OpenAI dropped a new model 🤯 but there's something fishy 👀
But before that, the news👇🏻
Yesterday, OpenAI launched new reasoning models - o1, and o1mini
What the heck is a reasoning model? 😅
1. Before I answer that, you have to understand how large language models work—you ask them a question, and they answer by predicting the next word.
2. So far, the only way we thought models could improve was - to train them on a bigger dataset.
3. But there was another hypothesis—if models could think before answering, analyze their answers, and then respond rather than immediately answer, that could also improve a model’s responses—and this is a reasoning model—a model that can think, analyze, and respond.
4. This is why the o1 and o1 mini take significant time (approx. 10-20 secs) before answering your questions.
Why is this a significant breakthrough? 😱
1. You have to understand that LLMs are designed to be non-deterministic in responses. This means there is a high probability that you might get different answers, if you ask the same question at different times.
2. By design, LLMs are vague - they are not designed to be correct but to be able to answer anything and everything. This can both be a bug and a feature.
3. For cases like coding, maths, etc., where you need more deterministic answers - this is where reasoning models champion.
What are the downsides? 🚩
1. These models will be really expensive.
2. You have to understand that these models' currency is tokens. Input tokens are the words you write, and output tokens are the answers you get.
3. E.g., if you want to get code for a simple website, your input token would be 200 words/tokens, and the output token from a model like GPT4 would be 2000 tokens/words (because it’s code).
4. However, with a reasoning model, the output tokens will be north of 20,000 because it will perform multiple code gen, check, test, regenerate code, etc.
5. Hence, you are spending 10X more on reasoning models vs existing models.
Now, why do I feel there is something fishy? 🐠
1. OpenAI is marketing this as a model that can now think to make it more human—but it’s not. It’s just doing a recursive test of what it generates, testing/predicting what is right or wrong, and then publshes the result. So it’s self-analyzing the first answer, then the second, and then generates the final answer.
2. You can also give any existing models the ability to think (as they say) by writing a system prompt (DM me, and I can share). Hence, it's not a breakthrough in research but just how to prompt the models to perform reasoning.
3. This is not a new model - like it’s being hailed as. It’s still GPT4, which has this new ability to do recursive testing and generate answers.
4. Besides coding, maths, etc., GPT 4o is still the best model in many cases.
What do you think about this new reasoning model?
Comment down below 👇🏻
Follow me Rohan, for more such updates on anything and everything AI ❤️