2023 Elad Gil predictions

startup macro, LLMs, defense-tech

Macro on startups

• “Startups are an act of desperation” - desperation went out of the ecosystem over the last few years with people being very status/money focused. now people are building for other reasons including impact/products/career acceleration

• 2022 was the warm-up to bad times… 2023 is the bad times - we’ll see mid-late stage, layoffs, blow-ups, down-rounds, shutdowns

ChatGPT demonstrates potential of LLM, but still early

• chatGPT is a key moment demonstrating the potential of LLM models

• who develops/commercializes this? maybe OpenAI, Google or someone new

• competition creates the need to reexamine assumptions about safety and reputation risks

• virtual sssistants (like adept.ai) understand language and take action

On AI models

• AI open vs closed - the massive models today largely closed (e.g. Chinchilla, GPT-3.5) while the image-gen models more open (Stable Diffusion)

• image models are easier from a cost/quality-expectation-threshold

• higher threshold of quality required on language-gen than image gen

• 2x2: image vs. language, big vs. small model

Costs to train LLM models today

• costs to train a GPT-4 competitor from scratch might be tens of millions to low hundreds of millions of dollars today depending on baselines starting from

• capital scale will hit billions shortly, limited window of time for startups to enter LLMs

• does the industry structure look more like the cloud world with multiple competitors or more like semiconductor world where just TSMC won?

Space and Defense

• SpaceX Starlink

• Anduril becomes the default defense-tech startup

• defense tech

  1. rise of drones

  2. rise of machine learning/machine vision

  3. cybersecurity

• why defense tech hasn’t worked well before

  1. no “why now”

  2. defense tech starts with a point product as a subcontractor

  3. need a lot of capital

  4. business model of cost-plus versus - Anduril sells gross margin model

  5. execution speed of deal-closing