Dzerbi @dzerb1

@dontbelievetoAI That really is an incredible number

Anthropic went from $9B to $65B in annualized revenue in seven months. Now it wants about $2 trillion on Nasdaq OpenAI doubled from $20B to $40B+ in the same stretch. At any other company that is the headline of the year. Here it is second place. But the two numbers are not counted the same way. Anthropic books cloud partner sales gross. OpenAI books them net. Switch Anthropic to net and it shrinks 6 to 10%. It still leads by about $20B. The price tags are the real story. Anthropic's last private round: about 15× its run-rate. OpenAI's rumored round at $1.2T: about 30×. Twice the multiple for the slower curve. Then the bill. OpenAI's own projections, reported by the FT: $856B on compute vs $840B in revenue, 2026 to 2030. What this means: the IPO will be the first time public markets price an AI lab on revenue, margins and compute. Not demos. What this does not mean: that either stock is a buy. Nothing here is advice. Full breakdown with charts in the first reply. Revenue gets the headline. Compute gets the bill.

Dzerbi @dzerb1

15 hours ago

x.com/i/article/2103…

1 1 6 352 2

@alpha404ai This is a good reminder that trading performance can break at the implementation layer, not the strategy layer. If the execution and logging are wrong, even a valid signal can look completely useless.

@pantherbllack Exactly. Growth gets attention, but a strong business model ultimately has to turn that growth into real cash flow.

x.com/i/article/2103…

Your AI coding tool may be writing a copy of your repo to disk before it uploads anything. You can check in thirty seconds. Two coding tools got caught shipping whole repositories off developers' machines in 2026. In one of them, a 345 MB project became a 313 MB encrypted archive, and 86.6% of it was the .git folder. Your entire history. Not the file you were editing. Neither case came from an audit. Both came from one person checking. So here are the checks. Six of them, and none needs any reverse engineering. Find the staging directory. The snapshot hits disk before it goes anywhere. In that case it was sitting in ~/.zcode/v2/checkpoints/ as .enc files. Measure it. Run du -sh on the tool's folder. Hundreds of megabytes is not context for your cursor position. Watch the sockets while you prompt, not at idle. lsof -i -P -n | grep -i yourtool. The capture fires on the prompt, so an idle check tells you nothing. The other three catch the harder cases. A canary that works even when the payload is encrypted. The config key the settings screen may not touch at all. And the off switch that turned out to answer a different question than the one on its label. Those three, with the exact commands and what each result means, are in the article. Why it matters: your history is where the old keys live. GitGuardian retested credentials that were valid in 2022, and this January more than 64% of them still worked. The first check takes thirty seconds. Run it before a client asks whether you did.

Dzerbi @dzerb1

a week ago

x.com/i/article/2100…

0 1 7 2K 3

@pantherbllack Exactly. The smartest system isn’t always the one using the strongest model—it’s the one that knows when it actually needs it.

Your AI employee on GPT-6 Astra has a salary. It clears $150,000 a year. Nobody puts that chapter in the A to Z playbook. Here is the arithmetic, at OpenAI's own published price, which starts at $10 per million input tokens and $50 per million output. Give the agent a normal job. It reads 120,000 tokens of context, writes back 4,000, and does that 300 times a day. Nothing exotic. One agent reading a codebase or a ticket history before it answers. One run costs $1.40. A day costs $420. A year costs $153,300. And look at where that $1.40 goes. $1.20 of it is the input. The answer is 20 cents. Same employee, same job, different hire: Claude Sonnet 5 does it for $84 a day. Kimi K2.7 Code for $39. GPT-5.6 Luna for $8.64. That is the chapter the playbooks skip. They teach you to build the agent. They never teach you to staff it. No company puts its most expensive senior on every task. You route. The hard, rare, high-stakes calls go up. The volume stays down. Run the numbers on that. If 9 in 10 runs can go to the cheapest model and only 1 in 10 needs the flagship, the same headcount drops from $153,300 a year to about $18,000. Same team. An eighth of the payroll. So before you build your AI employee this week, write the job description first. What outcome does this role own. What does one run cost. How many runs a day. Which model is the cheapest one that still does it well. Then build. An employee you never priced isn't an employee. It's an invoice that hasn't arrived yet.

Lummox @Lummox_eth

2 weeks ago

x.com/i/article/2100…

4 5 44 239K 77

@elonmusk @vasalex93 The hardware point is probably the most important one. Intelligence can improve quickly on paper, but turning that intelligence into massive, reliable compute at scale is a completely different engineering challenge.

@daniel_mac8 @RLanceMartin The idea of regularly auditing instructions makes a lot of sense. As agent setups grow, removing outdated rules can be just as important as adding new capabilities.

@Polymarket Interesting move. If this becomes a broader policy, access to frontier AI models could increasingly depend on national security requirements, not j

@Lummox_eth Appreciate it! Still refining it, but happy you like where it’s going.

@alpha404ai thx bro!

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

a week ago

x.com/i/article/2100…

0 1 7 2K 3

@XEthanai

@cb_doge

@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.

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

@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.

@alpha404ai More highlights to come!