Last week, I attended the Imagination in Action event at the MIT campus in Boston. It was a great experience.

One interesting sentiment I heard was their vision to make Boston an AI hub comparable to San Francisco. A speaker pointed out that Facebook was founded just 2 miles from where we were sitting in the MIT building, but Silicon Valley won out on the long-term headquarters. As an East coast resident, I hope their vision for Boston’s AI growth comes to pass.

The event was spread throughout the Media Lab on the MIT campus, featuring five separate tracks and a startup lobby where attendees could meet startup founders.

This event is unique compared to other conferences I’ve attended due to its rapid-fire sessions. Startup founders were given 90 seconds each to pitch their business, lining up along the side and waiting their turn. Their slides were added to a slide deck in advance so it flowed smoothly. It was exciting to see the diverse ideas and needs being met in various industries by these founders.

The largest room had back-to-back panel (interview) sessions throughout the day, taking 25 minutes each and continuing without breaks. The event did not include the typical 45- or 75-minute prepared talks common at conferences, nor a specific keynote or shared closing sendoff. This makes sense, considering that there were more people (1,800) than could fit into a single room in that building.

I jotted down a few highlights from the panel sessions.


Aravind Srinivas, the co-founder and CEO of Perplexity, joined us for the first panel interview. In just 18 months, Perplexity has grown from the ground up to a Unicorn ($1B) valuation! He broke some startup trends by completing his PhD, compared to the many other successful founders who dropped out of school early.

Perplexity is taking on Google for search, and if you’re not already doing so, you should bookmark and consider using it before many of your Google searches. Even the free version is highly useful.

His vision for the future is to imagine that we can talk to any device in a shopping mall, by a row of dresses, etc., point to something, and ask questions. AI will be everywhere. He also talked about focusing on the smaller 7 billion parameter models, which will be able to fit on mobile devices and be as good or better than the much larger models today. It will take small models for AI to be prevalent everywhere.

Stability AI and Healthcare

Another panel session focused on healthcare. The Stability AI CEO was also there. There is a lot happening in healthcare and medicine. One interesting AI story came from Geoffrey von Maltzahn from Flagship Pioneering, who mentioned the challenge of replicating top pharmaceutical products in every way without taking any existing inside knowledge. With GenAI and modern tech, they accomplished this in 3 months for every single product, a process that would have taken 10 years in the past.

Investing in AI

The next panel was about investing in AI. Dave Blundin from Link Ventures gave a story from another great investor who once said that there are only three things that he looks for when looking to fund a startup:

  1. He will only fund if the founders are best friends.
  2. Do they write the code themselves?
  3. “Do I test them?” {I didn’t capture that note well. I’m not sure what I meant to write}

He said the business plan is last, and the teams are first.

Mark Gorenberg from MIT mentioned that 70% of investments are spent on training and inference.

LLM Futures

Some key points from this session include:

  • “I don’t think it’s the peak yet.”
  • “Are larger models or small ‘fit for purpose’ models going to gain more traction?”
  • Markus Flierl from Intel joked that he likes all models since he’s a hardware guy.
  • Paige Bailey from Google DeepMind talked about the growth of coordination models. These models coordinate with other GenAI models to pick the best-fit model that balances effectiveness, costs, and latency. They will pick the best model for the situation. We’ll see more focus on coordination models.

AI Hardware

I wasn’t expecting to enjoy this session so much. It was great hearing from Dinesh Maheshwari from Groq, the primary panelist. If you’re not familiar with it, check out to see mind-blowing GenAI speed. This is achieved by their specially designed hardware. I briefly posted about GPUs and TPUs last week. GPUs have been the primary processing units used for GenAI, and TPUs are the potential upcoming processing units to replace GPUs. However, Groq’s LPUs (Logical Processing Units) are multiple times faster, as you can easily see by using

Dinesh claims that LPUs are at least 10x more power efficient than GPUs. He says: “Power is the new currency.” LPUs use an assembly line-based architecture, which only touches the data once. It’s not a hub-and-spoke architecture like the others.

A Q&A question came: “Are you worried about Nvidia?”. His answer was a confident No. He said that the scale and increase are very different. Groq’s architecture is many times faster and more efficient than the other architectures. He mentioned that they are just getting started.

Stephen Wolfram

Brilliant mathematician, computer scientist, theoretical physicist, and businessman, Stephen was publishing scientific papers at the age of 15. He’s been a pioneer and early thinker on many of these core topics, so he was engaging.

Some miscellaneous quotes or notes:

  • Referring to LLMs generating original material: “It’s trivial to do something original and creative. Just generate some random numbers. That is original.”
  • “LLM is like a wild animal. You try to harness the power.”
  • “Can it be tamed? I don’t think so. Not with the current transformers architecture.”

Yann LeCun (Joined via webcam)

Another machine learning pioneer and Chief AI Scientist at Meta, Yann, spoke to us on the day Llama 3 was released (for anyone following Llama 3).

Some notes I jotted down:

  • 4 issues with LLMs today
    • Don’t have detailed physical world understanding
    • Don’t have memory
    • Can’t reason
    • Can’t really plan. Can regurgitate plans but can’t plan from scratch

He talked about how progress takes a lot of time, including progress in machine learning and LLMs. It takes about 9 months for babies to learn how to walk, based on many falls, over and over again. He said that we went through the same progress with machine learning for many years and failed. Then, we found what worked with generative AI. Further progress will continue to follow the same pattern.

Next Breakthrough in AI

Some quotes from this panel session:

  • “Important progress will be made by models that can identify their own weaknesses and self-adjust”
  • “Human brain data from multiple sources will create a new layer of applications for brain activity, in alignment with foundation models.”
  • “For startups, the rapid pace of progress is a double-edged sword because you don’t know if what you are working on will be done more easily shortly. Moats are a concern.”

AI After Transformers

This session introduced something completely new to me. The current generative AI is based on the transformer architecture, but it has many limitations. MIT is building a new architecture called Liquid, or Liquid Neural Network (LNN). This competes with the current transformer architecture used by LLMs today. It has many benefits, like being factual with the knowledge that it has and not hallucinating, being much more expressive, it’s easier to guide the model’s decision making, and it uses much smaller models that are less expensive to run.

Part of the active team from MIT was on the panel, and they talked about the progress. It’s still being worked on, but it apparently holds a lot of potential.


As a side, it was packed. There was an overflow room and standing room only for many people.

John Werner, the host, was the obvious central person for this event, and for good reason. He is a unique individual with abundant energy and passion for what he does. Known as an incredible networker and influencer, he has an amazing ability to find and remember people’s strengths. He seemed to know everyone, including numerous fun facts about them. His omnipresence brought a lot of life to the day.

Regarding the conference at MIT, one person mentioned, “Sometimes I didn’t realize until later that I was in the right room, but I didn’t realize it yet.” He was referring to the potential for there to be future influential people in the room, even if we don’t realize how significant the current moment really is.

Generative AI is progressing like crazy right now, so I enjoyed the chance to catch the energy from an event like this. I believe that we’re just getting started. While there may be some hype in the current AI boom, I believe that we’ve hardly scratched the surface of where we’ll be in the next few months and years ahead.

Thank you to MIT and John Werner for hosting a great AI event this week. It’s not the first or the last IIA event, so we should expect to see more of this in the future.

Additional links:


Leave a Reply

Avatar placeholder

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