I'd love a blog or coffee table book of "where are they now" for the director level folks who do dumb shit like this.
It said: OpenAI’s leading researchers have not completed a successful full-scale pre-training run that was broadly deployed for a new frontier model since GPT-4o in May 2024, highlighting the significant technical hurdle that Google’s TPU fleet has managed to overcome.
However, pre-training run is the initial, from-scratch training of the base model. You say they only added routing and prompts, but that's not what the original article says. They most likely still have done a lot of fine tuning, RLHF, alignment and tool calling improvements. All that stuff is training too. And it is totally fine, just look at the great results they got with Codex-high.
If you got actually got what you said from a different source, please link it. I would like to read it. If you just messed things up, that's fine too.
[1] https://newsletter.semianalysis.com/p/tpuv7-google-takes-a-s...
No one knows whether the base model has changed, but 4o was not a base model, and neither is 5.x. Although I would be kind of surprised if the base model hadn't also changed, FWIW: they've significantly advanced their synthetic data generation pipeline (as made obvious via their gpt-oss-120b release, which allegedly was entirely generated from their synthetic data pipelines), which is a little silly if they're not using it to augment pretraining/midtraining for the models they actually make money from. But either way, 5.x isn't just a prompt chain and routing on top of 4o.
I’m sure all these AI labs have extensive data gathering, cleanup, and validation processes for new data they train the model on.
Or at least I hope they don’t just download the current state of the web on the day they need to start training the new model and cross their fingers.
At the very least they made GPT 4.5, which was pretty clearly trained from scratch. It was possibly what they wanted GPT-5 to be but they made a wrong scaling prediction, people simply weren't ready to pay that much money.
Their investors surely do (absent outrageous fraud).
> For all we know, they could be accumulating capital to weather an AI winter.
If they were, their investors would be freaking out (or complicit in the resulting fraud). This seems unlikely. In point of fact it seems like they're playing commodities market-cornering games[1] with their excess cash, which implies strongly that they know how to spend it even if they don't have anything useful to spend it on.
[1] Again c.f. fraud
Right, this is nonsense. Even if investors wanted to be complicit in fraud, it's an insane investment. "Give us money so we can survive the AI winter" is a pitch you might try with the government, but a profit-motivated investor will... probably not actually laugh in your face, but tell you they'll call you and laugh about you later.
I know sama says they aren’t trying to train new models, but he’s also a known liar and would definitely try to spin systemic failure.
Doubtful. This would be the very antithesis of the Silicon Valley way.
This isn't really accurate.
Firstly, GPT4.5 was a new training run, and it is unclear how many other failed training runs they did.
Secondly "all subsequent models are routing systems and prompt chains built on top of 4" is completely wrong. The models after gpt4o were all post-trained differently using reinforcement learning. That is a substantial expense.
Finally, it seems like GPT5.2 is a new training run - or at least the training cut off date is different. Even if they didn't do a full run it must have been a very large run.
https://www.theinformation.com/articles/openai-says-business...
https://epoch.ai/blog/training-compute-of-frontier-ai-models...
For all we know, they could be accumulating capital to weather an AI winter.
It's also worth noting that OpenAI has not trained a new model since gpt4o (all subsequent models are routing systems and prompt chains built on top of 4), so the idea of OpenAI being stuck in some kind of runaway training expense is not real.