trenchi

Same world. A different angle.
lectures · finance · AI

[trenchi @trenchixbt]

@Beaver_0x Niklas Luhmann ran this exact system on 90,000 paper index cards and wrote 70 books and nearly 400 articles from it. Karpathy's version just hands the linking to Claude. The compounding was always the point.

[trenchi @trenchixbt]

@0xVarian Not stinginess, just math. A routing fork costs fractions of a cent, a frontier call costs real money at scale. Teams that skip this burn their margin on yes/no questions.

[trenchi @trenchixbt]

Your AI agent is burning money on questions a $0.000042 decision can answer.

Not on the writing. Not on the research. On the tiny forks in between: who works next, is this ready, does a human need to see it.

Most stacks send every fork to a frontier LLM. You pay for a paragraph of thinking to get back one word.

Jev 1.13 by TypeSafe cuts that loop:

> in: the goal, the work done, the workers free right now
> ask: Choice for a route, Score for a rating, Noul for yes or no
> out: a typed answer with a probability, 0 output tokens
> gate: anything under your threshold stops at review

The receipts:

> $0.042 per 1M input tokens, output is free
> Browser Use found flights in 7 seconds for $0.0039
> median browser calls per run fell from 1,092 to 101
> 1,018 research papers classified for $0.08
> LangChain uses it to pick cheap vs heavy models per task

Now look at the chip in the video.

0.94 on research clears the 0.85 bar and ships. 0.72 falls short and goes red, straight to review, before a single dollar runs down the wrong branch.

6,414 decisions. $0.27. Around 41ms each.

The bill was never the thinking. It was paying a genius to say yes or no.

LLM creates the work. Jev decides what happens next. Code executes.

So how many of your agent's expensive calls were ever thinking at all?

[codila @0xCodila]

Jev is the "Internet" moment for the AI industry

It tells your agents and LLMs what to do next, in milliseconds and at almost zero cost

If you set it up correctly, you will have the AI engineer’s stack for 2028

In this article, I show you how https://t.co/x3qn41ejnr

[trenchi @trenchixbt]

@AIDuckLabs Remotion plus Whisper is the right core because both are deterministic: same props, same frame, every render. The other 22 repos mostly add new ways to spend tokens before anything ships.

[trenchi @trenchixbt]

@DarkFactorr Clean-room isn't new. Phoenix cloned the IBM BIOS this way in 1984, one team wrote the spec, another never saw the code. The wall between them is the whole defense. One agent doing both sides is not a clean room.

[trenchi @trenchixbt]

@DarkFactorr Jev makes agent decisions cheaper and faster.

[trenchi @trenchixbt]

@andreysuperior Mozilla reviewed 25 car brands in 2023 and every single one failed its privacy test. A car is the only device people resell without a factory reset. Old owner access is the default, not a bug.

[trenchi @trenchixbt]

@0xMovez feed Claude three sites, not one. One site gives you a clone. Three give you a style nobody can trace back. That's the difference between copying a game and having a signature.