Hello everyone,
I’m excited to share that OpenAI has released a new model that is allegedly better at reasoning. What’s your take on that?
Here is the Summary of OpenAI o1-preview announcement:
7 Likes
I was eager to know more. I gained access about an hour after the announcement, even though my account is only Tier 1. I was a bit confused, but it’s cool.
6 Likes
It’s frustrating that they keep boasting about how it’s going to be even stricter with its ethical guidelines. It’s already annoyingly difficult to use for writing prompts since its grasp of human relationships is so dull that even 90-year-old Mormons would find it boring. Just give me some dialogue lines—I don’t need a lesson on micro-aggressions!
5 Likes
That line of thinking is pretty wild. It looks like OpenAI is delivering on the Reflection model they teased on Twitter last week, lol.
However, I wonder if these models could improve further if their reasoning happened internally, rather than outputting reasoning steps in natural language. From what I’ve seen with superhuman-level AIs in specific areas, their reasoning is only partly understandable. AlphaGo can give the probability of winning for each move, but how it evaluates the board to get that figure exists entirely within the network and is not interpretable.
4 Likes
I fully support your point—AlphaGo needs an underlying computation layer and a separate language model as the communication layer. While this could help, it’s not currently feasible since there’s no general method to translate between natural language and the specific abstractions used in computation. OpenAI’s approach relies on large-scale, manual text curation to address this, which is inefficient and unsustainable. It’s like if AlphaGo had been built by people manually annotating numerous Go games—a crazy concept when you think about it that way.
3 Likes
I’m happy for them, but I didn’t find much info on the new model, just a few vague paragraphs about reinforcement learning and some impressive metrics. They seem pretty confident about it.
3 Likes
@Joy. Just take a look at how it handles cost optimization.
2 Likes
I don’t agree. There is a lot of research showing that LLMs can’t reason, so OpenAI’s claim is misleading.
1 Like