Introducing SubQ - a major breakthrough in LLM intelligence.
It is the first model built on a fully sub-quadratic sparse-attention architecture (SSA),
And the first frontier model with a 12 million token context window which is:
- 52x faster than FlashAttention at 1MM tokens
- Less than 5% the cost of Opus
Transformer-based LLMs waste compute by processing every possible relationship between words (standard attention).
Only a small fraction actually matter.
@subquadratic finds and focuses only on the ones that do.
That's nearly 1,000x less compute and a new way for LLMs to scale.
OpenAI JUST released how people are using chatgpt
each bar in this chart is a billion-dollar wedge if you build the right verticalized, trust-rich AI startup:
1/ tutoring + teaching (10.2%) - people want on-demand teachers more than almost anything else. a personal ai tutor that explains things your way, remembers your progress, and nudges you daily.
2/ how-to advice (8.5%) - this is consumer SaaS for micro-niches: “how to fix my resume,” “how to meal prep,” “how to set up my Shopify store.” every “how-to” is a wedge into a vertical AI agent.
3/ personal writing + editing (18% combined) - 8% of traffic is people asking AI to write emails, 10.6% is editing or critiquing text. this is demand for AI copilots inside every workflow tool: sales, legal, HR, PR.
4/ health, fitness, beauty, self-care (5.7%) - validated consumer wedge. people already trust AI with their bodies. AI trainers, AI nutritionists, AI skin coaches — with a human layer for accountability.
note: btw if you like seeing ideas/trends like this, you'll like @ideabrowser (free/paid plans)
5/ purchasable products (2.1%)- tiny % but huge monetization. people literally asking AI what to buy. the Amazon affiliate model gets reborn here. whoever builds the “AI-powered shopping layer” owns the new SEO.
6/ translation (4.5%)still unsolved. what people want isn’t raw translation, it’s contextual translation: tone, culture, slang. AI that doesn’t just translate but localizes your intent.
7/ computer programming (4.2%) - we know this, but it’s worth underlining: devs want copilots. the non-obvious play is vertical code copilots... Shopify dev copilot, Unreal Engine copilot, etc.
TLDR;
every bar is a behavior that people are already paying for elsewhere. translate those behaviors into vertical ai agents and startups.
the data doesn't lie.
"Introduction to Machine Learning Systems"
- FREE from MIT Press
- Authored by Harvard Professor
- 2048 Pages
To Get It Simply:
1. Retweet & Reply "ML"
2. Follow so that I will DM you.
Me as Defense Against the Dark Arts Professor: "Today class, we are going to take on a boggart. But first, are any of you afraid of clowns?"
Class: *three hands*
Me: "You three, get the fuck out."
"Another Life": in the future, emotionally unstable teenagers with zero fear for the consequences of their actions try to save the world using only science learned from watching bad science fiction shows