Home

The New Microchip

February 15, 2024

In 1959, Robert Noyce invented the microchip at Fairchild Semiconductor, ushering in the computer revolution that has been continuously unfolding ever since. The microchip made it possible to build smaller, cheaper, more powerful computers, paving the way for a generation of new consumer electronics devices - in the beginning, simple products like calculators and digital watches, eventually much more consequential technologies like the PC and the smartphone.

Nine years later, Bob Noyce and Gordon Moore went on to found Intel, Fairchild’s spiritual successor. Intel’s chips were an integral part of the personal computing revolution. They could be found in the Apple Macintosh and IBM PCs, in cloud data centers, and in all manner of other consumer electronics.

Gordon Moore famously predicted that the number of transistors in an integrated circuit would double roughly every two years - Moore’s Law. Thanks largely to his own efforts, this prediction came true, and computing power has improved at an exponential rate for the last fifty years, enabling applications today that would have been unthinkable a few decades ago.

One such application is the neural network. In Yann LeCun et al.’s seminal 1989 paper on backpropagation, which represents the practical birth of deep learning, they trained a tiny ~1,000 parameter neural network for handwritten digit recognition. The training took three entire days to train on a state-of-the-art workstation. Today it takes less than 2 minutes on a Macbook Air.

As much as anything else, Moore’s law paved the path from these tiny toy neural networks to today’s large language models.

I see a lot of parallels between the microchip and the LLM:

  1. I expect that LLMs and deep neural networks writ large will probably improve every single year - perhaps exponentially - just like microchips did. Just as compute got cheaper and more powerful, so will intelligence.
  2. I expect that LLMs will be a key ingredient in a generation of new consumer applications, just like microchips were a key ingredient for personal computers and smartphones.
  3. I expect, however, that the companies who invent those applications will not be the same companies who build the LLMs, just as Intel did invent the PC.
  4. I expect that building this next generation of applications will be largely a matter of creating better user experiences and helping average people to integrate this powerful technology deeply into their lives and their work, just as PCs helped average people harness the power of computers.

I think that point #4 represents a massive opportunity, and that the next crop of generational technology companies will be built by solving this problem.

To extend the analogy between microchips and LLMs, I think ChatGPT is to the LLM as the calculator was to the microchip - merely the first, simple consumer application of a technology that will pale in comparison to future applications like the PC. The PC for the LLM era has not yet been built.

ChatGPT, like a pocket calculator, is profoundly useful and a massive upgrade over what was previously available, but it’s also very limited if you think about it:

  • It only speaks when spoken to. It will never proactively do anything for you. It won’t summarize what you’ve missed at work while you were on vacation. It won’t sit on the phone with your utility company for you. It won’t pay bills on your behalf.
  • It doesn’t have access to data about your life and work, at least not by default. We’re starting to solve this with tools, function calling, and retrieval augmented generation (RAG), etc. but these are fairly limited and mostly just available to developers who program their own assistants.
  • It can’t truly learn over time and take into account large amounts of contextual information in the fluid way that a human can. RAG, fine-tuning, and enlarging context windows aren’t quite an adequate solution here. A human interlocutor doesn’t need any of these tricks to remember its last interaction with you. We just intuitively recall all relevant details because our neurons update in realtime.
  • It can only do one thing at a time. One of the most important features of modern computers is that they are multithreaded. You can listen to Spotify and check your email at the same time. Why shouldn’t your LLM assistant be multithreaded too? Shouldn’t it be simultaneously responding to your emails, managing your finances, and helping to get you that elusive reservation at Polo Bar?

So, to summarize, the next evolution of AI assistants needs to be:

  • Proactive
  • Integrated with external systems
  • Contextually aware
  • Multithreaded

This isn’t an exhaustive list, of course. Portability and multimodality are helpful too, which is why we’re seeing an explosion of new wearable AI devices like Rabbit R1 and the Humane AI Pin that you can talk to as you go about your daily life.

All of these attributes are even more important in a B2B or enterprise setting. In fact, I suspect that the upside for LLMs in B2B applications is even greater than the upside in consumer applications.

Giving your employees a creative assistant like Github Copilot or ChatGPT is all well and good, but the human being at the keyboard is still the rate limiting factor. The real value will be created by embedding LLMs deep inside of your business’s digital infrastructure and allowing them to proactively perform thousands of actions in the background.

Whenever logical processes of thought are employed - that is, whenever thought for a time runs along an accepted groove - there is an opportunity for the machine.

- Vannevar Bush, As We May Think, July 1945

Every business in America is full of processes that are formulaic enough that they can now be done by an LLM, and for formulaic processes, an LLM is a perfect fit. Their precise instructions can be tweaked and introspected. Their outputs can be conformed to the correct format. And, perhaps most importantly, they are lightning fast and dirt cheap compared to humans.

But an assistant like ChatGPT is not the right vehicle for automating these formulaic processes. We need something more - something that is proactive, integrated with external systems, contextually aware, and multithreaded.

I believe that solving these problems can unlock the power of AI for the American economy, the same way that graphical user interfaces unlocked the power of computers and web browsers unlocked the power of the internet. The company that invents these new capabilities has the potential to transform the business world and reap huge rewards.

And, as mentioned, this company is unlikely to be Open AI or Mistral or Google or Microsoft, because history has shown that the company that creates the platform isn’t usually the company that builds the killer app on the platform.

  • Intel built the microchip, but they didn’t build the PC (Apple, IBM, and Microsoft did)
  • Oracle built the database, but they didn’t build the CRM (Salesforce did)
  • Apple built the smartphone, but they didn’t build ridesharing (Uber did)

So that’s what I’ll be working on building. We’ll see what happens.