@J_Von_Random @danluu Based on understanding the architecture of a transformer model, and seeing the sheer amount of recursion that is needed to have an LLM operate.
English
ohnotheirtokenlimit
749 posts
> markets close
> “WE ARE BOMBING AGAIN!”
*5mins before markets reopen*
> “We’ll have a deal soon”
The Spectator Index@spectatorindex
BREAKING: CBS News reports that the Trump administration is preparing for new strikes against Iran
English
Honestly why stop at 100x engineer?
Just use more agents, you literally could be 1000x, 10000x, 100000x just by scaling
You could what you use to in an entire year in one second
English
@viktorg475 @danluu It’s only obvious if you don’t understand how transformer models work.
English
Why are so many people so sure that the big AI providers are losing money on inference? It reminds me of the comments about how Uber can never make money. Their unit economics were fine and they were only losing money because they chose to do so on customer acquisition.
English
@RyanJTopps @danluu And that is then cooking the numbers to get the most favourable number that they could!
English
@danluu Open AI’s operating margin is -122% so yeah it’s confirmed at this point
English
@CosmicSenate @Fooney_ @SocialistMormon @danluu I’ll tell you now, enterprise customers are billed higher, but the cost is still heavily subsidised.
They need enterprise to have no choice but use LLMs, THEN the prices go to what they need for a heavy profit margin.
English
@Fooney_ @SocialistMormon @danluu u have no idea what you cost them, just how much they charge enterprise users.. they for sure have high margins on the api token rates
English
@jintaeghong @danluu This is the raw details that they don’t look into.
Once you build even a small model (not run one, BUILD one with PyTorch), you realise just how many calcs are being ran for inferring one token. Then look at the number of tokens being spat out in LLMs and agent flows.
It’s insane
English
@danluu If you know the mechanism of the transformer model, you'll understand why it cost more than human when it comes to replacing human labor
English
@LandonExplr @danluu Give us the total spend for AWS, and then for any major LLM company.
English
@danluu AWS was 'unprofitable' for years too. Inference unit economics are fine at reasonable utilization. The loss claims always price hardware per query instead of amortizing capex over years.
English
And at that point, general estimates seem to be suggesting the costs will 40x to consumers of tokens.
They are running them cheap, like Uber did, to capture market and make it hard to leave.
This is also why they try to stop ppl becoming engineers; it makes LLMs the only choice
English
“How can they be so sure they are losing money on inference? Uber lost money too, to capture market”
Two points:
1. You say they cant be losing money but then explained that they are, but its for a reason
2. We agree with you. Once they get market share the token subsidy vanishes
Dan Luu@danluu
Why are so many people so sure that the big AI providers are losing money on inference? It reminds me of the comments about how Uber can never make money. Their unit economics were fine and they were only losing money because they chose to do so on customer acquisition.
English
Just straight up ignoring the hardware that requires COBOL, in this post.
How many EC2 instances are you gonna need to match the MIPS rating for a mainframe in banking?
How much is that cloud bill gonna cost?
U.S. DOGE Service@USDS
If you don’t know what COBOL is, it’s a programming language created in the 1960s. And yes, some of the federal systems Americans rely on today are still running on it. That means parts of the technology behind critical government operations were built before the internet, smartphones, or GPS even existed. When we talk about “legacy systems,” this is what we mean. America cannot build the future on infrastructure built for another era. That’s why bringing the best tech talent into government matters. That’s why work is underway to move these systems forward.
English
@siddak_b @ThePrimeagen Use another 50,000 agents to review the code
English
@ThePrimeagen 100000x slop that's impossible to review at that speed
English
“S you want everyone in the movie to wear crocs?”
“Yeah, there’s no way crocs will become popular, and if they do, then maybe the movie has become real life”
FactPost@factpostnews
A merch website run by Kash Patel is trying to trick visitors into installing malware. Customers are instructed to copy seemingly benign text into their terminal, but clicking copy and pasting installs software stealing personal data.
English
ohnotheirtokenlimit retweetledi
You can no longer Google the word ‘disregard’ techcrunch.com/2026/05/22/you…
English
@Polymarket I want andreesen to get liquidated sooo fucking bad. Wipe that smug smile off his egg head face.
English
NEW: Marc Andreessen declares AGI was achieved three months ago.
English
@Hermes4Megistus @TheMarketDog And that, ladies and gents, is the typical AI Bro. All claims, no actual substance.
English
@tokenlimitto0 @TheMarketDog No. You are just low iq and I don’t owe you anything
Muted
English
Imagine running a 16 bn annual loss and being valued 2tn.
*Walter Bloomberg@DeItaone
*SPACEX 1Q REVENUE $4.69B *SPACEX 1Q NET LOSS $4.28B *SPACEX: GOLDMAN, MORGAN STANLEY, BOFA LEADING SPACEX IPO
English
Serious question; Do you think Sam has been spending the last 24hrs prompting and re-starting chat threads with ChatGPT to try and produce an S1 document that doesn’t highlight the absolute shit state of the company?
Exec Sum@exec_sum
BREAKING: OpenAI is preparing to file for an IPO, possibly as early as Friday.
English