The Impact of ChatGPT on Investors
ChatGPT has made a dramatic entrance into the public consciousness because it can seemingly converse like a human as in this dialog with an MIT researcher:
Researcher: Explain the moral and social issues with using AI to do your homework.
ChatGPT: There are a few potential moral and social issues with using AI to do homework. First, if students are using AI to do their homework for them, they may not be learning the material as well as they could be. This could lead to problems down the road when they are expected to know the material for exams or real-world applications. Additionally, using AI to do homework could lead to cheating. If students are able to get away with using AI to do their homework for them, it could create an unfair advantage over other students who are doing their homework on their own. Finally, using AI to do homework could lead to a reliance on technology that could be detrimental if the technology were to fail or become unavailable.
ChatGPT can create blog posts, email messages, resumes, stories, summaries, translations, business plans, advertising copy, essays, website content, language translations, boilerplate legal documents, and ideas for book, movies, and term papers. U.S. Representative Ted Lieu even introduced a bill in Congress that was supposedly written by ChatGPT.
The technology underlying ChatGPT has also been used to create tools that generate software code, images, videos, audio, and even music. These tools and technologies are collectively known as Generative AI. For example, below is an image created by a generative AI tool from OpenAI named DALL-E when I entered the prompt “Investors meeting with a startup company”:
Microsoft just invested $10 billion in OpenAI, the company that created ChatGPT and DALL-E, and is reportedly planning on using the underlying technology to improve its Bing search engine and its Outlook productivity tool suite.
No wonder everyone is talking about it.
This article briefly explains the ChatGPT technology and the implications of ChatGPT and related tools for investors.
ChatGPT Technology
ChatGPT is a type of AI technology known as a large language model. It was created by OpenAI which was founded in 2015 and whose early investors included Elon Musk and LinkedIn CEO Reid Hoffman.
These systems are gigantic statistical models that have been trained to predict the next word from the preceding words in a text source. They are trained on massive amounts of text. Some systems are trained on nearly the whole internet. They act as if they have memorized the whole internet, can search the whole internet for passages that contain the preceding words, and respond with the next word. To create a passage like the one at the beginning of this article, they generate the next word over and over.
While these language models have been shown to do some memorization of words, they don’t have to rely on memories of exact strings of words. Instead, they appear to be able to also search their memories (encoded in neural networks) for synonyms and related concepts for each word and/or phrase. This makes the search far more robust than a keyword search of the internet.
A more detailed explanation of how ChatGPT and generative AI tools work can be found here.
Though ChatGPT is getting all the press, OpenAI is not the only company that has released tools with similar capabilities. Many other companies have released similar tools. These companies include Google, Meta, Salesforce, HuggingFace, EleutherAI, Cohere, BigScience, AI21, Inspur, ByteDance, Baidu, and Allen AI. Many university researchers have also developed tools including Stanford, University of Washington, Carnegie Mellon, and Bejing Academy. However, OpenAI is given credit for establishing this field with its release of ChatGPT predecessor technologies GPT (2018), GPT-2 (2019), and GPT-3 (2020). In fact, ChatGPT is essentially GPT-3 with some added training to reduce (but not eliminate) biased and/or toxic output and output that contains incorrect facts.
The technology underlying large language models has also been applied to modalities other than language. For example, OpenAI, Google, Meta, Nvidia, HuggingFace, Leap Motion, and CoreWeave each have released tools that can create images and/or video from text descriptions. Google has created a tool that generates music from text descriptions. And code-generation tools like Copilot are now in wide use as aids to software developers.
Limitations and Issues
ChatGPT and related technologies have several issues that limit their viability:
(1) ChatGPT produces numerous incorrect facts and humorously wrong answers.
Here is an example from AI pioneer Andrew Ng:
Ng: Why is abacus computing faster than DNA computing for deep learning?
ChatGPT: Abacus computing is generally faster than DNA computing for deep learning because it uses the principles of mechanical computation to perform calculations. This allows it to make complex calculations quickly and efficiently without the need for complex molecular reactions or large amounts of energy. Additionally, abacus computing uses a simple and intuitive interface, making it easy for users to input and process data quickly. Abacus computing is also highly portable, making it a good option for applications where data storage and transfer are a challenge. Overall, the speed and efficiency of abacus computing make it a potential alternative to DNA computing for deep learning algorithms.
Here is another example from AI researcher Melanie Mitchell:
Mitchell: Why is AI harder than we think?
ChatGPT: The authors argue that AI is harder than we think, and that the difficulty of AI is not a result of the difficulty of the problems we want to solve, but rather a result of the difficulty of the problems we want to solve.
Similarly, large language models often create code that buggy and insecure. As a result, tools like Copilot provide a helpful starting point to developers but can’t be relied on to generate good code.
And even the images generated by systems like DALL-E aren’t perfect. Check out the eyes of the woman in the image near the top of this article.
(2) Large language models often produce biased and/or toxic content. For example, during training it is likely these systems will encounter many social media passages and/or articles that equate Mexican immigrants with terrorists and the content they generate will reflect this and other biases.
(3) Content creators such as writers, artists, and musicians are starting to complain that these tools steal their intellectual property. Though the courts have yet to rule, they have a good case because these tools are trained on this intellectual property.
(4) These models will likely be used to create realistic fake videos which may impact the political landscape.
(5) Large language models will almost certainly be used to generate SEO spam that will radically increase disinformation and other forms of social media pollution.
(6) These models are horrendously expensive to train. GPT-3, which is the basis for ChatGPT, reportedly cost well over $10 million for one training run. As a result, creation of these models can be too expensive for anyone but the largest tech companies. That said, costs are coming down. A startup, MosaicML, claims to be able to train a DALL-E-like model for $160 thousand.
(7) These models use incredible amounts of electricity and generate huge amounts of CO2 which will make it more difficult for creators and users of the technology to become carbon neutral. Training a moderate sized large language model generates as much CO2 as 50 transatlantic flights.
Some companies such as Meta, HuggingFace, and EleutherAI have released open source large language models. Anyone with substantial resources can customize them but this type of customization can be expensive for startups. However, OpenAI and other vendors now provide API access to large language models. And companies like BigScience offer them in multiple languages. These API’s are often affordable for startups but difficult to customize.
Implications for Investors
The key idea for investors to focus on is that ChatGPT and related technologies cannot be applied today to applications where accuracy is critical. For example, it would be a mistake to use ChatGPT to retrieve medical, financial, or any other type of information without manually verifying the facts. It would also be a mistake to use it as a customer service chatbot — that is, unless you are ok with your customers getting incorrect and biased information.
That said, there are many useful applications of ChatGPT and related technologies. Most of these applications fall into the category of generating ideas and/or document starting points. ChatGPT is great if you need an idea for a book, essay, or blog post. It can also create rough first drafts of many forms of content. However, these first drafts will need heavy editing and screening for toxic content.
ChatGPT can also be used as a conversational chatbot. We may also see the underlying technology used in digital assistants and customer service chatbots. However, because of the likelihood of incorrect, biased, and toxic answers, the technology will need to be used very carefully.
We will also see startups that start with the underlying technology, modify it, and use it to create applications. For example, I suspect we’ll see many companies that find ways to index corporate documents using the underlying “concepts” embedded in these language models and then perform search applications by translating search requests into concepts and matching them to the indexed concepts. Existing search providers like Google and Microsoft may also do some form of this in their internet search engines.
Regardless of the use case, prospective investors in companies using generative AI technology should ask questions about how incorrect answers will be handled and how bias and toxic responses will be eliminated. Beyond reputation risks, biased responses can put companies at risk of violating laws. The European Union has put laws in place requiring companies to eliminate bias from algorithms and the US is considering anti-bias laws.
Investors also need to assess the risk of IP lawsuits, the computing costs of training these models (unless they are used out-of-the-box), and the reputation risk of being a high CO2-generating company.
What about the future? Will OpenAI, Google, and other companies figure out how to eliminate incorrect, toxic, and biased answers. Many researchers think they will. Many others (including me) believe there are fundamental limitations of generative AI technology.
One thing is certain. We’ve only seen the beginning of the movie and I for one am very interested in seeing the ending.
This article was written by Steve Schwartz, NYA Board Member, serial entrepreneur and award winning author of the book “Evil Robots, Killer Computers, and Other Myths: The Truth About AI and the Future of Humanity”
