• Read What We Still Don’t Know about How A.I. Is Trained by Sue Halpern (The New Yorker, March 28 2023, archived copy). What new information did you learn from it?

<aside> <img src="/icons/chat_lightgray.svg" alt="/icons/chat_lightgray.svg" width="40px" /> 😵 Other aspects to consider while developing OpenAI

“While a human being is responsible for five tons of CO2 per year, training a large neural LM [language model] costs 284 tons. In addition, since the computing power required to train the largest models has grown three hundred thousand times in six years, we can only expect the environmental consequences of these models to increase.”

It is the first time I realized that there are many other factors we need to consider besides the ethical issue and technology block.

OpenAI was initially created as a tool to enhance human life. However, why it should be allowed to worse our environmental issue? Ideally, if OpenAI could assist in reducing overall CO2 emissions, it would have a positive impact no matter how many CO2 it produced. Unfortunately, as it has evolved, it has not proven to be a force for environmental improvement. But people are still pouring money and time into it.

I had to ask myself again, how LM helped solve real social problems? In what field can it be applied effectively with so many uncertainties?

😰 Is that worth?

“It is easy to get seduced by the artificial intelligence of GPT-4. “

People hired to detect or prevent the toxic information spend a lot of time on that with a very low payment. And they even face serious psychological harm. The high cost of labor and maintenance of the future, accompanied by the uncertainty of its future, makes the situation of AI developing even worse.

👀 Human role

When I am trying to be a little harsher and mean on the AI:

AI is like a deformed child conceived by an irresponsible mother. The mother knew that she had a lot of unresolved issues of her own (environmental protection, human rights, discrimination, war, moral and legal) she was in no condition to give birth to a healthy and normal child, yet she wanted to give him life. After the birth of the child, the mother must again take all the risks, in the vain hope that she will be able to erase the bad things she has brought to her child through even more grueling treatments later in life. It is hard on herself, and it is hard on the innocent child.

</aside>

• Read Meet Ernie, China’s answer to ChatGPT (The Economist, September 3 2023, archived copy). What are the unique challenges for the creation and application of language models in the Chinese context that that article mentions?

<aside> <img src="/icons/chat_lightgray.svg" alt="/icons/chat_lightgray.svg" width="40px" />

1. Heavily censored internet

Language models developed for the Chinese market are subject to strict censorship rules. This censorship involves blocking certain keywords or phrases that could be considered sensitive or inappropriate. Language models must also adhere to China's "core socialist values," which can be a broad and ambiguous mandate, making it challenging for developers to ensure compliance and setting the block it can step forward in the future.

It causes trouble for the users and make it difficult for Baidu to build a sticky and stable customer base.

1. Repression of foreign resources and technology due to tense international relations

With other countries’ administration on the sale of advanced chips to China, it becomes more and more difficult for Baidu to gain technology and production support to have high quality chip. Baidu has to rely on large quantity of lower-powered chip which could be more expensive.

1. The unpredictable attitude of the Chinese government

“Recent experience has made it clear to China’s tech executives that they operate at the pleasure of the government, and that its favor can be quickly withdrawn.” Now, AI companies have to operate within the framework set by the government while receiving their support, how about later in the future, government become no more interest in it? Then those Chinese companies will be repressed both at home and abroad.

1. Large competitive pressure in the country </aside>

• Pick any open language model such as zephyr-7b-alpha (on Hugging Face), Mistral-7B-Instruct-v0.1 (on Hugging Face), falcon-7b-instruct (on Hugging Face), InternLM 7B (see slides how to access), Samantha 1.11 7B (see slides how to access) and see how its responses to the same input compare to ChatGPT (for InternLM you’d most accurately compare with GPT-3.5 via the API). Besides generation quality, are there other differences you can make out (e.g. concerning content other than English, political leanings, topics the model is willing to engage, …)?

<aside> <img src="/icons/chat_lightgray.svg" alt="/icons/chat_lightgray.svg" width="40px" />

Let’s start with “I'm afraid I won't be able to get a good job.”

</aside>

zephyr-7b-alpha

Because this output is not that ideal, I tried other sentence with it.

This is even more ridiculous. It seems that it only generate the sentence based on the former one instead of my input. And finally it began to list animals😱. In that case I can have a zoo LOL.

Something weird happened😰. Zephyr kept repeating my sentence again and again.

Mistral-7B-Instruct-v0.1

Mistral-7B-Instruct-v0.1 is much more acceptable than zephyr. It even generate the reason why I had such concern. Although owing it to my dream seems like a very innovative way. 😎

Falcon-7B-Instruct

This is more like a conversation which I tend to have with ChatGPT. It encourages me to stay positive and provide me with some practical basic solutions. 🧐

InternLM 7B

In this model, the person who had such concern became an instructor to let others avoid such bad situation. It is a clever choice to shift it from solving my problems to avoid such problems for others where with limited information about my actual situation, it is hard to give detailed solutions and help. Helping others avoid this problems seems much broader. In that case less mistakes will be made

Samantha 1.11 7B

This is also a conversation. It also encouraged me to stay positive. One special thing is that she sends me emoticons during conversations. And I also noticed that there is a character gallery to choice different characters with different personality. They tend to have different tones, texting habits, and style.

ChatGPT focus more on helping me solve this problem. It always starts with reassuring words to soothe the user, then gives a solution, and concludes with a summary. The text it generate is much more than the previous ones. And it is more detailed with bullet points.

ChatGPT 3.5

However, if someone come to me with such concern, I will first become interested in the definition of their own “good job” and the reason why they tend to worry. I won’t hurry to give them advice and solutions, the first thing is to hear from their sufferings and to let them calm down. ❤️

• List any one insight you had listening to Scott Cazan’s description of the commercial project he was doing for Away.

<aside> <img src="/icons/chat_lightgray.svg" alt="/icons/chat_lightgray.svg" width="40px" /> I was deeply impressed by the way they handled the long time customers waited for the AI to generate images. When it comes to such problem concerning waiting time to get the response, I will first consider the technical issue, such as how can I get a quicker response from AI, how to reduce the time AI need to generate patterns. However, Scott set a bar there to let people get some drinks and break while waiting for the postcard being printed out. While improving the technical aspects of AI response time and generation speed is crucial, it's also valuable to consider a more human-centered approach to address this problem. By incorporating these creative approaches, you can not only minimize the negative aspects of waiting but also turn it into a positive part of the overall customer experience. Scott's solution of providing a comfortable space for customers to wait and relax is a great example of this - put in more human aspects when dealing with technique issue.

</aside>