OpenAI tries to trademark 'GPT'. US patent office says nope
Plus: How you can set up your own AI chatbot on your device, and more
AI in brief The US Patent and Trademark Office has rejected OpenAI's request to trademark "GPT," which stands for Generative Pre-trained Transformer, the architecture powering its large language models.
OpenAI filed to trademark GPT last year, but trademark attorneys didn't approve its application on the grounds that the term was "merely descriptive." In other words, GPT only describes a function or feature of something rather than being a unique phrase for a commercial product or service.
Officials pointed out that GPT describes a family of neural networks, whose architecture is used in lots of AI models built by other companies. Its common use means that people will understand that it's a term used to describe generative AI technologies. OpenAI, however, reportedly disagreed and said that people wouldn't understand what "Generative Pre-trained Transformer" means.
The USPTO wasn't convinced, and ruled in its final action that the "applied-for mark appears to be generic in connection with the identified goods and/or services. A generic mark, being the 'ultimate in descriptiveness,' cannot acquire distinctiveness and thus is not entitled to registration…under any circumstances," it wrote [PDF].
Don't say anything to Gemini that you don't want people to read
Google will keep all conversations you have with its AI chatbot, previously named Bard and now known as Gemini, for three years – even if you delete its app.
"Conversations that have been reviewed or annotated by human reviewers (and related data like your language, device type, location info, or feedback) are not deleted when you delete your Gemini Apps activity because they are kept separately and are not connected to your Google Account. Instead, they are retained for up to three years," Google confirmed in its latest update of Gemini's privacy policy.
Users should be wary about what they say to the chatbot. Google advised people to not share any personal information that they wouldn't want reviewers reading, nor any trade secrets that could go into training future models. The rule seems to apply to conversations that have been inspected.
It's not clear, however, whether you can tell if someone has peeked at your exchanges with Gemini, as Gizmodo pointed out. Still, it might be better to be safe than sorry.
- US patents boss cannot stress enough that inventors must be human, not AI
- Nvidia reckons this itty-bitty workstation GPU won't run up your power bill
- Google debuts Gemini 1.5 Pro model in challenge to rivals
If you have a decent GPU, here's how you can run your own custom chatbot
You can now run an open source AI chatbot personalized on data stored on your devices, thanks to Nvidia's free "Chat with RTX" software.
However, it requires a Windows 10 or 11 system and a GPU as powerful as the GeForce RTX 30 Series with 8GB of video random access memory. "Chat with RTX" packages together tools like retrieval-augmented generation and Nvidia TensorRT-LLM software to connect large language models like Mistral or Llama 2 to your own PC files.
Why might you want to do this, you ask? Well, if you have a lot of content you want to search for, an AI chatbot customized on your data could come in handy. Information stored in txt, .pdf, .doc/.docx and .xml formats can be processed. You can also add other types of content like the URLs of YouTube videos and query the chatbot to find things like travel recommendations from your favorite vloggers or the latest tutorials.
The system runs locally, so your custom chatbot shouldn't be too slow in replying to you. Installing "Chat with RTX" is a bit tricky. Nvidia warns that the software must go into the default installation directory.
Welcome to the future, where everything looks the same
Following the release of OpenAI's text-to-video model, Sora, here's an interesting thread on X that shows just how generative AI models are given the same input prompt.
You can see side-by-side comparisons for Sora vs Midjourney's text-to-image software here.
Does it mean that all these companies have essentially trained their models on the same data? ®