- AI Fire
- Posts
- ⚡️ GooglePT Haunts DeepSeek’s Graveyard
⚡️ GooglePT Haunts DeepSeek’s Graveyard
Free 1-year Full ChatGPT Access for Students
Wendy
September 30, 2025
In partnership with
Read time: 5 minutes
GPT-5 just helped prove a quantum theorem. And in China, DeepSeek launched a model that slices AI costs in half, while U.S. labs are still feeding GPUs like firewood
What are on FIRE 🔥
IN PARTNERSHIP WITH SECTION
On November 6, join Section for a half-day of micro-workshops designed to turn you from an AI prompter to a workflow redesigner. You’ll learn how to build LLM automations, design agentic workflows, and try your hand at vibe coding. Walk away with practical frameworks and a certificate of completion.
AI INSIGHTS
While U.S. labs solve long prompts with more Nvidia GPUs, China’s DeepSeek doesn’t have that luxury. They dropped DeepSeek-V3.2-Exp that cuts costs in half without losing performance.
Remember the OG Transformer from 2017? It checks every word against every others in a sentence. With 10K words, that’s 100M comparisons. Sparse attention flips that:
It picks only the ones that matter
It builds a tiny neural module called the “lightning indexer”
It holds up in benchmarks vs. their previous dense model (v3.1 Terminus)
Public API prices already slashed, up to half the cost for long inputs
Actually, aparse attention isn’t new, OpenAI pioneered it with sparse transformers in 2019. Google dropped Reformer in 2020.
But the extent to which GPT-4 or GPT-5 actually use sparse attention? No one knows. OpenAI’s been mum. DeepSeek is the first major lab to go, “Hey, here’s our implementation and it works.”
Why it matters: I believe this is a full-on efficiency flex. Every time you chat with ChatGPT, it still reprocesses all prior tokens. Even if it already saw them. If GPT-5 uses sparse attention, it’s clearly not doing enough to fix that slowdown.
PRESENTED BY 1440
Join over 4 million Americans who start their day with 1440 – your daily digest for unbiased, fact-centric news. From politics to sports, we cover it all by analyzing over 100 sources. Our concise, 5-minute read lands in your inbox each morning at no cost. Experience news without the noise; let 1440 help you make up your own mind. Sign up now and invite your friends and family to be part of the informed.
TODAY IN AI
AI HIGHLIGHTS
🆓 Why spend money when so much is offered for free? An AI author, Thomas Perry, tested the 5 best free & popular AI chatbots now on the market. Here's his verdict.
🧠 USC just gave Free full access to ChatGPT for students, staff, and faculty under a $1.5M deal. It's now valid for 1 year with an option to renew. Claim yours here.
🤖 A guy just built an AI ad machine using Nano Banana, n8n & Claude that cranks out TikTok, Facebook & Instagram ads nonstop. Get his complete system here.
📑 As usual, Anthropic just published another banger. Today’s one is on how context engineering differs from prompt engineering. Get this must-read for any AI lovers.
🎥 OpenAI just launched Sora 2, an AI video app that makes 10‑sec clips from copyrighted material unless owners opt out. Disney already bailed it. Try it here.
🖱️ Cursor now lets you control your browser, capture screenshots, enhance UI, and debug client issues in Claude Sonnet 4.5 (just release yesterday). See it here.
💰 AI Daily Fundraising: Assort Health secured $76M for its voice AI platform. With 14 languages, it has handled 42M patients & boosted revenue eightfold.
AI SOURCES FROM AI FIRE
NEW EMPOWERED AI TOOLS
💼 Wing Assistant gives you a full-time virtual teammate for admin, ops & sales, no hiring needed. Just a trained assistant to delegate everything that slows you down. Already behind 5,000+ teams.
✨ Imagine with Claude gets software as you type in real-time
🎬 CrePal creates short movies with just a prompt with AI director
⚡ Creavit Studio generates pro social media content in seconds
AI QUICK HITS
🛡️ Google Drive uses AI to spot ransomware & swiftly restore files
👩🔬 20 top AI researchers left OpenAI, Google, Meta for a startup
☁️ Lovable introduced Cloud & AI to build apps by prompting
🔍 Google AI Mode can now help you search & explore visually
📢 Amazon unveils new Echo devices, powered by its AI, Alexa+
AI CHART
Scott Aaronson, a legend in quantum computing, just dropped a paper that includes a key technical step in the proof came from GPT-5-Thinking!
First, QMA is like the quantum version of NP, a world where a “quantum witness” convinces a verifier of a “yes” answer. For years, there’s been a (very annoying) ques:
→ Can we always make QMA verification perfect?
In classical settings, we can amplify correctness up to 1. But for QMA, it wasn’t clear. Scott wanted to prove that this new amplification result is actually optimal.
But he got stuck on one thing & asked GPT-5-Thinking. After a few bad ideas (and course corrections), GPT-5 suggested a function that Scott said:
→ “If a grad student had given it to me, I’d have called it clever.” “I guess I should be grateful that I have tenure.”
→ GPT helped break the mental block. That’s the part that matters:
The suggestion was non-obvious and genuinely useful
This is just the first of what will be thousands of examples like this
Goldman Sachs warned the AI bubble might burst soon, taking down data centers, Google, Meta, OpenAI, Nvidia, and more. And guess what’s next? They’re pivoting hard to quantum computers.
We read your emails, comments, and poll replies daily
How would you rate today’s newsletter?Your feedback helps us create the best newsletter possible |
Hit reply and say Hello – we'd love to hear from you!
Like what you're reading? Forward it to friends, and they can sign up here.
Cheers,
The AI Fire Team
Reply