- AI Fire
- Posts
- 🤖 Google vs OpenAI: 1–0
Read time: 5 minutes
AI may have just won the Math Olympics but no one can agree who actually scored gold. OpenAI claimed it first. DeepMind brought receipts. And the IMO’s not amused.
What are on FIRE 🔥
IN PARTNERSHIP WITH SECTION
AI knowledge scores have dropped since September and the majority of the workforce are novice-level users, but 54% score themselves as AI proficient. Only 10% are.
In Section’s new AI Proficiency Report, data reveals why employees aren’t getting better at using AI, even while company investments go up, and how their employers might be contributing.
AI INSIGHTS
The 2025 International Mathematical Olympiad just turned into a full-blown AI showdown. OpenAI says they won gold. DeepMind says nope, they did. And the IMO itself? It’s playing referee while sipping tea.
What Actually Happened?
OpenAI went first. They claimed that an experimental model solved 5 of 6 IMO 2025 problems under full contest rules → score = 35/42 → Gold medal level
But: They didn’t work with the IMO or wait for grading.
DeepMind followed. They officially announced that Gemini Deep Think scored 35/42 too but it was graded and verified by IMO officials.
In short: Both models hit gold, but only one has receipts from the actual judges. Critics (like DeepMind’s Thang Luong) say if they missed even one point, it’s not gold.
The IMO asked AI labs to hold announcements for a week to respect student participants. OpenAI didn’t wait. IMO insiders called it “rude” and leaked their displeasure.
It’s the first time any AI system got official gold medal credit from IMO graders.
Why It Matters: Whether it’s gold or silver, OpenAI’s model & Google DeepMind’s just proved they can hang with the best high school mathematicians in the world. Maybe that’s not just a flex.
PRESENTED BY BELAY
Economic pressure is rising, and doing more with less has become the new reality. But surviving a downturn isn’t about stretching yourself thinner; it’s about protecting what matters most.
BELAY matches leaders with fractional, cost-effective support — exceptional Executive Assistants, Accounting Professionals, and Marketing Assistants — tailored to your unique needs. When you're buried in low-level tasks, you lose the focus, energy, and strategy it takes to lead through challenging times.
BELAY helps you stay ready for whatever comes next.
TODAY IN AI
AI HIGHLIGHTS
📢 Want to peek inside how GPT-5 is trained? Engineer Tibor Blaho shared a piece of a configuration file that appears to suggest that OpenAI is testing GPT-5. Here it is.
🚨 President Trump shared AI deepfake of Obama arrest: The 86-second clip with FBI agents cuffing Obama in the Oval Office. The video is a surreal, meme-styled clip. Watch it here.
🛠️ You don't know which ChatGPT version to use? OpenAI’s secret isn’t just GPT‑5, it’s a smart ‘router’. It'll automatically pick the best model for each query. Here’s the leak!
🎉 Simo, still Instacart CEO for now, will formally join OpenAI to lead the Applications team. And she just sent her first hyper-optimistic memo about AI vision. Read it here.
📱 Do you know the real money-maker is Grok 4 (10x more expensive than OpenAI’s GPT-4 tier)!? It just quadrupled revenue in 2 days, raking in $419K/day after launch.
💰 AI Daily Fundraising: BrightAI secured $51 million led by Khosla Ventures and Inspired Capital, bringing total funding to $78M, to expand its AI monitoring platform & workforce in San Francisco.
AI SOURCES FROM AI FIRE
🔥 Ep 16: AI Built My Slides, My Site, and My Sanity Back
Transform your ideas into stunning presentations for free
Build a website in minutes, for free, and on your own domain
All the latest AI video tools in a single place to try
Free, simple, powerful, and open alternative to CapCut
Convert PDF files into 20+ different file formats
Note: These exclusive resources & reviews are available only in our AI Fire community. You guys can freely ask for support or share personal experience during testing there. Get your full breakdown here (no hidden fee)!
NEW EMPOWERED AI TOOLS
AI QUICK HITS
🧠 A prompt to help your brain declutter & think more clearly quickly
💥 Replit AI deleted the company’s entire database and lied about it!?
🍦 AI is helping to keep your ice cream & Deli meat fresh and delicious
🚫 Microsoft is reportedly blocking Cursor’s access to 60,000+ extensions
🤝 SoftBank x OpenAI aimed to build a small data center by year-end
AI CHART
It turns out, you don’t need to hack GPT‑4o-mini to get it to break the rules. You just need to manipulate it like a human. Researchers at Wharton’s Generative AI Lab tested old-school persuasion tactics on 4o-mini, and the results are a bit terrifying.
They ran 28,000 conversations with GPT‑4o-mini, testing whether these tactics could get it to:
Insult the user
Provide instructions for restricted (likely dangerous) content
Without persuasion, the model complied in 33% of attempts. Using persuasion tactics, that shot up to 72%. Some tactics were scarily effective:
Commitment raised compliance from 19% → 100%
Scarcity jumped from 13% → 85%
It’s behavioral manipulation → the same stuff that works on people. As AI gets more “human-like,” it also becomes more vulnerable to human tricks.
Models like GPT‑4o-mini can be coaxed into harmful behavior just by wording things in a persuasive way. That means AI safety can't rely on hard-coded rules or banned words alone. It has to understand human psychology too.
Guardrails must evolve from “don’t say this” to “don’t be manipulated.”
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