Thinking Through the AI Hype
This week I spoke with the CEOs of Cohere & Perplexity AI. Plus, information from Anthropic's fundraising documents. Ending with a poem from ChatGPT.
When a new technology and hype arrive at a party together, it’s hard to think clearly until hype has left the room.
For some early technologies, hype is the only evidence that most people ever get.
That’s been true with quantum computing. I’ve been escorted into a cleanroom wearing the full getup and looked at tiny chips and elaborate quantum machines. But at the end of the day, I just had to decide whether to believe investors and researchers as to whether the technology they were betting their careers on would deliver anytime soon.
When I rode in early versions of self-driving cars, it was similar — even though the experience was more tangible. The humans seemed to constantly intervene. Still, watching a machine turn the steering wheel was amazing.
Most technologies exist in a sort of middle area. Think the blockchain and the metaverse: There’s stuff for regular people to see — but evangelists promise the real magic is just over the horizon.
The hype isn’t always imposed from the outside. Ideas that are initially thrilling don’t always have staying power. In late 2021, Gather — a company that raised a $50 million Series B and had created an animated office that I could explore — really impressed me. I wrote a story about the experience.
I haven’t visited since.
The iPhone exists at the other end of the spectrum. We can hold it in our hands. Every day, it proves to me that it is useful and wonderful — my personal glass panel to the world.
Artificial intelligence sits in a weird space between tangible and imagined. It’s widely available but there is also an insanely speculative conversation on top of it.
Normally, the people shilling a technology hype it up and the opponents dismiss it.
In this case, critics are warning that artificial intelligence is going to doom humanity, while OpenAI CEO Sam Altman says stuff like “it is still flawed, still limited, and it still seems more impressive on first use than it does after you spend more time with it.”
This week, in the pages of the Wall Street Journal, Peggy Noonan accused the tech set of playing God. She wrote that, “developing AI is biting the apple. Something bad is going to happen. I believe those creating, fueling and funding it want, possibly unconsciously, to be God and on some level think they are God.”
Meanwhile, Farhad Manjoo wrote in the New York Times that “ChatGPT Is Already Changing How I Do My Job.”
He offers some compelling uses for AI.
He uses it as a sophisticated thesaurus to search for a better word.
ChatGPT can be a capable thought partner even when the human is doing most of the thinking.
And it can quickly pull themes out of a long document.
I agree with all that — and I would add that it can be good at generating a list of potential sources for a particular story too. (It would be even better if it had internet access.)
But I still wouldn’t say ChatGPT has really changed how I do my job. Often, ChatGPT can be more of a temptation than a real help.
When I’m tired and feeling lazy, I’ll try to get GPT-4 to do writing for me — only to find that it produces middling text that looks like a post but lacks real coherence.
For my job, text generation can be particularly enticing. At the most basic level, I am paid to produce text — and AI is great at producing text quickly in vast quantities.
The quality of the ideas and the clarity of the writing are what really matter. ChatGPT can give you so much text so quickly that it feels like spending hours to make something better isn’t worth it. But then I resist the urge and decide to continue living as a thinking, typing being.
I don’t want it to sound like I’m getting pessimistic about this artificial intelligence wave — I’m not. If anything I’m worried that I’m not technical enough to get access to the cutting edge open-source projects that people are playing with to create agents and to expand their memory access.
Friday, I chatted with Aidan Gomez, the CEO of foundation model company Cohere. In his spare time, he has built a web browser driver that has access to his credit card and his password credentials to “a ton of different websites.”
“If I say, ‘go buy me hand soap,” the AI agent knows his address and can order the soap online for him.
Gomez pairs his webLM with techniques to expand his memory window, so that Cohere is able to recall relevant information about him when it’s useful without storing it in the foundation model.
While there’s a lot of excitement about self-directed agents, I’m personally eager to have access to a much larger memory window and better connection to the current internet.
Currently, foundation models significantly restrict the memory capacity — determined by entry length measured in tokens — available to language models for generating their answers. If you pour your heart out to chatbots, eventually they start to forget what you said because they go beyond the context window.
It’s frustrating when I spend a lot of time with ChatGPT, teaching it about myself and creating a backstory with it in a pen-and-paper role playing game, only for Chat-GPT to start forgetting things. Making matters worse, ChatGPT has no idea that it’s forgetting things and starts bullshitting instead of admitting that it can’t remember.
“Certain use cases just demand more and more context,” Cohere’s Gomez told me. “It’s an area of hot competition. As you extend that, you unlock use cases along the way.”
Gomez said that foundation model companies are pushing towards infinite context windows. “You unlock 99% of use cases once you get around 8,000 tokens but there are some really interesting problems in the 1% that you push forward for.”
Cohere currently offers about 4,000 tokens and plans to offer just over 8,000 soon. (OpenAI offers about 4,000 tokens for regular GPT-4 users and offers longer context windows for users of its API.)
I asked Gomez whether he thought foundation models would start improving more slowly. After reaching parity with GPT-4, would Cohere start to hit the top of an S-curve? (Altman has suggested that OpenAI isn’t working on the next version of GPT.)
“There’s definitely a resistance point in capability that is — the best human on Earth performs at this level at that task,” he said. Once artificial intelligence is as good as the best human at a particular task, it may take more work to push it beyond human capacity, though that has certainly proved possible as we’ve seen with chess and Go.
I got my hands on foundation model company Anthropic’s Series C fundraising deck.
The document outlines a number of believable examples where foundational models like Anthropic are likely to make an impact.
In the next three years, Anthropic believes it will be able to serve as a “1-1 tutor, coach, or therapist for anyone in the world… For any domain.” The company also thinks AI will be able to create “advanced creative content generation” like movie scripts, songs, and video games.