🚀 GPT-o3 Finally “Thinks with Images”

Free AI Tutorials | AI Academy | Advertise | AI Agent No-code n8n

Plus: Curated List of Awesome MCP Servers

Read time: 5 minutes

OpenAI ship week is still going on. Its latest “o-models” can even zoom in, analyze a blurry image using ‘chain of thought.’. Plus, Google hires researchers for “post-AGI” work. AGI might already be achieved internally? AGI is no longer a far-off dream?

What are on FIRE 🔥

• 🚀 o3 and o4-mini can “think with images”

• 🌟 AI Highlights

• 📚 AI Sources From AI Fire

• 🏅 AI Tools

• ⚡ 5 AI Quick Hits

• 📃 AI Cheat Sheet

• 📊 AI Chart

• 💼 4 AI Jobs

AI INSIGHTS

🚀 o3 and o4-mini can “think with images”

Another day of OpenAI ship week. They just announced o3 as “the most powerful reasoning model to date” and o4-mini as “a competitive trade-off”. They may be the last stand-alone AI reasoning models before GPT-5.

=> “For the first time, our reasoning models can independently use all ChatGPT tools – web browsing, Python, image understanding, and image generation”. Is it an all-in-one GPT agent? Full tools support includes parallel tool calling.

Key Capabilities and Highlights:

• Visual Thinking: o3 can zoom in, crop, read even blurry, reversed, or low quality images… with a scanner. This could be the time when it beats the “Pokemon game”.

• Next-gen API Use: Developers can now plug these into their workflows to handle complex problems.

• Benchmark-Leading Accuracy: Both models outperform previous versions from PhD-level science to competition math. Smashes SWE-bench Verified with 71.7% – leaving o1 (48.9%) and Claude 3.7 Sonnet in the dust.

• Codex CLI: A terminal coding assistant built for developers. Multimodal (accepts screenshots/sketches). Fully open source.

A plot twist for the history books: OpenAI originally planned to bundle o3 into GPT-5. They split it off early to avoid getting steamrolled by Google’s Gemini 2.5 Pro and Anthropic’s Claude 3.7. In the coming weeks, OpenAI will release o3-pro just for Pro users.

Why It Matters: OpenAI’s o3 and o4-mini are the first models that actually feel agent-like. “o-series” is actually a quiet prototype for AGI behaviors. Is traditional “chat” AI dying? Coders, analysts, consultants, students – many of their steps can now be offloaded entirely.

TODAY IN AI

AI HIGHLIGHTS

🧠 Microsoft just dropped a hyper-efficient AI model called BitNet. It runs fast on regular CPUs, beats Meta, Google, Alibaba models of similar sizes in tests, but doesn’t work on GPUs yet.

🔍 Google is hiring researchers for “post-AGI” work. AGI might already be achieved internally? AGI is no longer a far-off dream?

🕵 Gamma AI platform abused in phishing chain to direct unsuspecting users to spoofed Microsoft login pages. All just through a seemingly innocuous email with Fake PDF attachment.

🚨 DeepSeek steals U.S. data, spreads CCP propaganda, and illegally uses over 60,000 Nvidia chips. 85%+ of DeepSeek’s responses are manipulated without disclosure to users.

🔥 OpenAI already has its own coding tools but is in talks to acquire Windsur for $3B because Cursor’s likely tied to Amazon. It feels a lot like Facebook buying Instagram in the early mobile era, doesn’t it?

🧠 Microsoft researchers report “Longer responses aren’t a sign of deeper reasoning, they often indicate confusion”. Time ≠ intelligence.

💰 AI Daily Fundraising: Auradine just raised $153M in Series C funding, hitting $300M+ total. They build energy-efficient AI and Bitcoin infrastructure and just launched AuraLinks AI to boost next-gen data centers.

AI SOURCES FROM AI FIRE

IN PARTNERSHIP WITH GREAT LEARNING

The AI & Data Science Boom Is Real. Get Job-Ready – Now at 50% Off!

AI is evolving at lightning speed. But let’s face it – you’re not a machine trained on trillions of tokens. So how do real people get job-ready, fast?

We’ll let you in on what top learners know (but rarely share):

✅ Learning from academic experts? That’s just 50% of the equation.
✅ The other 50%? Real-world projects + expert mentorship.

That’s how you Learn Right. Learn Fast and go from beginner to skilled in no time.

What You’ll Get:

🚀 Expert Guidance: Learn from top university professors and industry pros.
🛠 Hands-On Projects: Work on real-world challenges and code your way to confidence.
🤖 24/7 AI-Powered Mentor Support: Instant feedback, mock interviews, and guidance tailored to you.
📈 Career-Powering Certificates: Showcase skills that hiring managers are actively looking for.

This isn’t just another course. It’s your launchpad to a thriving career in AI and Data Science.

⏳ Enroll Now — 50% Off! Start Today for Just $40 (Limited-Time Offer)
❌ No strings. No pressure.
✅ Just click, preview the courses for free.

NEW EMPOWERED AI TOOLS

1. 👩‍💻 OpenAI Codex CLI is an open‑source, local coding agent powered by the latest o3 and o4‑mini models.

2.  🔎 Anthropic’s AI now searches your entire Google Workspace without you.

3. 🤖 Potpie AI builds task-oriented custom agents for your codebase.

4. 📞 ServiceAgent handles customer inquiries, and captures every lead 24/7.

5. 🐾 DocsHound turns product demos into full chatbots or product education.

AI QUICK HITS

1. 💊 Google DeepMind Labs rethinks drug discovery for personalized medicine.

2. 🤖 Hugging Face launches open source AI robots to fight Tesla Optimus.

3. 🤔 AI-generated design just won an Oregon State University T-shirt contest?

4. 💥 Following ChatGPT, Anthropic app will do voice internally with 3 options.

5. 🌎 AI just helped scientists find 44 stars having Earth-like planets.

AI CHEAT SHEET

AI CHART

🛡️ Google’s Breakthrough in Fighting Against “Prompt Injection”

🧠 CaMeL – first serious solution that doesn’t rely on “more AI”.

“Prompt injection” has haunted developers since chatbots went mainstream in 2022. Despite many attempts, no one has fixed this core flaw – like whispering secret commands to hijack a system. Until now, perhaps.

What is prompt injection, anyway?

=> A prompt injection is when malicious text hidden in content (e.g. emails or docs) tricks an AI assistant into doing something it shouldn’t.

So, AI agents are being embedded into real systems. Then, a prompt injection could now mean real damage: moving money, leaking documents, or sending emails to the wrong person.

Do you know that Apple’s Siri still avoids complex assistant tasks partly because of this exact risk?

Google DeepMind’s CaMeL – a new approach to stopping prompt-injection attacks. It doesn’t try to detect prompt injection with more AI, it applies decades-old software security principles. It splits responsibilities between two AIs (one can act, one can’t) and tracks the flow of data like plumbing, blocking actions if they involve untrusted sources.

=> This is a shift in how we secure AI agents from naive trust to explicit control. Google’s move will likely push OpenAI, Anthropic, and Meta to adopt similar defenses – or… risk falling behind in enterprise trust.

AI JOBS

• Meta: AI Scientist, Supply Chain

• Stanford University: Data Scientist, GUIDE-AI Lab

• Amazon: AI Tutor, Rufus, Amazon

• Google: Strategic Program Delivery Lead, Content and AI

We read your emails, comments, and poll replies daily

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