Sam Altman says OpenAI won't train GPT-5 for “some time” and the letter to pause AI development missed “most technical nuance about where we need the pause”
This is in line with my current belief that GPT-4 is pretty close to where LLMs will peak. … Tweets: Brij Singh / @brij : Capabilities over versions! Moving forward, @OpenAI may shift focus away from model versions to avoid the “AGI - are we there yet?” question. Just like no one knows Google's search version number https://twitter.com/... Yann LeCun / @ylecun : Before we can get to “God-like AI” we'll need to get through “Dog-like AI”. Brian Merchant / @bcmerchant : I would add that making a statement like ‘we are not training GPT-5’ in the context OpenAI has helped nurture—that doing so would risk meddling w forces too powerful to bear—only serves to underscore that *it* is the company with the capability to do so (so maybe check out GPT4) https://twitter.com/... Paul Kedrosky / @pkedrosky : Pro tip: You're not reassuring people properly if you tell them that you are already doing the things they are worried about, but not calling it what they thought you were calling it. OpenAI's CEO confirms not training GPT-5 “for some time” https://www.theverge.com/... James Vincent / @jjvincent : Sam Altman has confirmed OpenAI is not training GPT-5 and 'won't for some time.' if you're worried about AI safety, though, I don't think this is a meaningful statement. i wrote a little about why, and the AI world's treatment of vague metrics: https://www.theverge.com/... https://twitter.com/... Grady Booch / @grady_booch : “We are doing other things on top of GPT-4 that I think have all sorts of safety issues that are important to address and were totally left out of the letter.” What things, we ask? So much for “open”... https://www.theverge.com/... Dr. Marie Haynes / @marie_haynes : There's still a lot to learn about GPT-4 before OpenAI moves on to GPT-5. https://www.theverge.com/... Adam Singer / @adamsinger : I actually do appreciate the AI critics, and their caution should be taken seriously. But if I were Sam or working at OpenAI I would still continue to build and do so as responsibly/ethically as humanly possible. The AI bears will *never* be satiated. Progress must proceed... https://twitter.com/... Shakeel / @shakeelhashim : “I think the focus in these discussions should be on capabilities: on demonstrations of what these systems can and can't do and predictions of how this may change over time.” strong strong agree https://twitter.com/... Barry Schwartz / @rustybrick : Self pause on AI? https://twitter.com/...
Context & Ripple Effects
The story lands amid an unsettled response to the proposed training pause: some signatories and AI experts disputed the letter's proposal, while discussion of GPT-4's capabilities had already elevated the stakes around frontier-model development.
OpenAI's position distinguishes a pause on a named successor model from safety work across the broader development stack. That distinction remained consequential as later coverage tied a delayed GPT-5 timetable to compute-allocation constraints and documented continued releases beyond GPT-4.
First-order effects
- OpenAI publicly lowers expectations for an imminent GPT-5 training run, shifting immediate attention from a version-number milestone to work around GPT-4 and its safety implications.
- The pause-letter debate is recast from a blanket stop request into a dispute over which technical activities should be constrained and how safeguards should be targeted.
Second-order effects
- Competitors and AI customers have less reason to treat the next model label as the sole competitive benchmark; product capability and deployment practices become more salient measures.
- Safety advocates face pressure to specify interventions beyond a training freeze, while labs can argue that work not captured by a model-name pause still requires oversight.
Third-order effects
- If frontier labs increasingly decouple releases from simple version sequencing, governance will need to focus on capabilities, training scale, and deployment conditions rather than model names alone.
- The episode points toward operational AI assurance: durable oversight may depend on testing and controls across development and use, not one-off voluntary pauses.
The trend: Frontier-AI governance is moving from headline model-release moratoriums toward capability- and operations-based controls, even as labs continue to advance their systems.