😺 This new AI subQ might kill the transformer.

Q: Subquadratic launched SubQ, a 12M-token LLM on a sub-quadratic architecture with $25M seed funding,

Subquadratic launched SubQ, a 12M-token LLM on a sub-quadratic architecture with $25M seed funding, running 52x faster than FlashAttention at 1/5 the cost of frontier models.

Q: Google was sued for $1.5M over a false AI Overview and Character.AI was sued by Pennsylvania for cha

Google was sued for $1.5M over a false AI Overview and Character.AI was sued by Pennsylvania for chatbot medical impersonation — both on the same day.

The Neuron·Wednesday, May 6, 2026·9 min read

AI/ML Technology Startups

Share𝕏 in

AI Summary

Subquadratic launched SubQ, a 12M-token LLM built on a sub-quadratic architecture (SSA) with $25M in seed funding, claiming 52x speed over FlashAttention and 1/5 the cost of frontier models. Anthropic released ten finance agents and reportedly committed ~$200B to Google over five years for cloud capacity. AI faced two lawsuits in one day — Google for $1.5M over a false AI Overview, and Character.AI for chatbot medical impersonation.

Key Facts

✓Subquadratic launched SubQ, a 12M-token LLM on a sub-quadratic architecture with $25M seed funding, running 52x faster than FlashAttention at 1/5 the cost of frontier models.

✓Google was sued for $1.5M over a false AI Overview and Character.AI was sued by Pennsylvania for chatbot medical impersonation — both on the same day.

✓A free four-line prompt can convert raw notes into a fully backlinked Obsidian vault using atomic markdown files with wikilinks, requiring no RAG or vector databases.

Author Takes

BullishThe Neuron

SubQ vs. transformer replacement claims

The receipts look different this time compared to Mamba — PhDs from Meta, Google, Oxford, and Cambridge are behind SubQ and API access is live today, suggesting this could be a genuine architecture shift.

SkepticalThe Neuron

Marc Andreessen's system prompt

'World-class expert in all domains' is cargo-cult prompting from GPT-3.5 days that does literally nothing, while 'make answers as long as possible' is actively counterproductive, producing padding rather than completeness.

BearishThe Neuron

RAG and agent scaffolding as technical debt

If SubQ's architecture truly holds 12M tokens cheaply, much of the RAG and orchestration scaffolding stops being load-bearing — you skip chunking, embedding, and orchestration and just ask.

Contrarian Angle

Strip Flattery, Keep the Anti-Sycophancy Protocol in AI Prompts

Marc Andreessen's system prompt shows that 'world-class expert' framing is cargo-cult prompting, but explicit anti-sycophancy instructions ('never praise my questions', 'do not anchor on my numbers') measurably change model behavior.

Counterintuitive: flattery-style system prompts do nothing, but behavioral anti-sycophancy rules produce verifiable output changes most users never implement.

SubQ 12M-token context replacing RAG / chunking pipelines

SubQ's native 12M-token context window could replace RAG, embedding, and orchestration scaffolding by letting users skip chunking and just query directly.

Engineers switching from RAG / chunking pipelines to SubQ 12M-token context

More from The Neuron

😺 Google is killing the prompt box

Google announced Gemini Intelligence for Android, Magic Pointer (a context-aware cursor powered by Gemini), and a new premium laptop category called G

May 13

😺 Your AI bill is creeping up. Here's why.

Cerebras is going public Thursday at a $33B valuation after upsizing its IPO to $4.8B, backed by 20x oversubscription and a $20B+ OpenAI compute deal.

May 12

😺 You're more AI-ready than your boss

The Neuron's May 11, 2026 edition covers Microsoft's 2026 Work Trend Index showing workers are more AI-ready than their organizations, a new AI tool p

May 11