Today, OpenAI revealed in a press release that "OpenAI has recently begun training its next frontier model". It was commonly believed that GPT-5 had begun training in 2023, shortly after GPT-4 finished. For example, the market for a release this year went from 68% to 50% on the OpenAI press release today, and markets like this one implied a substantial chance of training beginning in 2023 earlier this year.
This market resolves YES if any reputable news source, such as The Information, indicates that OpenAI did in fact begin a training run for GPT-5 previously and abandoned it to start this training run. This could indicate that they ran into some sort of issue or simply that they decided to restart and train a cheaper model using the innovations made for GPT-4o, if those save a significant amount of compute.
"GPT-5" here means "a scaled up version of GPT-4"; if it is referred to by a codename but they mention that it has many more parameters than GPT-4 and is the model that would have been expected to be called GPT-5, I would still resolve YES.
Related questions
https://x.com/kimmonismus/status/1795458986289869310?s=46
https://www.businessinsider.com/openai-launch-better-gpt-5-chatbot-2024-3
Some potential evidence for this theory
@RemNi Right, this market is asking whether they tried to train gpt-5 with the older architecture and gave up on the process. I went and did some research and it doesn't seem like they ever claimed that a gpt-5 run was in progress until now, but it's still possible
@SaviorofPlant Was it really "commonly believed" that they started a training run in 2023? Twitter AI hype accounts aren't a good source of truth. Is there any evidence that this belief was widespread among reputable sources?
I'm not disputing the premise of this market, there could indeed have been a discarded unimodal training run that plateaued early.
@RemNi Not sure about reputable sources, but I think there was a substantial number of people who thought that on this site (see: markets for GPT-5 release date)
@SaviorofPlant you mean like this one? https://manifold.markets/RemNi/will-gpt5-be-released-before-may-20-370feb584056
@RemNi Yeah, 30% for a May release strongly implies a chance that training started in 2023. The market for a release this year went from 68% to 50% on the OpenAI press release today
@SaviorofPlant I think that citing those previous trends on manifold more explicitly in the description would make it a bit clearer in that case.