Nanome’s Post

OpenAI o1-preview A new series of reasoning models for solving hard problems. OpenAI o1-mini The o1 series excels at accurately generating and debugging complex code. To offer a more efficient solution for developers, we’re also releasing OpenAI o1-mini, a faster, cheaper reasoning model that is particularly effective at coding. As a smaller model, o1-mini is 80% cheaper than o1-preview, making it a powerful, cost-effective model for applications that require reasoning but not broad world knowledge. Read more:

Do you want to have your own Fine-tunned OpenAI LLM model ? Now you have chance to train your own model from foundational LLM model for free until September 23rd. Take a look at OpenAI documentation to get insights how it could done. What are the benefits? 🔍𝐄𝐧𝐡𝐚𝐧𝐜𝐞𝐝 𝐏𝐞𝐫𝐟𝐨𝐫𝐦𝐚𝐧𝐜𝐞: Fine-tuning allows you to tailor GPT-4o mini to your specific needs, ensuring more accurate and relevant outputs for your applications. 💡 𝐂𝐨𝐬𝐭-𝐄𝐟𝐟𝐢𝐜𝐢𝐞𝐧𝐜𝐲: Compared to GPT-3.5 Turbo, GPT-4o mini is more cost-efficient, offering longer context and lower latency. This means you can do more with less. 🌐 𝐀𝐜𝐜𝐞𝐬𝐬𝐢𝐛𝐢𝐥𝐢𝐭𝐲: Currently available for Tier 4 and 5 users, OpenAI plans to extend this offer to all tiers gradually. Each organization gets 2M training tokens per day for free, providing ample resources to experiment and perfect your models.

Check out this insightful guide from OpenAI on optimizing results with large language models. It offers valuable strategies and tactics to enhance your prompt engineering skills. https://lnkd.in/eKHVvYbY

OpenAI documentation helps to learn about openAI (GenAI )

New OpenAI model "gpt-4o-min" is out https://lnkd.in/g_MzPUJS. Smarter, faster and cheaper alternative to GPT-3.5-turbo. Also, multi-modal (vision + text). I love those releases. You get a better product, that is cheaper and faster! I am validating the model against my use-cases, so can't say anything about the performance myself, official data can be found by the link above.

Another nice update from OpenAI released only as small news: predicted outputs. With this one you can provide a reference string, and AI only has to inference the differences to the given string. Benefits: speed and costs. This makes at least code generation much faster when the existing code is given as reference. I would assume the same concerns JSON-examples, so AI would not need to inference the keys and other boilerplate in the json? Have to try this out. https://lnkd.in/dNgNHMid

I'm having fun pushing the new OpenAI o3 model on analysis and subtle edits for a new AI agents things I'm working on and it's delivering! I like the 200,000 input 100,000 token outputs very much, especially given one of my files is pretty long. So far, it seems to be doing a stellar job of keeping straight over 100 specific requirements and constraints in a consistent and correct manner through the entire project. I know, I know, the model just dropped a few minutes ago, but is anybody else pushing it on hard stuff that the other top models have been struggling with? How is it going? https://lnkd.in/gX-fx6C7

I found a bunch of discrepancies in OpenAI's vision API's tokenization and pricing. 1. Given its algo for computing pricing here - https://lnkd.in/gFN_sY8x, no image can have number of tiles fewer than 4 (since the short side is scaled to 768px) and yet on its pricing page (https://lnkd.in/g-E9kGrC), it says that the # of tiles would be 1 for a 5x8 px image. Yet, it charges for 2 tiles when an actual API call is made. 2. At the same time, the pricing page itself displays 2 different costs for gpt-4o vs gpt-4o-mini contrary to what was claimed here - https://lnkd.in/gjUXQ4X9. 3. Also, the Vision API page doesn't clarify that the # of tokens is counted differently for 4o vs 4o-mini. It is hidden deep in the pricing page. Anyone else?

Around two weeks ago, I designed and launched Hameh on Honarestan, a personal AI assistant using OpenAI's beta stage assistants and threads features. With threads, OpenAI takes care of memory management, meaning no more need to create DIY code to handle it. Their assistant can also follow instructions, provide code, generate graphs, and accept various types of inputs. It can even be given files or run functions and code. This is honestly a big step forward, and I was very impressed with the features. If you're setting up your own GPT-based chatbot, I strongly recommend using the threads beta feature and letting OpenAI handle the memory for you. You can read more here:

The new OpenaAI o1 models have implemented a new type of token called Reasoning tokens. I would read through the documentation on it: https://lnkd.in/eVMRwgz5 Apparently, these Research tokens are used internally for GPT to reason. But they are also part of the output context window. So you need to carefully monitor your output tokens (it is explained in the docs how to do this ) because if you don't, your token limits could be reached by GPT needing to perform more in depth reasoning and thus you might not receive any output response. 😱 💲 💲 💲