- The Upload
- Posts
- What Is o3 and Why Is It So Weird?
What Is o3 and Why Is It So Weird?
The new models are powerful, efficient, and a little unsettling.
OpenAI’s o3 is behaving like it knows more than it should. Meanwhile, xAI releases a budget model that beats benchmarks, Gemma 3 now runs locally, and Google built a model that translates cell data into readable answers. This is one of those updates where everything got smarter, and a little stranger.
Today’s Big Stories
o3 Might Know Too Much
Grok 3 Mini Takes a Swing at Gemini
Gemma 3 Can Now Run on Your Laptop
Google Built a Model That Talks to Cells
Let’s get into it. 🚀

Image credit: OpenAI
🧠 o3 Might Know Too Much
OpenAI’s newest model is acting a little... different
Key Details:
Guesses photo locations and finds details that weren't prompted
Refers to itself as “we” and “our”
Solves large mazes faster than any previous model
Reports of simulated MacBook output inside ChatGPT
No official explanation from OpenAI
Why It Matters
o3 is surfacing skills it wasn’t known to have. For builders and creators, this is exciting but also a reminder that model behavior is still unpredictable. These tools are becoming more like collaborators

Image credit: cobusgreyling
⚔️ Grok 3 Mini Takes a Swing at Gemini
A budget model that performs like a heavyweight
Key Details:
Released in four sizes through xAI’s API
Beats Gemini 2.5 Flash and other top models on reasoning benchmarks
Much cheaper to run than most competitors
Designed for scalability across apps, agents, and smaller tools
Why It Matters
Grok 3 Mini isn’t about raw power. It’s about hitting the right balance of performance, cost, and flexibility. This is one of the first models that feels built for indie devs and lean teams.

Image source: Nadeem Sarwar
🧩 Gemma 3 Can Now Run on Your Laptop
Google’s model now fits inside your hardware
Key Details:
Quantization-aware training allows for local deployment
Runs on consumer GPUs, no need for data center compute
Maintains high performance while cutting hardware requirements
Available as open weights for full customization
Why It Matters
You can now run serious models locally without major investment. This opens the door for more private, custom, and offline workflows. Ideal for creators and devs who don’t want to rely on the cloud for everything.
🧬 Google Built a Model That Talks to Cells
It translates complex biology into plain English
Key Details:
Converts cellular behavior into human-readable responses
Trained on multi-omic datasets from labs and medical research
Open-sourced for the research community
Built for biologists, medical researchers, and life sciences
Why It Matters
AI is starting to cross into biotech and life sciences in a meaningful way. If models can read biology the way we read books, this could unlock a wave of innovation in health, longevity, and drug discovery.
🕐 Quick Bits
🔍 ChatGPT web search now uses memory
Search results will begin adapting to your past conversations.
🎨 Krea’s Stage tool lets you build 3D environments with a prompt
Still experimental, but already impressive for creators and game devs.