From Israel to AI: Elad Gil, investor
Elad Gil is 47 years old and very rich. Born in Israel, he grew up in California where he crushed it in school - getting two degrees from UC San Diego (Math and Molecular Biology) and then a PhD from MIT. He started at McKinsey, moved to Google, and then built his first AI startup which Twitter bought. After working at Twitter he started creating and investing in companies.
Gil has an almost magical sense for spotting winning companies. He was one of the first investors in pretty much every cool tech company in last 15 years: Airbnb, Airtable, Anduril, Coinbase, Deel, Figma, Gitlab, Gusto, Instacart, Notion, Opendoor, Pinterest, Square, Stripe... the list goes on. No surprise that other rich people are lining up to give him their money to invest.
Gil is extremely knowledgeable about AI: both classical ML/NLP and GenAI of today. One of his recent investments is Braintrust, an important AI company that builds AI evaluation tools and other AI infrastructure. He's also invested in other hot AI companies: Character AI, Harvey, Mistral, and Perplexity.
You can hear him share his AI wisdom on "No Priors" - probably the best podcast about AI analysis out there - which he co-hosts with the brilliant investor, beautiful Sarah Guo.
In late October Elad gave an interview to fellow investor, billionaire and podcaster, co-founder of Palantir, Joe Lonsdale.
Here's a balanced version that keeps Elad's voice but breaks things into digestible chunks:
AI is, in my opinion, dramatically underhyped.
"We're seeing massive actual revenue and impact of AI without that much adoption. Example: last quarter for Azure was $28 billion, and they say 10-15% was from AI products. $2.5 billion quarterly. Take Klarna - they reduced their customer success team by 700 people by adopting AI and suddenly had 24-7 availability in 30 languages with a higher net promoter score, faster response time, higher customer success.
People misunderstand what is going to be the end product [of AI industry]. I think the end product is units of cognition.
You're going to be effectively paying for or renting something that's going to think or do things on your behalf. That could be legal assistants, could be a customer success team, a software engineering team. This is a revolution in terms of units of cognition.
"Right now in services, it's not pure cognition yet. It's doing things to save people tons of time and make each person able to do three times or four times as much work. For example, an auditor - you can bring up right away there's 10 screens they're most likely to look at, and that saves them five minutes each time. That's huge.
"5-6 years ago I wrote that there'd be three eras of intelligence. The first era is people. The second era is a hybrid era where we will have a mix of humankind and machines.
Eventually intelligence will be dominated by machines, in terms of sheer number of brain equivalents that will exist out there.
"If you look at the areas that AI can transform, it's probably $5 trillion in headcount. Our best estimates show about 40% of that could be transformed by AI right now. We think already you can at least double the productivity, which would be a trillion dollar pullout - that's huge.
"We had these prior waves of machine learning - neural networks, RNNs, GANs and others. But they were just good at pulling out statistical associations between large data sets. The current wave, the transformer architecture, is actually understanding and generating language and images. And I think LLMs are not even close to an asymptote.
Legal firms say that the nature of a law firm will change.
"Now a firm hires 50 associates every year, and 5 of them will make it to partner. But in a couple of years, they can just hire 5 associates to begin with - then who becomes partners? Maybe you will have one senior partner, two associates and 30 bots.
I see 5 layers for the AI value stocks:
1 - The bottom layer is NVIDIA and other chip manufacturers.
2 - The next one is the data centers.
3 - Then it's the models, whether OpenAI, Anthropic or Llama.
4 - Tools for deploying AI, "AI infrastructure" like data infrastructure.
5 - And level 5 is the company owning the workflow.
I've mainly been focused on the top three of those layers.
The exciting part is that eventually, each person will have a custom tutor, helping them learn deeply at their own pace. AI is perfectly suited for that. Research from the 80s shows that kids who receive one-on-one tutoring learn dramatically faster - by one or two orders of magnitude.