The US entrepreneur Reid Hoffman co-founded the career network Linkedin in 2002. Since then, Hoffman has been active as an investor and has invested in AI startups such as Tome, which can transfer content into presentations, and at times in the AI research company OpenAI, which made its voice assistant ChatGPT available to the public in 2022. In March 2023, Hoffman resigned from his position as shareholder to prevent conflicts of interest with other AI investments.
A year earlier it was announced that the founder was building a new startup with Inflection AI and was working on his own language model called “Pi”. According to the website, the AI will answer questions “from general knowledge to personal relationships.” Hoffman also works as a writer: In his new book (translated) “Ideas at the touch of a button – how we can develop our full potential with the help of artificial intelligence,” the tech pioneer experimentally deals with OpenAI's latest version GPT 4. Hoffman analyzes for various areas of life – from education to justice to social media – how the language model works and what advantages it can bring personally and in the world of work.
A book excerpt from “Ideas at the touch of a button – how we can develop our full potential with the help of artificial intelligence” by Reid Hoffman, published by Börsenmedien AG 2023.
Just as I expect GPT-4 and similar technologies to transform the world of work, I also expect them to transform the way I work. I've had the opportunity to experiment with GPT-4 for a few months, and while I know I'm still in the steep part of my learning curve, I think I have enough experience to offer some advice on using these tools give.
When disruptive technologies emerge, most people are tempted to use them as a replacement for an existing technology or technique. This seems sensible because it minimizes the time and effort required to use the new technology, but this approach is actually a trap: rarely is the new technology an exact analogue of what came before. When the Internet emerged, pioneering services like Yahoo! the online telephone directories. Directories were used to find things, so the seemingly logical first step was to create an online directory.
Over time, we discovered that the better approach was to create a new tool: the search engine. We are still in the online phone book phase of LLMs. They are unlikely to be a direct replacement for search engines in many of their use cases, but they will provide people with new ways to gather relevant and useful information. Here are three key principles that I have found useful in my own attempts to use GPT-4 in my work.
Principle 1: Treat GPT-4 like a student research assistant, not an all-knowing oracle.
If you have ever worked with a research assistant (or been one yourself), then you know that such an assistant is both powerful and limited in some ways. In some ways, GPT-4 far surpasses any human research assistant: it has access to an incredibly broad knowledge base, it's lightning fast, and it's available whenever you need it (as opposed to having to study for midterms). It also has many of the other disadvantages of a human research assistant: it is not an expert, its understanding of a given topic is fairly superficial, and it makes mistakes. If it's wrong, it's even worse than a human research assistant, because a human often has the common sense to warn you if they're unsure about the quality of their work.
Despite these drawbacks, what is truly remarkable is that everyone has access to near-instantaneous, on-demand research support. Just make sure you cross-reference the results with deeper sources and real human experts.
Principle 2: Think of yourself as a director, not a carpenter.
When we work with tools, we are trained to view them as an extension of ourselves. A carpenter swings a hammer and gets one step closer to completing a house. He knows what the hammer will do every time he swings it and works accordingly. This metaphor no longer works when it comes to tools like GPT-4. Instead of manipulating a physical tool to achieve a predictable and planned result, using GPT-4 is like staging a scene in a movie. As a director, you work with an actor to get the best performance from them. You don't tell him: “Bend your neck 15 degrees and then look at the other person after 2.5 seconds.” Instead, you ask him to convey a certain feeling to the audience: “Convince us that you are in love .” GPT-4 is not a human actor, but the directing process is similar.
You have to tease out the desired result and perhaps try many variations on a prompt, just as a director might need to do multiple takes of a scene. Ultimately it is a collaborative process.
Principle 3: Just try it!
In most areas of our work, we are taught to plan ahead and avoid mistakes. That's because implementing a plan takes a lot of time and other resources – it's not for nothing that the carpenter's saying goes, “Measure twice, cut once.” But what if implementing a plan was cheaper and quicker than that Thinking about it? This is the confusing paradox of GPT-4 and LLMs. In far less time than it takes to discuss a plan, GPT-4 can easily generate a complete answer for you to review. If you don't like the answer, you can throw it away and generate another one, or simply generate three variants to have more choices.
For example, I've gotten into the habit of trying out some simple instructions every time I have an important task ahead of me. For example, sitting down to write this chapter.
Reid: What ways are there for investor, author and philanthropist Reid Hoffman to use LLMs to work faster, more efficiently and effectively?
- Automated research: LLMs can be used to search and find relevant information quickly and precisely. Hoffman could use this technology to quickly find research and data useful for his investments and philanthropic efforts.
- Automated content creation: LLMs can be used to create content such as articles and reports, saving Hoffman time and effort in research and writing.
- Automated Analysis: LLMs allow large amounts of data to be analyzed quickly and accurately, allowing Hoffman to make better, faster decisions.
- Automated Communications: LLMs can be used for automated communications tasks such as responding to emails and scheduling meetings, freeing Hoffman to focus on more important tasks.
- Automated Investment Decisions: LLMs can be used to make investment decisions based on complex data and algorithms, allowing Hoffman to make informed decisions more quickly.
As usual, GPT-4 provides some immediate food for thought. Research and analysis is definitely a good use case for GPT-4 as it literally knows most of the internet around 2021. Its knowledge is far more extensive than that of a human, and future instruments are likely to be even more extensive. Automated content creation is interesting for instant first drafts, but I'm skeptical when it's claimed that it can be used for important texts without human intervention – at least for my work.
I would want to review and refine any articles or reports created with it, although it could still significantly speed up my throughput. I am a little more skeptical about the other suggestions, such as automating my communication. People send me emails asking for decisions and letters of recommendation, and I'm not quite ready to hand those tasks over to AI yet.
When it comes to investment decisions, I think AI could help me make informed investment decisions faster by making me think and helping me consider all potential data points. However, it seems to be an exaggeration to call this process an “automated investment decision”. Despite these minor objections, GPT-4 helped me tackle this chapter by creating a sensible taxonomy of my work with pointers to where I could delve deeper.