LinkedIn·Wednesday, 19 August 2026·7d ago
AI still gets expensive the moment you ask it to read everything. That is the bottleneck a Miami startup, Subquadratic, says it has broken.…
IndieStudio
60 followers
AI still gets expensive the moment you ask it to read everything.
That is the bottleneck a Miami startup, Subquadratic, says it has broken. Its SubQ architecture claims a 12 million token research context window, processing 52 times faster than FlashAttention at one million tokens, for under 5 per cent of the cost of a frontier model on comparable long-context work. Those are company-reported figures, and researchers have publicly asked for independent proof.
The headline number is not the interesting part, though. The operator question underneath it is: what changes when AI can cheaply read more of the business at once? Whole codebases. Contract rooms. Support histories. Product docs. Company memory.
But accepting more tokens is not the same as reasoning over them. Long context is not a vanity metric. A model that can read more is not automatically a model that understands more.
So judge these tools by what they can prove: what did it find, what did it reason over, what did it cite, and what did it actually act on.
Don't ask how many tokens it accepts. Ask what it can reliably do with them.
#AI #LLMs #AIinfrastructure
♥ 1
View on LinkedIn Cross-referenced
Related on the wire
Ten competitor checks, one Monday digest. Here is a simple way you could structure it: choose the pages, fetch them weekly, compare the…
Ten competitor checks, one Monday digest. Here is a simple way you could structure it: choose the pages, fetch them weekly, compare the…
AI agents are turning websites into infrastructure they can consume. Cloudflare's new Monetization Gateway is not just a publisher story.…
Prompts are instructions. Controls are enforcement. Google's managed Gemini agents can now run scheduled work with checks before and after…
Voice AI is not just getting smoother. It is becoming a live work interface. OpenAI’s GPT-Live can listen and speak at the same time, which…
Cursor Origin brings code hosting into the same environment as the coding agent. That can remove handoffs across repositories, reviews and…