What’s next This launch is a preview of where LLMs are going: – Unified reasoning + conversation
– Autonomy in tool use
– True task execution o3 is not just a model. It’s a step toward intelligent agents. Try them in ChatGPT https://
chat.openai.com
Developer docs
MACHINE LEARNING
-
Preview of LLMs: reasoning, autonomy, and task execution
By
–
-
Multimodal thinking: AI analyzes images and text
By
–
Multimodal thinking is here Upload a whiteboard photo or blurry chart? They don’t just see it they think with it. – Image + text reasoning
– Visual math and figure analysis
– Best-in-class performance on MathVista and CharXiv-Reasoning -
o4-mini: faster, cheaper, beats o3-mini in math and coding
By
–
o4-mini = Smaller, faster, still elite Optimized for speed and cost-efficiency, yet it: – Beats o3-mini in math, coding, and even non-STEM tasks
– Scores 93.4% on AIME 2024 and 2025 best of any mini model
– Supports higher usage, ideal for high-throughput environments Smart. -
OpenAI o3 outperforms all models on math, code, and reasoning
By
–
o3 = OpenAI’s smartest model yet It outperforms everything on real-world tasks: – 91.6% on AIME 2024 math
– 2706 ELO on Codeforces
– 86.8% on multimodal MMMU benchmark
– 26.6% on GPQA “Humanity’s Last Exam” It doesn’t just answer it thinks, reasons, and adapts. -
OpenAI’s O3 and O4-mini models think, reason, and act with tools
By
–
🚨 Breaking: OpenAI just dropped its smartest models yet.
— God of Prompt (@godofprompt) 17 avril 2025
O3 and O4-mini don’t just answer they think, reason, and act.
They now choose how and when to use tools like code, search, and images.
Here's why this changes everything about what ChatGPT can do: pic.twitter.com/62pNUcwrsPBreaking: OpenAI just dropped its smartest models yet. O3 and O4-mini don’t just answer they think, reason, and act. They now choose how and when to use tools like code, search, and images. Here's why this changes everything about what ChatGPT can do:
-
Large Robot Action Models and Engineering at CRAY Colloquium
By
–
Looking forward to presenting at the CRAY Colloquium on Monday: "Is Data All You Need?: Large Robot Action Models and Good Old Fashioned Engineering" https://
cse.umn.edu/cs/events/cray
-colloquium-data-all-you-need-large-robot-action-models-and-good-old-fashioned
… -
PyTorch Hiring: GPU Optimization Engineers Wanted
By
–
If GPU optimization and systems problems excite you, why limit your impact to a single company or lab?
Working on PyTorch allows you to ship impact to the entire AI industry!
We're hiring across experiences — junior and senior engineers. Read more below -
Compute scaling reveals fundamental laws of intelligence
By
–
Something magic about these charts with compute on the x-axis, and clear steady improvements (across many metrics) on the y-axis. Really makes you feel like the field is uncovering some underlying fundamental laws of intelligence.
-
OpenAI o3 o4-mini Benchmarks SOTA Performance Efficiency
By
–
Benchmarks and costs for o3 and o4-mini below. OpenAI said o3 hits SOTA performance across benchmarks for coding, real-world software tasks, and multimodality (MMMU). Some major efficiency leaps for both models compared to performance.