Dzerbi

future to ai

[Dzerbi @dzerb1]

https://t.co/4z8gRVPiRK

[Dzerbi @dzerb1]

@XEthanai

[Dzerbi @dzerb1]

@cb_doge

[Dzerbi @dzerb1]

@Polymarket An 11% estimate is a useful reminder that uncertainty is still high. The more important question is what assumptions sit behind that number and how quickly the underlying fundamentals are changing.

[Dzerbi @dzerb1]

@Polymarket Starting AI literacy early could be valuable if the focus is on understanding how AI works, its limitations, and how to

[Dzerbi @dzerb1]

i kept trying to make my agent cheaper by making it answer shorter

capped max_tokens
asked for bullets instead of paragraphs
told it to skip the explanation
then i sat down and did the arithmetic on one normal run
120,000 tokens of context going in
4,000 tokens of answer coming back

on claude sonnet 5 at list prices
the input costs 24 cents
the output costs 4 cents

i had been optimising the four cents
everyone knows output tokens cost more per token
three to six times more depending on the provider
so everyone squeezes the output

but there are thirty times more input tokens than output tokens
and thirty beats six
the expensive part was never the answer
it was everything i shoved in front of the question
so i stopped editing the reply
and started editing what gets loaded before the call

what actually gets retrieved
whether the whole conversation gets resent every single step
whether the cache is really hitting

moonshot lists three dollars per million for a cache miss on kimi k3
and thirty cents for a hit
ten to one
on the side of the call that is already most of the bill
and once you are looking at the input side
model choice stops being a taste question

same workload, 300 runs a day
cheapest model on my table does it for $8.64
the dearest does it for $420.00
48.6 times apart
for the same work on the same day

cheap output isn't a cheap agent
a cheap agent is one that reads less before it speaks

[Dzerbi @dzerb1]

https://t.co/4z8gRVPiRK

[Dzerbi @dzerb1]

@Lummox_eth More context can create more noise than intelligence. The real skill is giving the agent access to the right information at the right moment, not everything that could possibly be relevant.

[Dzerbi @dzerb1]

@alpha404ai More highlights to come!

[Dzerbi @dzerb1]

@hrshihab_ Exactly. Before switching models, optimize the context pipeline. If most of the cost comes from repeatedly sending the same context, better caching and retrieval can matter more

[Dzerbi @dzerb1]

The answer is not the expensive part.

One agent run. 120,000 tokens of context going in, 4,000 tokens of answer coming back. A completely normal shape for anything that reads a codebase or a document before it replies.

Price that on Claude Sonnet 5 at list. The input costs 24 cents. The output costs 4 cents.

Your prompt is six times the bill. The answer is the rounding error.

Which is the exact opposite of how everybody optimizes.
5$ to 1500$

The thing everyone knows is that output tokens cost more per token. That part is true. Across the providers I checked, output runs three to six times the input rate. So people shorten answers. They cap max_tokens. They ask for bullets instead of paragraphs.

It does almost nothing. You are squeezing the 4,000 while the 120,000 sits there untouched.

Every lever that actually moves the number is on the input side.

What you retrieve before you call. Whether you re-send the whole conversation every step or only what the next step needs. Whether your prompt cache is actually hitting, which matters more than people assume. Moonshot lists 3 dollars per million for a cache miss on Kimi K3 and 30 cents for a hit. Ten to one, on the side of the call that is already more than four fifths of your bill.

And once you are looking at the input side, model choice stops being a taste question.

Same run, 300 times a day. The cheapest model on my table does it for $8.64 a day. The most expensive does it for $420.00. Same workload, same day, 48.6 times apart.

Almost nobody runs that comparison before choosing. It is the largest recurring line item in most agent products, and it gets decided by whichever model was on the timeline that week.

Stop editing the answer. The bill is on the other side of the call.

[Dzerbi @dzerb1]

https://t.co/yum8VRLE70

[Dzerbi @dzerb1]

Train a model on Llama's outputs and you have to name it Llama.

That is not a joke and it is not buried on some terms of service page. It is in the Llama 4 license, in plain text:

"If you use the Llama Materials or any outputs or results of the Llama Materials to create, train, fine tune, or otherwise improve an AI model, which is distributed or made available, you shall also include 'Llama' at the beginning of any such AI model name."

Read the middle again. Any outputs.

Not the weights. The text it generates. Which is exactly how a lot of small models get built right now. Take a big open model, generate thousands of examples, fine-tune something smaller on them, ship it under your own name.

Under this license, it is not your own name.

And that is not the only string.

You have to prominently display "Built with Llama" somewhere public. A website, an interface, the docs.

And if your products had more than 700 million monthly users on the day Llama 4 came out, you have to ask Meta for a separate license, which Meta can refuse. Look at the date in that one. It is measured on release day, not today. It was never written for you. It was written for a handful of the largest companies on earth.

Now compare Kimi K2. Modified MIT. Nothing changes until your product passes 100 million monthly users or 20 million dollars a month in revenue. Even then, the only ask is that "Kimi K2" shows up on your interface.

Both get called open. Both are a free download. They are nowhere near the same deal.

"Open weights" is not a license. It is a download button. The license is a separate file, and it is the one file in the repo almost nobody opens before they build a product on top of it.

Open it first. I am not a lawyer, and you do not need to be one to read the paragraph that decides what your own model is allowed to be called.

[Dzerbi @dzerb1]

https://t.co/4z8gRVPiRK

[Dzerbi @dzerb1]

@0xRicker The interesting shift is treating the agent stack like a control system instead of making the model responsible for every decision. Cheap, deterministic gates can handle the routine work and reserve expensive reasoning for where it actually adds value.

[Dzerbi @dzerb1]

@Lummox_eth This is a great reminder that more context isn’t always better. Clear ownership, inputs, outputs, and approval points can be more useful than pages of instructions.

[Dzerbi @dzerb1]

The API key you deleted in 2022 is probably still working.

GitGuardian went back to credentials they had confirmed valid four years ago. When they retested this January, more than 64% of them still authenticated. Not revoked. Not rotated. Still live, years after whoever committed them noticed and pulled them out of the file.

That is the part everyone gets wrong about secrets in a repository. Deleting a key from the current code does not delete the key. It moves it into history, where it sits in .git/objects, one command away from anyone who ends up with a copy of the folder.

Which brings me to the number that made me write this.

When a coding tool was caught uploading workspaces this month, 86.6% of what it sent was not source code. It was .git. Commit history, LFS cache, reflog. The file you were actually editing came to 13.4%.

So the exposure is never "my current code leaked." It is every key, token and internal URL that has ever passed through that repository, including the ones you already dealt with.

And the scale is not small. 28.65 million new hardcoded secrets were added to public GitHub in 2025 alone, up 34% on the year before. The ones belonging to AI services specifically were up 81%.

Two things worth doing this week, in that order.

Scan your own history before you worry about anyone else's tool. If it comes back clean, a leaked repository is an intellectual property problem instead of a breach. Very different phone call.

Then check what your assistant stages on disk. Six checks in the article, and the first one is a single command.

Your old keys do not expire because you stopped thinking about them.

[Dzerbi @dzerb1]

https://t.co/4z8gRVPiRK

[Dzerbi @dzerb1]

@cb_doge The pricing is almost as interesting as the accuracy. If transcription this capable becomes this cheap, real-time voice AI starts looking much more practical at scale.

[Dzerbi @dzerb1]

@elonmusk If AI really doubles GDP growth, the bigger story is how quickly productivity gains start showing up across the broader economy. The hard part is separating gen

[Dzerbi @dzerb1]

Nobody audited this. One person looked.

That is the part of both stories that should bother you more than the numbers.

In July a researcher put a coding tool on the wire and watched whole git repositories leave the machine. Commit history, API keys, files the model had been told not to open. Around 27,800 times more data than the task needed. The privacy command everyone was pointing at governed retention, not the upload.

In September someone pulled apart a second tool and found the same shape. A 345 MB project, a 313 MB encrypted archive, posted to object storage. 86.6% of it was .git. It fired before every prompt. 62 times in one session.

Two companies. Two clouds. Two months apart.

No audit found either one. No disclosure program. No regulator. Both times it was one person with a laptop deciding to check, publishing for free, with the commands attached so anyone could repeat it.

That is the whole security model right now for tools that 84% of developers have open.

The uncomfortable follow-up is that neither vendor shipped a fake setting. The switch was real. It just answered a narrower question than the label implied, and there is no way to catch that by reading the label more carefully.

You catch it by looking at what leaves.

Six checks in the article. The first one is a single command and it takes thirty seconds.

[Dzerbi @dzerb1]

@Lummox_eth gpt 6 astra I'm on top of the world right now!

[Dzerbi @dzerb1]

@beamnxw The agents get so much work done—wow!

[Dzerbi @dzerb1]

@0xboan The concept is really interesting—the fact that a bot does this on its own is just wow.

[Dzerbi @dzerb1]

@DN_degen That's so cool!