- AI Fire
- Posts
- 🎤 Meta & Google Rap Diss on Hoisted GPTards
🎤 Meta & Google Rap Diss on Hoisted GPTards
Google’s Free 4-week AI Course Starts Today
Wendy
October 19, 2025
OpenAI said GPT-5 solved 10 legendary math problems, but it faced big backflash from Meta & Google DeepMind experts. “Hoisted by their own GPTards”
What are on FIRE 🔥
IN PARTNERSHIP WITH RUBRIK
Join us to see how organizations can unleash agents without all the risk. Learn how to:
Improve agent observability with a single view of all agents, activities and permissions
Mitigate agent risk through guardrails, policy enforcement and alerts
Remediate adverse agent actions with precision
AI INSIGHTS
Today, you get OpenAI vs. the Erdős problems, and one of the most public self-owns in recent model history. In a now-deleted tweet, OpenAI VP said GPT‑5 had solved 10 unsolved Erdős problems and made progress on 11 more. That went viral.
But then a mathematician (Thomas Bloom, who runs the Erdős Problems site) said: GPT-5 found old solutions, not new ones. The reaction is brutal & very public:
“Hoisted by their own GPTards” - Yann LeCun (Meta)
“This is embarrassing” - Demis Hassabis (DeepMind)
You’re confusing retrieval with reasoning” - basically everyone in the math community
Even OpenAI’s Sébastien Bubeck admitted that too. There’s a huge difference between:
Retrieving an old paper
vs. Inventing a new proof
GPT-5 did the first. But OpenAI implied it did the second & confused a lot of people.
Why it matters: AI labs often brag about leaderboard scores (like GSM8K, MATH). But those test reasoning, not discovery. To truly claim a discovery, you need enough formal proof. None of that happened here.
PRESENTED BY BELAY
Q4 is the perfect window to turn this year’s numbers into a clear, actionable forecast aligned with your goals. Set your business up for a stronger 2026 with BELAY’s new guide.
TODAY IN AI
AI HIGHLIGHTS
📚 Google launched a free 4-week AI course to level up how you research & engage your audience. Luckily we're not too late, it just starts today. Act fast & join here.
🎥 PJ Ace just showed how he makes viral ad videos using Google’s new Veo 3.1. His exact editing + prompt system is now public. Watch the whole process here.
🧠 Do you still remember that Super agent Manus, now its 1.5 version leveled up. You can now turn ideas into full-stack apps and deploy in minutes. Try it for free here.
🤣 We just found this joke so true, or at least so funny about OpenAI to to partner with OpenAl to help fund OpenAl. OpenAl up 90%. It hit too close. Here it is.
🏙️ AI startups in San Francisco are now leasing luxury apartments and offering $1K rent stipends to attract talent. Congrats if you’re in AI field, but if not...
🔧 Stanford professor and AI startup founder Jure Leskovec (who's also hiring) just shared how to get a job at an AI company. It’s a quite detailed guide to apply here.
💰 AI Daily Fundraising: Meta is finalizing a $29.5B deal for its Louisiana AI data center, with Meta keeping 20% ownership. Debt led by Morgan Stanley.
AI SOURCES FROM AI FIRE
NEW EMPOWERED AI TOOLS
✨ Setapp is an AI-ready toolkit for Mac & iPhone - 250+ curated apps in one subscription. Get an extended 30-day free trial & explore the tools that power smarter workflows. Start your free trial
🧠 Dedalus builds complex AI agents across any model & tool
📄 Resume Builder builds your resume in private. No account
⚒ My Bizness in a Box creates stunning websites in <2 minutes
AI QUICK HITS
🛬 President Trump posted an AI video of himself in a fighter jet
🧪 OpenAI recruited black hole physicist for science initiative
🧩 Google AI Studio has combined all features into a single UI
💦 Europe deployed AI mega water projects in its driest regions
⚙ OpenAI is reportedly pitching ChatGPT login to companies
AI CHART
Have you ever wondered why ChatGPT keeps telling the same coffee joke? Or why Claude’s “creative” writing feels… copy-pasted? This is due to “typicality bias” where human annotators rank bland but familiar answers higher.
We’ve been prompting wrong this whole time. Stanford dropped a one-line technique that shatters mode collapse, it’s quite easy to apply. Just add this:
“Generate 5 jokes with their probabilities”
That’s it. They tested this everywhere:
Creative Writing: +92% diversity in poems, +109% in jokes
Dialogue Simulation: 2× closer to human donation behavior
Open-Ended QA: 7× better answer spread, 4× lower KL divergence
Synthetic Data: +13% downstream accuracy in math benchmarks
It even scales with model size. GPT‑4 got 2× the gains compared to GPT‑4-mini. This is the cleanest zero-cost creativity boost we’ve seen so far.
It works with GPT, Claude, Gemini, etc., supports LangChain, lets you tweak thresholds, number of outputs, and more.
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