开发者对 DevDay 的 Astra 愿望清单
四条诉求：持续学习、实时 computer use、更持久的 Astra agent、以及自动研究助手的量化数据。
DevDay wish list 🎁 1. Continuous learning research. I want to hear about any serious effort toward continuous learning, ideally parameterized learning, fast weights, adapters, or some other form of persistent model updating rather than just increasingly sophisticated context compaction. Even an early research direction here would probably be the thing I care about most. 2. Much faster computer use. I’d like computer use to move closer to a continuously perceived desktop rather than the current screenshot -> reason -> act loop. macOS already lets ChatGPT’s Work with Apps feature query compatible apps through the Accessibility API, while OpenAI’s computer-using agent has historically operated from screenshots, (barbarically slow on windows) so I’d love to see a hybrid approach using structured accessibility information plus a low-latency visual stream/video encoding to make desktop control feel essentially real time. 3. Hardware trailer or a more persistent Astra agent framework. Any actual hardware reveal would obviously be interesting, but short of that I’d love to see Astra become substantially more persistent agents that can remain active across longer periods, maintain state, and continue working without every task feeling like a new isolated run. 4. Actual numbers on the auto research assistant. I want OpenAI to quantify how much of the next-model R&D pipeline is now being completed by AI experiment generation. Similar to the kind of visibility Anthropic have started giving into AI-assisted model R&D. OpenAI has already published early data on agents accelerating research, but I’d especially like to know what fraction of pipeline delivery is autonomous and if it’s toward specific research goals I’d like to see what those are and if I (again) hope for any of that effort to directed toward problems like continual learning. 5 Higher compute limits. This one is cheesy but giving advanced users more access to the highest reasoning settings. 6. Access to the next step beyond Astra. This is the least likely one, but I know OpenAI already has a substantially stronger internal successors to Astra, I’d love to see some version of that next compute tier exposed publicly. I understand that economics, safety, and competitive positioning all affect when that happens, but.. Letting users touch the next SOTA model would be an incredible DevDay drop.