"Vision" is the current bottleneck imo
热度趋势
趋势数据积累中
百分比基于当前可用热度信号，而非评论数或独立用户人数。
这条记录涉及编程工具或代码能力更新，适合开发者评估工作流变化和可复用价值。
目前人工智能发展中的瓶颈在于“视觉”能力，尤其是在处理涉及视觉元素的任务时。尽管人工智能在某些领域表现出色，但在视觉评估方面却面临挑战，这类似于盲人难以创作一款以视觉元素为重点的视频游戏。人工智能处理视频的方式，通常每秒仅处理1-5张图像，这与人类的视觉感知方式大相径庭。…
Blind people are obviously just as capable in some areas, but ask them to make a Minecraft clone (or an original video game with a focus on visual elements) and they'll really fucking struggle, not with the coding, but with the loop of assessing what they've done so far.
You just can't unit test everything in advance reliably, sometimes you have to load in the game and notice the sword is rotated incorrectly in the player's hand.
With one person with vision on their side they can do it, and that's basically the pairing a lot of vibe coders amount to, you are the set of eyes for the LLM, still a small amount of providing "common sense", taste, opinions, goal alignment, etc. but in terms of technical capability it's obviously very far along and imo not the bottleneck.
Memory is an issue too. "Dexterity" is ofc an issue with those trying to make robot irl workers. There are a whole host of bottlenecks but to me vision feels like a big one, to instantly notice issues, yeah if you ask ChatGPT what's wrong if you show it a picture of a player with a sword held blade first it'd probably notice, but can it notice it unprompted for a video? Could it notice if the sword was only rotated incorrectly in the Z axis such that the flat side was being "used" for swings?
Needs to be more efficient and more intuitive, rather than relying on the reasoning power they've cultivated imo.
Especially for video, norm for video is to give 1-5 images per second depending on the LLM, price, plan, company, etc. but that's not how people see videos. And it's not even how we see images, tokenisation of text probably isn't TOO far from how the brain does it, but for images it's probably WAY off.