Optimistic Notes Heading into 2023

Contrary to what I wrote in 2020, I’m much more optimistic about the future these days. Reviewing what I wrote 2 years ago, however, my optimism then was mostly misplaced. If anything, we saw an overemphasis on financial engineering and FinTech since 2021, and the acceleration we saw during COVID seems to be fleeting as most metrics are returning to their pre-COVID trend lines. However, after the economic turmoil of this year, which will continue in 2023, there is one standout reason to be optimistic for the future: Artificial intelligence.

Artificial intelligence, particularly generative AI, has seen tremendous strides in 2022, building on top of the gold rush of AI and machine learning research in the last 5 years. Lensa is a great example – a mobile app to generate profile pictures, built on top of Stable Diffusion, an open source image generation model released earlier in 2022. ChatGPT is even more interesting, as it was a new product built on primarily 2 year old technology. ChatGPT was eye opening for many people, as it was an incredibly user friendly interface to experience modern large language models (LLMs). Having your model respond in first person seemed to break many people’s brains, but the people who were doomsaying about the singularity or the threat of AI due to ChatGPT are probably closer to the truth than those who ignore it or minimize the impact similar technologies will have over the next decade. Still, Lensa, a consumer app available in mobile app stores, is probably the first significant consumer product leveraging generative AI for a key feature, and will be the first experience with generative AI for many.

ChatGPT being built on a fine tuned version of technology from almost two years ago is actually evidence of a critical aspect of the current state of AI – research is leaping well ahead of the current state of products. Current best in class AI products are toysMidjourney is incredible to see in action, but still seems mostly to be used due to the novelty, not for productive purposes. Text editors like Sudowrite or Lex are starting to get at productive use cases, but still very early days and in search of a critical customer. The fact that the killer app of AI in 2022 was a demo from an AI research group, in my mind, is evidence of a failure of innovation from product developers and designers.

Research in machine learning is moving incredibly fast. Attention is All You Need was only written 5 years ago, and there are still incredible gains found just by experimenting with taking a technique that worked in one domain and trying it in another (like taking techniques from BERT and applying to TTS instead of NLP). Relatively simple training techniques are bringing huge gains (ChatGPT was built on GPT3.5, essentially just GPT3 with reinforcement learning from human feedback [RLHF]). This is the metaphorical gold resting on the ground in California in the 1840s.

With any gold rush, there are two ways to capture some of the gains society will be presented with – you can sell pickaxes, or learn to swing them.

With research moving fast, there are incredibly opportunities to build great products and experiences. If the pace of progress in the last few years continues, there will be a new Moore’s Law to rely on. Products can be designed without needing to be concerned about today’s constraints, as by the time the product is built and used at scale the models will have surpassed the performance of what was possible early in the product development cycle. There will be lots of work on the research side as well, but R&D efforts will need to transition from being purely research focused – ensuring models play nice with each other and updated models can be used by existing products will be critical. We’ll also need to move away from published papers and citations as a measure of success. Groups that have researchers working hand in hand with product developers, balancing the tradeoffs of exciting research opportunities with delivering value to users, will be best positioned to succeed in the next few years. This will require leaders with a balance of technical acumen and ability to understand their customers. There will be many who try, but few who will succeed at scale.

But for those without the interest or skill to work on building products, there is still the opportunity to become the best creator or user of these new AI products. AI products present new loops and workflows that will take some getting used to. Becoming an expert in the early days of these products will be a surefire way to accelerate one’s growth – whether the use of AI products is known to others or simply an input that increases productivity. In a decade or less, everyone will need to leverage AI products to keep up, so there is only a short period of time where simply being comfortable using AI products will be an advantage. However, there will always be those who can extract more value and productivity gains from the same tool, and building the skillset early will provide lasting value. Working with generative AI products is different than the normal mental model we’ve all developed of working with computers – generative AI is not guaranteed to be correct or deterministic. Learning how to leverage this is a skill in and of itself – like learning how to get value out of a teammate who is incredibly responsive but just makes something up the majority of the time.

I encourage my students at Baruch to use generative AI products. I’m able to do this while the school doesn’t have an official policy yet, but I’m doing my part to help academia evolve to support generative AI products faster than the calculator was adopted in the 20th century. I think everyone needs to start learning to use these products, and, besides encouraging the next generation to use them, in 2023 I’ll be doing my best to help build user focused products that turn these exciting opportunities into real productivity gains and value for users.

Leave a Reply

Your email address will not be published. Required fields are marked *