Stream of thoughts on AI

Getty Images sues Stable Diffusion…. in the UK. With their IP friendly courts and innovative precedents such as Temple Island Collections Ltd v New English Teas Ltd (that extends copyrights to photograph concepts in addition to their expressions), I’m pretty sure the claimants will breeze through the courts with a solid win (just like Amber).

ChatGPT 個 G 係 「Generative」，其實本身係一個「文字生成器」。作為一個「語言模型」(language model)，佢主要訓練嘅係「語言能力」，唔係認識現實世界嘅任何嘢。與其當佢係一個「會俾錯答案嘅Google」，我覺得應該當佢係一個鳩噏大師（褒義）。好聽少少叫「文妓」，正式啲可以話係「文膽」。即係你叫佢寫啲嘢，佢可以文通字順、行文流暢，就算顛倒黑白都冇問題；但內容係咪準確，就閣下自理。所以最合適嘅用法，就係本身得兩句嘅內容，你想嘔五百字大方得體嘅說話出嚟嗰陣用佢。如果上面嘅講法都太抽象，或者咁講大家會明：

⋯　你當佢係19才子就啱㗎喇。

GPT can be implemented in less than 100 lines of code. I’m not sure anyone actually understands the implications of this.

We have through experimental science proven that, there exist relatively simple systems where if you feed them sufficient amounts of data, they will become as intelligent the data allows. The threshold is at most some tens of billions of repeating “units”, possibly less, organized in a way that follows a gradient and settles on local optimum state (which many physical processes do naturally). This is obviously less complicated than the simplest lifeforms we know on Earth!!

So, we have basically proven that life is not a prerequisite for intelligence, and that intelligence can be structurally less complex than life. Think about all the possibilities. Elementals, Nature Spirits, Gaia, God…

The extent that language models can write code but can’t do math is evidence that programming (of the kind language models do well) is a language skill, not a “math” or “logic” skill (to the extent that math or logic is not a language skill).

I have long argued that software engineers should take courses in creative writing instead of linear algebra and calculus (if one must choose). We haven’t proven this is better yet, but at least we’ve seen (IIRC, read it somewhere but I don’t have references, appreciate if anyone can suggest some) that the natural language ability of language models are improved after being trained on computer language corpora….

Note: the qualifications/reservations in the brackets are important.

This might have been obvious in retrospect - if we adopt the Turing Test definition of intelligence, a language model that accurately “predicts the next word” of human text is obviously going to be intelligent.

見親啲個 CV 有「Engineering」呢個字嘅人講 ChatGPT 乜乜柒柒就覺得好笑。你估真係有個電字就代表你識？

In terms of entropy, Intelligence ~= Compression ~= Encryption ~= Random Noise

The difference between these four things is literally a matter of subjective interpretation.

Now, look at nature, look at all those sources of random noise. They say it’s just random noise, nothing intelligent is going on. It’s as if you look into a TLS session and conclude that humanity is just transmitting zetabytes of white noise every day 😃

PS: what does the so called “laws of thermodynamics” really imply with the entropy increase then? In general I think the prevalent interpretation by physicists of entropy is a bit too narrow-minded. Randomness is merely a measure of our lack of knowledge or capacity to understand.

GPT-4 參加考試。基本上科科成績都係 80+% percentile，包括考律師試有90%， 比 GPT-3.5 進步咗唔少 （證明讀多啲書真係好緊要 [誤]）

好在我玩 codeforces 同 leetcode 可能仲有機會贏佢⋯

去到原版 ChatGPT/GPT3.5 嘅年代，我已經感覺到佢 gen 出嚟嘅嘢比一般普通人嘅水平為高。坊間不斷話 AI 有呢樣嗰樣限制，根本係啲人 hallucinate 出嚟嘅論述，潛意識令自己好過啲。GPT 基礎技術來自 2017 年一份論文 (“Attention Is All You Need” [利申冇睇過])，大家就攞住一個開發咗五年未完全成熟嘅新技術鳩噏佢未來幾十年會有咩限制，唔去諗下佢會越嚟越勁，呢種心態其實真係幾好笑下。

「哎，課溫，窩襟拗汪波歐化囉多s」

https://openai.com/research/gpt-4

睇啲人討論「後 AI 年代應該教個仔咩技能」覺得人類真係幾可愛。

本身我哋上一代 (你當我係講緊 boomer generation啦) 已經唔多識教我哋呢代人要學咩技能，今日嘅 80/90 後又點可能覺得會「研究」到一個正確嘅方向，唔會老點仔女呢？

個 inconvenient truth 係，AI 亦已經取代埋「幫仔女舖路做成功人士」呢種「工作」。覺得自己叻過仔女嘅同時，又想個仔女成功過自己，本身係幾矛盾嘅諗法嚟。

大部份人係冇乜「先知力」嘅，世界嘅嘢你估唔撚到㗎喇。

Language models are going to be pocket sized. LLaMA and its derivatives already fit inside a phone (currently runs very hot and slow but things are just a couple optimizations away).

Super-human AGI models are going to be the new super computers. Imagine a beowulf cluster of these solving the hardest problems that have stumped humans for decades (P=NP?) An implied intuition of the Turing Test is that general language ability ~= intelligence. I can’t imagine these efforts not working out at least to some extent given the pace of progress.

A couple (10+) years ago the book “Thinking, Fast and Slow” by Daniel Kahneman popularized the idea of two modes of thinking: “System 1” and “System 2”. “System 1” is fast, instinctive and emotional; “System 2” is slower, more deliberative, and more logical.

GPT and other language models are basically similar to “System 1”. The models don’t have a way to pause to think (any pauses you encounter are due to high load from other users…) . I find it actually quite amazing that the AI models can produce high quality results without “pausing to think”, most experts can’t do that without some deliberation and planning before they give an answer to a complex question.

When we pause to think, we often brainstorm some likely answers and evaluate the merits of each of them, kind of like having a mini-debate inside one’s head. I wonder whether this technique can work on GPT models too – i.e. ask it to debate itself critically for a couple thousands of words, and then come up with a final answer. It would also be interesting to allow it to request online resources to look up factual data, before answering questions. It’s also possible that you could ask it to use different “personalities” to look at its answer from different “perspectives”, so that mistakes are further minimized.

I imagine that as long as the token limit is raised, it would be trivial to prompt the AI model to incorporate all the techniques above and loop through them a couple times to refine an answer. That answer is likely to result in super-human reasoning abilities (at least in many topics/fields/subjects), I think.

That is why I think those betting against AI capabilities is going to lose hard. If even an amateur like me can think of ways to trivially improve its performance, one can’t help but wonder how many crazy things the bleeding edge researchers are doing…

The only question is whether they’d even let us know. Good thing about USA is that people like to blow whistles, so… chances are still good.

Re: https://arxiv.org/abs/2303.12712

This is why the Ivory Tower and associated industries are rotten to the core.

Only 9 months ago Blake Lemoine was thoroughly ridiculed, with so called experts going so far as to claim that he was suffering from the “ELIZA” effect. The level of certainty they made the claim was mind boggling, only surpassed by the US health officials’ claim that wearing masks don’t help prevent covid.

Now, when respected researchers who couldn’t tell you what a “soul” is claims the same thing, they can pretend they’re the first to do so.