# 疑似GPT-6？OpenAI总裁谈新模型Spud

*机器之心* · https://www.youtube.com/watch?v=1C1CozmFq34

## Summary

OpenAI总裁Greg Brockman在访谈中回应了关于新模型Spud的传闻。他解释说Spud是一个新的预训练基础模型，承载了约两年的研究成果，但强调OpenAI的进步是一个不断加速的引擎，任何单一模型都只是其中一步。他认为Spud将能解决更难的问题，更好地理解指令和上下文，既提高能力上限也提高使用下限。

## Key points

- Spud已完成预训练，Sam Altman曾告诉员工几周内会有一个非常强大的模型。
- Brockman认为Spud不是重点，它只是新的预训练基础模型，后续还有强化学习和后训练过程。
- 过去18个月他主要精力放在GPU基础设施和支持大规模训练运行的团队上。
- Spud承载了大约两年的研究成果，即将开花结果。
- 新模型将能解决更难、更开放的问题，理解指令和上下文的能力更强。
- 他提到所谓的大模型气味：模型真正更聪明时，用户能明显感觉到它更顺从、更懂你。
- 他期待模型既提高上限（如物理应用、更长时间跨度的问题），也提高下限（日常任务更有用）。
- 每次发布都只是后续更多改进的早期版本，进步引擎会越来越快。
- 团队相信这个模型可以真正加速经济发展，进展比许多人预期的更快。

## Notes

### Spud是什么
- Spud是OpenAI的新模型，据The Information报道已完成预训练，Sam Altman曾告诉员工几周内会有一个非常强大的模型。
- Brockman强调Spud本身不是重点，它是一个新的预训练基础模型，是后续改进的起点。
- 这个模型承载了大约两年的研究成果，即将开花结果。

### OpenAI的开发流程
- 预训练：产生新的基础模型，是公司许多人共同努力的巨大工程。
- 强化学习：让已学到世界知识的AI应用这些知识。
- 后训练：让模型在各种情境中练习解决问题，完成行为和可用性的最后一公里。
- Brockman过去18个月主要精力放在GPU基础设施和支持训练框架团队的大规模运行上。

### Spud能做什么
- 能解决更难的问题，更加细致入微，更好地理解指令和上下文。
- 他提到大模型气味：当模型真正更聪明、更有能力时，它们更顺从你，你能感觉到。
- 当AI没听懂问题需要你解释时，总是令人失望；新模型会减少这种挫败感。
- 量化上会有很多提升，质上会出现新事物：以前不用AI的任务，现在会不假思索地使用AI。

### 提高上限与下限
- 提高上限：已经看到物理等领域的应用，新模型能解决更开放的问题、更长时间跨度的任务。
- 提高下限：对任何想做的事情，模型都变得更有用。
- 团队相信这个模型可以真正加速经济，进展比许多人预期的更快。
- 每次发布都只是后续版本的早期形态，OpenAI的进步引擎会越来越快。

## Transcript

[00:00] foreshadowed a couple times already that you have some good models on the way. What is Spud? Uh the information said uh that you finished uh pre-training Spud, and Sam Altman, the CEO of OpenAI, has told the staff that uh they should expect to have a very strong model in a few weeks. This was a few weeks ago. And the team believes it can really accelerate the economy, and things are moving faster than many of us expect expected. So, what's Spud? It's a good model. Um but I think that it's really not about any one model. Okay, right? The the the way that our development process works is you have pre-training, so you produce a new base model that then is the foundation that we build further improvements on top of. And that

[00:45] is always a huge effort across many people in the company, and that's where I've actually been spending most of my efforts over the past 18 months has been really focused on our GPU infrastructure, on supporting the teams that do all of the training frameworks to scale up at these big runs. But then there's a reinforcement learning process. So, you take this AI that has learned lots of things about the world, and it applies that knowledge. And then we do a post-training process where you really say, "Okay, now you know how to solve problems. You practice it in in all these different contexts." And then here's kind of the last mile of of behavior and usability. So, I think of Spud as a new base, as a new pre-train,

[01:22] and that we have had this uh I'd say it's like we have maybe 2 years worth of research that is coming to fruition in this model. It's going to be very exciting. And I think that the way that the world will experience it is just improved capabilities, and that for me it's never about any one release, because as soon as we have this one release, it'll be an early version of what we have coming. We'll do much more of each of these steps of the improvement process. And so, I think that where we're going is is almost just we have this engine of progress that just moves faster and faster, and that Spud is just one step along the way. So, what do you think it'll be able to do that today's models can't? I So, I think

[02:04] it's going to be able to solve both much harder problems. I think it will be much more nuanced. It'll understand instructions better. It'll understand the context much better. That there's this thing called big model smell that people talk about, where it's just like there's something about like when these models are just actually just much smarter, much more capable, that they bend to you much more, and you feel it. Right? When you ask a question and the AI doesn't quite get it, it's always so disappointing. Right? When you have to like explain it, you're like, "You really should be able to figure this out." And so, I would just think of it as in some ways just qualitatively there will be like quantitatively lots of shifts,

[02:44] right? And qualitatively there will just be new things where you would be frustrated before you never use an AI for it, and now you just use it without without thinking very much. And I think that that that is what we're going to see across the board. I'm super excited to see how it raises the ceiling, right? We've already seen these physics applications, things like that. And I think we will be able to just solve like way more open-ended problems, way longer time horizons. And then also very excited to see how it raises the floor, where just for anything you want to do, it's just so much useful more useful for you.