewer tokens: expensive model for decisions, cheap models for volume, and the effort dial checked before every run. My full routing setup is in the guide below.

One more course number worth keeping: ChatGPT earns about $10 per user per year, Google earns $100, and the instructor's call is that ads close the gap. So the question that decides the next two years: would you accept ads in your chat window for a free frontier model?

Bookmark this. Follow @limalemonnn

[Lima @limalemonnn]

@Neuron_404 The workflow works, but 15 seconds isn't what brands pay studios for. They pay for revisions, brand guidelines, IP checks, and sound. The clip is 20% of the deliverable.

[Lima @limalemonnn]

@carbonyxxx keep grinding

[Lima @limalemonnn]

A CEO got tired of one subscription. He live-coded the replacement in v0. Kept the parking lot, dropped the vendor.

Rauch told it at Stanford as a routine story. His own people rebuilt Salesforce on top of Salesforce. His words: software is basically free now.

The split is clean. Survives: the database, access control, any SaaS with an open API. Dies: the interface you rent. Tobi from Shopify named the trap: one prompt away means you never pay for generic again.

The guide above cuts your agent's bill. Every dollar off that bill lowers the bar for replacing the next subscription. First SaaS on your list?

Bookmark this. Follow @limalemonnn

[codila @0xCodila]

https://t.co/CAbiVW92hs

[Lima @limalemonnn]

@carbonyxxx We are living through incredible times

[Lima @limalemonnn]

@Pixel_Neuron Why use SaaS if there are cheap tokens?😅