Yesterday marked a significant day for OpenAI, as they began the full-scale training of GPT-5. I gathered reliable information from various sources, including OpenAI’s public comments and an exclusive interview with a hardware CEO. I’ve also added my analysis. I’ll share a practical tip for using GPT and a discovery I enjoyed. The first clue about GPT-5’s training came from OpenAI’s co-founder, Greg Brockman. OpenAI usually trains smaller models before a full-scale run. They use these to gather insights for the larger model.
Brockman’s recent tweet hints at maximally harnessing computing resources, indicating the training of their biggest model yet. Jason Wei, another OpenAI researcher, tweeted about the excitement of launching massive GPU training. However, this doesn’t mean GPT-5 will be available soon. GPT-4 took about 3 months to train, followed by safety testing.
OpenAI updated their blog, noting the Red Team Network applications are closed. Red teamers are likely starting safety testing on the new model. Even before full training, models undergo various checkpoints. Brockman mentioned the possibility of releasing these checkpoints before GPT-5.
OpenAI might incorporate a feature in GPT-5 allowing it to lay out reasoning steps in solving challenges. This could improve understanding and verification of its processes. Sam Altman of OpenAI hinted at significant improvements in reasoning and reliability in future models. He mentioned multimodality and a focus on audio and visual data integration.
From my interview with Gavin Uberti, CEO of Etched AI, GPT-5 might have about 10 times the parameters of GPT-4. This could result from a larger embedding dimension, more layers, and more ‘experts’ in the model. A larger embedding dimension means more detailed understanding of each data token. More layers allow for deeper pattern recognition.
GPT-5 Release Date
Regarding GPT-5’s release, I predict it might be around November 2024. The training, followed by extensive safety testing, could explain this timeline. OpenAI might avoid releasing it during the contentious U.S. election period. Also, competition from other AI models like Gemini Ultra and Llama 3 might influence their release schedule.
Despite extensive testing, the exact capabilities of GPT-5 remain uncertain until its training is complete. Even OpenAI executives like Greg Brockman and Sam Altman acknowledge the unpredictability of such large-scale AI models.
Finally, a practical tip: GPT-4, and likely GPT-5, can handle typos well. So, don’t worry about correcting minor errors in your queries. This could save time and still yield accurate responses from the AI.
Credits: @AI Explained.
Read related articles: