A useful post summarizes the news about OpenAI's latest advanced model, o1:
🔹 The improvement in quality stems from the model's ability to reason before providing an answer. While the reasoning process itself won't be shown, there will be a brief summary with a high-level overview.
🔹 Previous models could reason as well, but with less effectiveness. OpenAI has focused on enhancing the model's ability to arrive at the correct answer more frequently through iterative self-correction and reasoning.
🔹 o1 is not intended to replace gpt-4o for all tasks. It excels in math, physics, and programming, follows instructions more accurately, but may struggle with language proficiency and has a narrower knowledge base. The model should be viewed as a reasoner (akin to "thinker" in Russian). According to OpenAI, the mini version is comparable to gpt-4o-mini, with no major surprises.
🔹 The model is currently available to all paid ChatGPT Plus subscribers, but with strict limits: 30 messages per week for the large model and 50 for the mini version. So, plan your requests carefully!
🔹 If you have frequently used the API and spent over $1,000 in the past, you can access the model via API with a limit of 20 requests per minute.
🔹 However, costs are high: the junior version of o1-mini is slightly more expensive than the August gpt-4o, but you’re paying for reasoning (which you won’t see) that will be substantial. Thus, the actual markup could range from 3 to 10 times, depending on the model's "thinking" time.
🔹 The model handles Olympiad-level mathematics and programming problems with the skill of international gold medalists, and for complex physics tasks resistant to Google searches, it performs at a PhD student level (~75-80% correct).
🔹 Currently, the model cannot use images, search the internet, or run code, but these features will be added soon.
🔹 The context for models is still limited to 128k tokens, similar to older versions. However, an increase is anticipated in the future, as OpenAI claims the model currently "thinks" for a couple of minutes at a time, with aspirations for longer durations.
🔹 As with any initial release, there may be some simple bugs where the model fails to respond to obvious prompts or leads to jailbreaks. This is normal, and such issues should decrease in 2-3 months once the model transitions from preview status.
🔹 OpenAI already possesses a non-preview version of the model, which is currently being tested and is reportedly better than the current release—see the attached image for details.
🔹 The new model operates without needing prompts; you won’t have to ask it to respond in a thoughtful, step-by-step manner, as this will be handled automatically in the background.
Welcome to Strawberry Era! 🔥
Source |
Artificial intelligence 🤖
عرض المزيد ...