👩💻 OpenAI announced a new "smart" model - o1
OpenAI Submitted by a new model called "o1", oriented to solve complex mathematical problems. The trick of the model is that it is able to think before answering, forming long chains of reasoning, which is especially important when performing tasks that require in-depth analysis. It outperforms GPT-4o in many ways in the context of logic and complex problem solving.
🤯 Mathematics Competition (AIME 2024)
- GPT-4o solves about 13,4% of the tasks, which is a rather low score.
- o1-review shows improvement solver 56,7% tasks.
- o1 significantly outperforms both models with resulting in 83.3% accuracy.
🤯Programming (Codeforces)
- GPT-4o scores only 11.0% in programming competitions.
- o1-review shows significant improvement with result of 62%.
- o1 shows the result – 89%.
🤯Doctoral Level Questions in the Natural Sciences (GPQA Diamond)
- GPT-4o reaches 56.1% accuracy with problem solving
- o1-review increases the accuracy to 78.3% and the o1 version to 78.0%.
- Hired human experts with PhD average 69,7%
Conclusions: o1 outperforms GPT-4o in everything tasks presented - be it mathematics, programming or natural sciences.
🤖 Basic The innovative killer feature of "o1" is the use of the chain of thought. The model can "think" in front of itself how to respond, making her approach to tasks more consistent and structured. o1 learns to break down tasks to simpler steps, correcting mistakes and changing strategies if necessary. In the case of GPT-4o, there is no such approach.
🤖 Altman writes that this is the beginning of a new paradigm: AI is now able to perform complex general reasoning meetings. 1-preview and o1-mini are available today on ChatGPT for plus and team users.
Given that experts with PhDs have already shown themselves to be worse than o1 in 2024, Leopold's projections to 2030 seem quite realistic.
👆 Meanwhile China:The best neural network for math has been released by the Chinese - their Qwen-2.5 Math solves problems on par with the OpenAI o1 hype !- The model works on the same reasoning method as o1 and the ability to execute Python code has been added - multiplication and division are no longer a problem (see screenshot);- They have released three versions at once: for parameters 1.5b, 7b and 72b. The first two will work even on the weakest computers;- It works for FREE.You can use Homework Killer online or install it on your computer locally.