OpenAI has announced o1, a new language model that, compared to the company’s GPT models, takes more time to answer and 'thinks more deeply'. OpenAI remains vague about how this works exactly. The model had the codename Strawberry and is called o1, with a smaller version named o1-mini, the company reports. Paying customers already have access or will get it soon, while free users will gain access to the smaller version in the future. The model is not as versatile as the current GPT-4o; it cannot upload files, generate images, or browse the web. The new o1 model distinguishes itself in reasoning. It first produces a long 'chain of thought' before answering, OpenAI explains. Since the model tries to arrive at answers in multiple ways, it is said to make fewer mistakes. This should also reduce hallucinations, although hallucinations are still possible with the new model. OpenAI chose not to use the abbreviation GPT for this model. The "1" in its name indicates that the reasoning makes it a different type of model: it is the first version of a new series. Compared to GPT-4o, the model scores better on various tests and is more resistant to jailbreaking. However, using it via the API is 3-4 times more expensive than using GPT-4o.

On Wednesday, October 2, 2024, OpenAI will update the default version of GPT-4o to the latest model, GPT-4o-2024-08-06. This new version introduces several benefits, including a 50% reduction in input token costs and a 33% reduction in output token costs. Additionally, it supports structured outputs via JSON, a feature that is especially useful for developers needing precise and predictable output formats. Developers already using the GPT-4o API will be automatically switched to the new model starting October 2, unless they specifically opt for the older version, GPT-4o-2024-05-13. This upgrade not only lowers operational costs but also improves the efficiency and reliability of the model, especially for applications requiring fast, real-time text responses. For those interested in exploring the new model before the update, it is already possible to specify the parameter gpt-4o-2024-08-06 in the API. https://platform.openai.com/docs/models/gpt-4o

Hi all - not sure if this is a bug, common problem or I'm doing something wrong. But I have created a custom GPT to help with my job, uploaded instructions on how to respond to specific prompts. This works GREAT... Until it doesn't. GPT seems to erase its memory of my instructions for the prompt. One of the prompts, for example, is "Company Research" + [company name] This then outputs research on a company in a specific format This works perfectly when I am in editing mode for the GPT. I then save & update the GPT, exit it and it then works perfectly for a while after this. But somewhere along the line it randomly forgets the instructions for the prompt and this is driving me insane. Can anyone help with this?

OpenAI is launching GPT-4o Mini, a smaller version of its GPT-4o model released earlier this year. For now, the smaller AI model only supports text. Support for images, video, and audio will follow later. OpenAI has not released details itself; the American broadcaster CNBC reports on the arrival of the new AI model. GPT-4o Mini should be available to users of the free version of ChatGPT and paying customers starting Thursday. At the time of writing, the Mini version is not yet available. GPT-4o Mini will replace GPT 3.5 Turbo.

Meta, formerly known as Facebook, has released Llama 3.1. The latest version of its large language model has up to 450 billion parameters, although there are also smaller variants. It supports eight languages. Llama 3.1 supports not only English but also French, German, Spanish, Italian, Portuguese, Thai, and Hindi, as shown on the HuggingFace page. It is an open-source model that anyone can freely download and use. However, users must share some data via HuggingFace before they can download the model, such as username and email address. The model can process up to 128,000 tokens, compared to 8,000 for Llama 3. The model with 450 billion parameters is new, but there are also updates for the models with 70 and 8 billion parameters, according to Meta. Llama 3 was released a few months ago. Llama 3.1 is expected to perform similarly to GPT-4 and Claude 3.5 Sonnet, as claimed by Meta based on benchmarks and human evaluations. The company is also working on smaller models to run more easily on devices. Additionally, Llama 3.1 currently only accepts text and files as input and output, but other forms, such as images, are expected to be added.