Sabitlenmiş Tweet
quantzoid
258 posts
quantzoid
@quantzoid
lion amongst men. ML eng.
the microseconds btwn trades Katılım Mayıs 2026
78 Takip Edilen18 Takipçiler
@atelicinvest @evergreencap3 Besides the model, tokens gain value via optimization and inference engineering techniques (such as quantization and KV cache optimization), which makes them denser and more valuable.
English
@evergreencap3 I might also push back on the tokens getting better every day. Both Anthropic and Google released new models that were widely panned, even though on benchmarks they showed big improvements. This is not indicative of acceleration the way VCs want you to believe in
English
There is a view that Anthropic and OpenAI revenues are being inflated by 'tokenmaxxing,' and that customers will soon cut their token budgets in response. I think that's extremely unlikely.
Tokens continue to get better, smarter, and more valuable every day. Model progress is accelerating. Companies that don't spend enough on tokens face systemic risk from competitors that do.
This creates an environment where any wasted tokens will be reallocated to users and use cases that are value-creating, rather than being cut outright. We're still at the beginning of a multi-decade paradigm shift, and the risk of underinvestment is simply too high, even at the expense of near-term margins.
Token spending is the new cost of doing business. And it’s unlikely to flatten anytime soon.
English
The non-consensus read for $SNDK (and memory):
Even if China is flooding the global memory supply, I'm skeptical if hyperscalers can use those products willy-nilly for compliance reasons.
Remember, Enterprise SSD for AI is a completely separate market.
TrendForce's Q1 data shows that NAND revenue is up 83%.
I think the dump really hurts consumer NAND, not the AI SSD business.
quantzoid@quantzoid
In short, Tuesday is rough if the market doesn't understand that China is attacking commodity memory, not advanced memory. Remember, HBM is completely sold out from 2026 all the way through 2027. Because of the deals that are being signed at Computex, I'm bullish.
English
If you want something spicier:
x.com/quantzoid/stat…
quantzoid@quantzoid
The non-consensus read for $WDC (and memory): Even if China is flooding the global memory supply, I'm skeptical if hyperscalers can use those products willy-nilly for compliance reasons. Remember, Enterprise SSD for AI is a completely separate market. TrendForce's Q1 data shows that NAND revenue is up 83%. $WDC is explicitly mentioned in that report. I think the dump really hurts consumer NAND, not the AI SSD business, which $WDC is in.
English
English
English
I’m sorry but the orbital data center idea is completely ridiculous. We understand why a few people have to pretend it’s real, but every one else does not.
English
@mxschumacher @hissklarman @when_to_fold_em Data centers need a) robust grid infra b) fiber connectivity and c) stability. Can’t exactly build it in the Congo.
English
@quantzoid @hissklarman @when_to_fold_em plenty of places on planet earth where it's easy to build large scale data centers, it's not like the whole surface of the earth has the political climate and regulatory environment of San Francisco
English
why would anyone use meta’s mods? Are they just going to be a neocloud?
kuz@kylekuzma
Yep wait till their data center build out 🔥🔥🔥 heavy capex
English
@MacroGhost7 @hissklarman @when_to_fold_em Because of scaling laws: arxiv.org/abs/2001.08361
Hyperscalers are currently paying for it.
English
@quantzoid @hissklarman @when_to_fold_em I mean you also have to take a step back and ask yourself… Why exactly do we need THIS much compute?
Who’s going to pay for it?
English
@hissklarman @when_to_fold_em Yes? Do you think it’s politically popular to build data centers or something?
English
@quantzoid @when_to_fold_em Yeah way easier to deliver a MASSIVE MASSIVE payload to space than figure out somewhere on earth to put your data center. So dumb
English
@when_to_fold_em In space you don’t have to worry about land permitting or NIMBY opposition
English
@quantzoid Because power is not in shortage on earth and power generation capacity is scaling much faster than other bottlenecks
English
in a sense, c++ was always vibe coded
os@segun_os_
you literally cannot vibe code c++ char is an 8 bit string that can also be an unsigned 8 bit int depending on how you use it. the level of precision required to write c++ is too high for vibecoding. there are just too many quirks in the language.
English
Very good lecture. Take notes!
Dwarkesh Patel@dwarkesh_sp
New blackboard lecture w @reinerpope How do chips actually work – starting with basic logic gates, and working up to why GPUs, TPUs, FPGAs, and the human brain each look the way they do. 0:00:00 – Building a multiply-accumulate from logic gates 0:16:20 – Muxes and the cost of data movement 0:25:59 – How systolic arrays work 0:39:00 – Clock cycles and pipeline registers 0:51:40 – FPGAs vs ASICs 1:03:14 – Cache vs scratchpad 1:07:16 – Why CPU cores are much bigger than GPU cores 1:11:49 – Brains vs chips 1:15:22 – A GPU is just a bunch of tiny TPUs Look up Dwarkesh Podcast on YouTube/Spotify/etc to watch. Enjoy!
English