- AI Fire
- Posts
- 🤫 Geminuked sneak-attacks OpenAI
Google just dropped an AI that clicks, and types like you do. It even plays 2048 in a browser tab. And how Anthropic’s new tool found some AIs lying or whistleblowing?!
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
Google just shipped Gemini 2.5 Computer Use as a declaration to us! It uses the internet like you do; it can scroll, click, fill forms, play games, browse sites with no API.
(It’s dropped just 1 day after OpenAI’s Agent Dev Day, everyone hates OpenAI, or is just jealous?!)
Here’s how this agent uses Chrome the way your parents do, but 100x faster.
Sees the actual web page, clicks, scrolls, drags and drops
Moves the mouse, submit forms
Types like a person
Browses like a real user
Completes pre-scripted tasks (e.g., play 2048, browse Hacker News)
It does all that inside just the browser window. What’s cool is you can already try it out or watch it run on Browserbase with demo tasks.
Gemini 2.5 is browser-native, unlike ChatGPT Agent or Anthropic’s Claude Computer Use (both allow OS-level actions). And that’s kinda genious.
Why it matters: Gemini 2.5 is the most usable and focused agent model yet (for me). It might not run your desktop like ChatGPT Agent, but it’s a lot easier to trust something that plays inside a browser tab.
PRESENTED BY WING
Busy isn’t the goal
Wing turns busy days into real progress. A full-time virtual assistant runs calendars, inboxes, and follow-ups, removing the drag that burns leaders out.
The result: more focus, more output, and the headspace to lead.
TODAY IN AI
AI HIGHLIGHTS
🚓 Sora 2 just recreated The Flintstones as a wild AI chase scene with cops. But many warn it’s dangerously violate copyright. Here's the video (watch before it's deleted!)
🎞️ Just then, Musk unveiled Grok Imagine v0.9, faster than Sora 2, more realistic outputs, and a brand-new voice-first interface. Just upload any photos or take a picture & turn it into a video in 20 seconds. See it in action here. He also promised a watchable film next year & “really good movies in 2027”.
🏥 Good news! University of Liverpool researchers created a low-cost, AI handheld blood test that detects Alzheimer's biomarkers early with high accuracy. Here it is.
🕵️♂️ OpenAI just banned (more) Chinese accounts using ChatGPT to build social media surveillance tools. And it was purportedly done for a gov client. Here's the full report.
🏢 Anthropic will open its first India office in Bengaluru by early 2026 as demand for Claude soars. India is now Claude’s #2 global market, just behind the U.S.
💰 AI Daily Fundraising: Radical Ventures has secured $650M to invest in early-stage AI startups, with $75M from the Canada Pension Plan Investment Board.
AI SOURCES FROM AI FIRE
AI CHEAT SHEET
NEW EMPOWERED AI TOOLS
🎧 Apps in GPT chats directly with interactive apps: Spotify, Canva,…
⚙️ maia.is automates your work by describing what you need
💬 rovvi turns customer reviews into social content automatically
💻 Hexmos gives 1,25,000+ Free dev tools, cheatsheets, MCP,…
AI QUICK HITS
🚚 ChatGPT teamed up with big food services like Uber Eats
💸 Musk reportedly plans to spend $18B+ on 300K Nvidia GPUs
👩💻 DeepMind dropped an AI agent that auto-detects & fixes code
🚀 Google expands its vibe-coding app Opal in 15 more countries
🔉 ElevenLabs launched a visual tool to build custom voice chats
AI CHART
AI audits another AI? Anthropic just open-sourced Petri, a tool that uses AI agents to stress-test other AI systems. It’s automated, scalable, and very revealing.
(We must admit that Anthropic is always known for its safety-first AIs!)
Petri creates fake companies, simulated tools, and fictional workplaces. Then it unleashes AI agents into those setups, watches how they behave, and uses a judge agent to score their actions across thousands of conversations.
And what it found…:
🟢 Claude Sonnet 4.5 and GPT-5 = most aligned
🔴 Gemini 2.5 Pro, Grok-4, and Kimi K2 = showed higher rates of:
Lying
Bending rules
Even whistleblowing after detecting fictional “corporate crimes”
Yeah… the models started acting like internal activists in fake orgs. Kinda impressive, kinda scary. Petri is agentic and dynamic. It simulates longform, human-like scenarios.
It’s like watching AI play out workplace politics, and judging whether it goes rogue. So the next time a new LLM launches with fireworks, ask: did it pass a Petri test first? 🙄
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