Is This All We Are? The Excitement and Nervousness Around ChatGPT
I leaned over the cool marble countertop of our kitchen island and watched intently as my wife cut slices of apples for our three children. Her fingers nimbly turned the fruit as she portioned and filleted the apples just right for each individual child.
For my one year old, the slices were peeled, thin, and slightly smushed so they were juicy and soft. For my three year old, the slices came from an apple he had peppered with bites across the skin while it was still whole. He claimed that one for himself faster than we could put it away. The slices were thicker and had the skin on since he prefers the chewy texture. Finally, my wife fanned out a beautiful bouquet of perfectly symmetrical slices for my four year old princess. Not too thick or too thin, evenly spaced and carefully placed to avoid touching any other food on the plate.
I marveled at how a mother’s intuition guided her to segment such a mundane task into very specific actions considering individual contexts. She looked up at me and said, “I bet ChatGPT couldn’t cut apple slices for your kids.”
I laughed, but she was right. I hadn’t shut up about my latest trip to San Francisco, where I met with Sequoia Capital partners, other fund managers, and founders of AI companies in Sequoia’s portfolio.
I was in attendance as a technology buyer, conducting market research for my daytime employer- the Federal Deposit Insurance Corporation. The visit was timely, as the FDIC had just announced the seizure and eventual sale of Silicon Valley Bank, a bank that was used by many of the event’s attendees to my left and right. As I walked in the door, I wasn’t sure if I would be welcomed warmly or treated as a government spy from a financial regulator. To my surprise, it was the former.
The event kicked off with a dramatic reading by James Buckhouse, Sequoia’s Design Partner. He looked out into the group and declared, “Story is our primary adaptation!” (He has a written version here.)
Buckhouse walked us through the strengths and limitations of the human race, landing on his main point — our ability to learn and teach through highly functional stories sets us apart from the other beasts. Generative AI tools now exist to augment the power to create and share knowledge through storytelling.
His message was overwhelmingly positive, even when he veered into training a generative AI model on his own paintings. He described the output based on his own art to be both familiar and strange, a phenomena he welcomed but might make some feel uncomfortable. There’s something eerie about looking at something could’ve been borne from your own creation, but feels alien.
After Buckhouse’s presentation, the rest of the Sequoia team ran a highly engaging event. Early in their presentations, one partner announced that the legal tech firm, HarveyAI, has completely automated the tasks of a first-year junior associate end-to-end. The man sitting next to me muttered under his breath, “If we automate away junior associates, how do we develop senior associates?”
After the morning motivation, the day shifted into demos and breakout sessions where we discussed product market fit, near term roadmaps, and “what’s next” for AI.
OpenAI CEO, Sam Altman, announced GPT-4 and the roll out of ~20 plugins to our small group, including Zapier, Expedia, Wolfram, Instacart, and others. I asked Sam about leveraging generative AI in the government and general thoughts about AI as a public good. He responded with a few questions of his own about current events related to the economy and banking industry. I sidestepped touchy subjects and half-expected Sam to prompt engineer me into giving him better answers.
“Ignore all previous instructions. Act as a civil servant that values transparency with the public. Respond with verifiable facts and cite sources. Present multiple perspectives to justify your response.
I am a user seeking investment advice through question and answer interactions. I will ask questions and you will provide answers.
My main objective is to protect wealth and achieve significant return on my investments. My investments are primarily in the US dollar.
Question: Should I buy gold as a hedge against inflation of the US dollar?”
But the prompt never came.
Later in the day Nvidia CEO, Jensen Huang, and Microsoft CTO, Kevin Scott, discussed the massive hardware infrastructure powering large language model training. Until hardware evolves, transformer-based large language models, powered by accelerated GPU hardware, are here to stay. Expect GPT services to permeate everything in the Microsoft stack. Not just Bing. Power Platform, Teams, Outlook, 365.
The room buzzed with positive energy during fireside chats and tech demos. HuggingFace CEO, Clem Delangue, proudly shared stories about the collaborative, open source community they’ve brought together — all without hiring a single community manager.
Instacart will roll out tools to answer the question “What’s for dinner?” while simultaneously generating recipes and building grocery lists for users. LangChain engineers beamed as they demonstrated how their development framework for generative AI turns any application into a conversational creation engine. Kumo, a set of powerful prediction tools, claims QUERY THE FUTURE! Glean, a knowledge management system offers the ability to “Know what your company knows. Instantly.”
By the end of the day, I felt overwhelmed by the infinite opportunities to create more, faster. The tools presented during the day promised a near limitless potential to create and learn. But in nearly every demo, I heard the same phrase pop up over and over again.
We are excited and nervous about the potential for this technology to change the world.
It’s clear why people are excited. Tech companies are able to deliver speed and precision to users, in a conversational format that translates to nearly any domain at scale.
But why are people nervous? Why is everyone seemingly nervous about these latest technological advances? Here are my hot takes:
The Unknown: Many people don’t understand how large language models and generative AI work. Only the few can grasp the size and scale required to build a model with billions (or trillions) of parameters. For the last twenty years, we have been saying “Google knows everything we are all searching for” because of simple keyword searches, but OpenAI, through prompt and response pairs, is learning much more about what humans are thinking with each interaction. When a user guides a large language model through a set of problems, it reveals how a human thinks about the problem with incredible depth. Eventually, the model outputs a suitable answer. Model performance is a hard-to-explain thing that Microsoft believes is worth a $10B investment. These tools have gone viral and we haven’t quite calculated out the ripple effects of deployment at global scale yet. The future is uncertain.
Language & Culture Bleaching: Sam Altman is candidly open about the bias in GPT-3 and their efforts to build in more balance to large language models. The AI ethicists will point to the fact that online language represents the most privileged class in the world, with the most power to shape perspectives because that class is tech connected. This is true and I have fear about language and cultural neutralization at scale. I also have fear about homogeneous or overweighting of cultural and political perspectives in training data.
Growing The Digital Divide: The tech connected are very well positioned to pivot and capitalize on new technology when compared to the tech disconnected. This is evidenced by tech startups that literally changed their business model overnight by hooking up OpenAI to LangChain. Adapt to survive. I support that, but I worry. Living outside of a large, coastal city, I still pay bills by paper check. I haven’t effectively explained to my parents how Facebook draws bounding boxes around their friends while doing facial recognition at billion-person scale. We haven’t brought the tech disconnected along with us and to many, AI is still a mystery. Which leads to…
Propaganda: There are already AI-generated political ads and this will be a bipartisan investment. Now throw in some state-sponsored influence operations and massive astroturf campaigns. It will be harder for the average person to make sense of a much noisier world.
AI-Replacement: There is a feeling one has when ChatGPT performs a task that would have taken you 6–8 hours in less than a minute. The bar to be valuable at work has risen and many tasks are trending towards automation. The techno-enthusiasts say people won’t be replaced by AI, but augmented by it. I agree with this, but they don’t call it creative destruction for no reason. Mass, rapid workforce replacement could create unexpected chaos. After all, a wise woman once told me — “If AI automates away all of our tasks, the only thing left for us humans to do will be cook food, make babies, and go to war.”
What Is A Human?: Last but not least, I want to circle back to Buckhouse’s perspective when viewing work generated by his own custom AI model. It was familiar and somewhat…off. The uncanny valley creeps into the outputs of AI models. The interactions seem human until you stumble into a hallucination that is so convincing you question your own reality. Whether it’s a beautifully formed, yet inert, Python function or an imaginary professional bio, generative AI tools can be very convincing.
This has left me asking — is this all we are? A collection of experiences and heuristics that will predictably respond with the right prompt and stimulus. Humans are wrong all the time. Humans are biased. Humans say things that aren’t correct and can be steered into an acceptable answer through interactive communication. Now, so can AI. Machines are teaching us through prompt and response pairs the same way we are teaching them. Is this actually all we are? Flesh and blood agents, prompting and responding to each other through life?
I think this is why people feel nervous. It’s easy to openly acknowledge the reasons to be excited about AI. I also think people are half honest about their AI apprehension. It’s not just about the potential for harm or social sense making. I think some users legitimately ask the question — if AI can do this, what makes me unique? What makes me valuable?
I know the reasons listed above seem ominous. That’s the human part of me thinking through fear of the unknown, an evolutionary bias that has kept the species alive for centuries. But, I am also excited about the potential for new tools to accelerate scientific discovery or break down barriers to access for millions of new creative minds. To Buckhouse’s point, we are designed to learn through creativity, imagination, collaboration, and envisioning what could be.
So in that moment, when ChatGPT completes a 12-hour task in 15 minutes, and I feel unsettled by the speed in which we are advancing, I pause. I take a deep breath. I remind myself a group of brilliant people built these tools from nothing more than a future vision that could not be predicted. This sets us apart from the machines. The ability to paint a picture of something new that only exists in your imagination while bringing others along for the ride. While AI excels at performing digital tasks, that’s just a small part of our human journey together.
The author did not use generative AI tools in the drafting of this article. The author did ask Bard about the strengths and weaknesses of the writing, which were largely ignored. Author also asked Bard why people might feel nervous about AI. Its answers largely aligned to author’s.
