Kevin Weil’s Post

People don't realize o1 from OpenAI isn't even smart enough to tell you how to properly use ITS OWN API. I wish I was joking. This is the problem with probability models, they learn distributions then you need them to forget them once you update things. (Screenshot is from my own CLI tool (named anton) for interacting with o1 via API calls).

Whisper V3 Turbo model launched! Model distillation is happening not only in LLM space but all AI models. As large models get smarter, smaller models are also getting this advantage through distillation! #generativeai #whisper #openai

Another moment of OpenAI being O̶p̶e̶n̶A̶I̶ ClosedAI. A lot has been happening since the model o1 came out. Many are trying to understand the ability behind such a model's superb performance by tricking o1 into revealing its thinking process. Most likely they have Reflection baked in with some complex CoT, which iteratively improves the response of the LLM. Although this is great for OpenAI as they are the first to achieve this level of success, thankfully these simple techniques are also being replicated by other open-source competitors too. Such simple techniques should be openly shared with the public instead, OpenAI has started sending Violation notices to the developers who are trying to reveal the "special prompts".

With all this excitement around OpenAI's o1 reasoning model, its important t remember that this has been around in open source for a long time : this is an implementation of Langchain's "self ask ask agent" with web search surfaced in a Langflow workflow. And the great thing about doing it this way is that it doesn't hide the intermediate steps like the OpenAI implementation, so you can still tune and debug. Along with the web search, we can also use this process with documents and data, which isn't available yet from OpenAI

💥 OpenAI's o1-preview and o1-mini are now live on Dify v0.8.1 Discover these powerful new models from OpenAI, designed for complex reasoning tasks and built to tackle problems more intelligently. We've also fixed some bugs you might have encountered with Workflow Parallel processing to enhance your experience. Give it a try and check out the full changelog here: https://lnkd.in/gBeSzkGW

OpenAI Assistants! Again, probably not using them for what they were designed but they are extremely powerful. Especially the ability to specify a JSON Schema for the response. Our accuracy and conistancy in the results structure just went up tenfold and helped us reduce the complexity in our own code. https://lnkd.in/eC9eeVZg

What would be the best way to leverage this access? If you’ve been wanting to test the new OpenAI API - reach out and let’s chat on how we can leverage this pilot access together. Jeff Shi

Qwen-2-72B tops Hugging Face’s new Open LLM Leaderboard! Here’s the code to use Qwen-2-72B with OpenAI Client using Portkey -

Does anyone know if there are plans for JSON-LD (and perhaps Schema.org) support in Semantic Kernel and OpenAI? How do folks currently work with these together? I note the JSON Schema support in OpenAI's Structured Outputs - something like that for JSON-LD would be excellent. https://lnkd.in/evKHk24w

have you ever tried to recall what code changes you made in 5+ files, to write a commit message? install `npm i -g opencommit` and simply run `oco` it generates commit messages with openAI or any other known llm provider, even ollama running locally. emojis are optional