Reflectoputer

+1  

A program for personal computers that builds compressed model of their user, by observing and compressing data about them, and asking questions to understand the user, to help user automate what the user does, to become like deep mirrors of the user.

YAML 想法

The primary reason why we don't keep camera, audio, screen recorders and key-loggers on all the time, is that, while it is readily possible today, we lack disk space and processing power to compress the data collected, and it requires extra energy use, while we have limited battery life.

However, it is orders of magnitude cheaper to use pre-trained neural nets than it is to train them from scratch, and those models can cheaply compress complex pictures, videos, audios, etc. and store them as sequences of features on very tiny amount of disk space to keep a rich life-log about its user.

The primary reason why computers don't proactively ask us questions, is that they don't have a directive to model us. Reflectoputer would have a goal to build maximally accurate model of your self. So, additionally, -- use psychological research to build a generic human model, which captures most of what humans are like. (Human diversity is arguably quite limited (like the diversity of cats is), or like what MBTI tests show, covering broad set of aspirations with just 16 different core types of personalities with a few abstract dimensions to plot them), and all humans share a large bit of common traits that could be redundant, and so, could be used to compress model about humans (and get computers optimize for searching for something like a "sufficient statistic" for that model). Use the remaining peculiarities of the user's personality that doesn't fit the model -- as something that reflectoputer proactively asks questions in order to fill the gaps in the parametrization (or fitting) of human model (with user data). For example, certain questions from personality tests devised by psychologists can be quite useful to fill those gaps.

Then, observing repetitive works done by user, try to simulate inputs based on the model to gain confidence in automating user actions. If the sequence of actions produced by user coincides closely with the sequence predicted by the model built from past data, automatically ask user to permit automation of the user. (E.g., when confidence level exceeds certain threshold.) Do all that with local compute and data, without ever using cloud services.

This way, personal reflectoputer builds model of the user self by observing, recommending and asking, as well as helping user to build their life story.

Reflectoputers would be useful, as they would learn to automate you as you use them. Additionally, if you were to die, others could query your model. You wouldn't even have to write an autobiography.

While this may be quite WKTE in science fiction, the particular path of using pre-trained models to cheaply compress data to save space, in combination with building user's personality model through asking in addition to listening (e.g., current intelligent assistants do not proactively and randomly ask things about you, and they don't try to learn you and help you without explicit commands, but reflectoputers would), are the parts the combination of which may have not been readily pursued by academic research in the particular way I describe here. So, this idea is a concrete idea for a research paper and a computer program. I may be wrong, and it all may have been already done. Search is in progress.

Credits: Mindey of HalfBakery.



(別通知) (可選) 請,登錄

一面深鏡子?

A deep mirror?


如影隨行……

如影隨行……


好吧,主要功能是它不會連接到雲,因此,它只是您的個人,您可以刪除收集的數據/模型,然後從頭開始或修改模型。

//如影隨行……

不過,與陰影進行比較確實是不錯的,這是什麼電影[Omniscient](https://www.imdb.com/title/tt11542920/),不是嗎?雖然,問題出在中央計算機。反射式計算機將不會有。

Well, main feature would be that it would not be connected to the cloud, so, it's just your very -puter, and you could delete the collected data/model, and start from scratch, or modify the model.

// Go with your shadow ...

Comparison to a shadow may be good indeed though, -- what was the movie Omniscient, wasn't it? Though, the problem was the central computer. The reflectoputers wouldn't have that.


是說不連網的人工智能嗎?


//這是否意味着未連接的人工智能?

的確如此!它無需發送數據到Internet就可以工作。但是,它將下載經過預先​​訓練的模型,並且我認爲實際用戶如果願意的話應該可以共享自己新學習的模型的一部分。但是,不會代表該程序進行任何外部通信。

// Does it mean non-connected artificial intelligence?

Indeed so! It would work without sending data to the Internet. However, it would download the pre-trained models, and I think the actual user should be able to share parts of the newly learned model of their selves if they wanted to. However, no out-going communication would be happening on behalf of the program.


好吧,這是新的。然後我們可以使科幻小說中的類人機器人成爲現實。

Well that's new. Then we can make humanoid robots in science fiction works real.


語言