AI is in a bubble, so was the Internet

Seeing beyond the Trough of Sorrow

It seems like we’ve finally reached the “trough of sorrow” in the AI innovation cycle.

Chamath Palihapitiya, for one, can be seen on the All in Podcast becoming increasingly skeptical of the return on investment which AI native companies will be able to generate in the coming years.

Comparisons of NVIDIA to Cisco have been prevalent in the mainstream media, on X, and right here on Substack.

And Eric Newcomer recently expressed his opinion on Molly O’Shea’s Sourcery podcast:

“I hope AI doesn’t go the way of blockchain, where we build out all this infrastructure, and expect the value to accrue on the application layer, but then the application layer never really establishes use cases.”

I’ve expressed my opinion on this topic before, and it reflects the general views of Jamin Ball and Brad Gerstner of Altimeter Capital:

Unfortunately two competing statements can be true:

AI is a major platform / tech shift and will change the world

AI is in a bubble

—Jamin Ball, Clouded Judgment

There are three main reasons for this: 

1. We’re waiting for highly accurate AI agents (though Devin get’s close)

2. We’re waiting for AI products to have effective memory. 

3. The 90/10 problem (accurate 90% of the time, wrong 10%) 

4. We need to solve the data wall // problem

5. Then we need to solve the energy problem.

These are all major roadblocks hindering the continued growth and scaling of artificial intelligence models.

Indeed, the models of the future may not even be LLMs or include transformers. For more on that topic click → here.

AI is a major platform // tech shift and will change the world.

Just look at the last major platform shift, the internet.

(I’m counting both the smartphone mobile and social media tech shifts as being application layer technologies that were enabled by the internet).

1. The World Wide Web was introduced to the public by CERN, in 1991, significantly expanding internet use from closed server universities and government agencies.

2. The introduction of the browser in 1993 by Marc Andreesen, (Mosaic, then Netscape) being particularly easy to use and install, is often credited with sparking the internet boom of the 1990s.

3. The commercialization of the internet accelerates with the launch of websites like Amazon and eBay in 1995.

4. The IEEE approves the 802.11 standard, marking the official creation of Wi-Fi, in 1997.

5. Then the dot.com bubble bursts in the early 2000s due to:

   1. The overvaluation of companies: with investors ignoring traditional financial metrics and fundamentals. Companies were valued based on speculative future earnings rather than actual profitability.

   2. Drying up of capital: As the bubble reached its peak, the flow of easy capital that had been fueling the rapid growth of internet startups began to dry up

   3. Rising interest rates: The Federal Reserve raised interest rates several times in 1999 and 2000, making speculative investments less attractive compared to interest-bearing assets like bonds

   4. Lack of sustainable business models: Many dot-com companies lacked viable long-term business plans and were unable to generate profits, relying instead on investor capital to stay afloat.

   5. Shift in investor sentiment: As the bubble began to deflate, investor confidence in the tech sector rapidly eroded, leading to a self-reinforcing cycle of selling and further devaluation

(Stop me when this sounds familiar 😊)

AI is a major platform // tech shift and will change the world

AI is in a bubble

But as we all know, the internet continued to grow in the early 2000s, with the rise of social media and eventually, online businesses.

6. In 2001, Wikipedia is launched, followed by social media platforms like LinkedIn (2002) and Facebook (2004).

7. In 2006, the BT Group establishes Openreach in the UK to improve broadband competition and infrastructure.

8. In 2008, the demand for internet data skyrockets, leading to the introduction of fiber-optic broadband, offering significantly faster speeds.

9. By 2010, the internet becomes essential in everyday life, with advancements in home broadband speeds and the introduction of 4G mobile internet.

10. By 2015, Broadband is recognized as a utility in the UK, highlighting its importance alongside water and electricity.

11. By 2017//2018 the rollout of full-fiber broadband begins, offering ultrafast speeds of 100Mbps and higher, with some connections reaching 1Gbps.

The internet was a major platform // tech shift that changed the world

The internet was in a bubble

The internet took 30 years to become ubiquitous

So if all that is true, what is the road map to AI becoming ubiquitous?

The Five Tiers

Conversational AI (Tier 1)

This stage encompasses current AI technologies like ChatGPT, which can engage in basic conversational interactions with users.

Examples:

• Chatbots that assist with customer service.

• Virtual assistants like Siri or Alexa.

• LLMs with graphic user interface designs like Claude, Perplexity, and Llama.

Reasoners (Tier 2)

AI systems at this stage can solve problems comparable to a highly educated human, such as someone with a PhD. These systems can handle more complex problem-solving tasks.

Examples:

• AI capable of solving advanced mathematical problems.

• Systems that can analyze and interpret complex scientific data.

• We are approaching this level, working on enhancing the reasoning capabilities of AI models.

• See my recent article on Strawberry // Q* → here for more.

Agents (Tier 3)

These AI systems can perform tasks on behalf of users over extended periods, demonstrating a higher degree of autonomy and task management.

Examples:

• Personal AI assistants that could manage schedules, book appointments, and handle various tasks without constant supervision.

• AI systems that can autonomously conduct research projects.

  
  Steps to Reach This Level: To transition to this stage, we need to develop AI models that can maintain context over longer periods and manage more complex, multi-step tasks autonomously. New architectures, other than LLMs may be needed.

Innovators (Tier 4)

At this stage, AI could contribute to significant innovations, such as developing new technologies or aiding in scientific discoveries.

Examples:

• AI systems that could design new drugs or materials. (Protein folding models like Alpha Fold)

• AI that could generate novel solutions to engineering problems.

• AI that that could ingest everything we know about physics and mathematics and create novel theories which bridge quantum mechanics and general relativity.

  
  Steps to Reach This Level: Achieving this stage would require AI to not only perform tasks autonomously, but also to exhibit creativity and innovation, generating new ideas and solutions independently.

  For me, personally, this level would indicate “AGI”, as a generalized intelligence would be necessary to create new, good explanations for phenomena that are not yet fully explicable.

Organizations (Tier 5)

The final stage involves AI systems capable of performing tasks equivalent to those of an entire organization.

Examples:

• AI managing all aspects of a company's operations, from logistics to human resources.

• AI systems that can run large-scale projects or enterprises autonomously.

• Think “smart AI” from the science fiction Halo franchise.

  

  Cortana, a “smart” AI

  
  Steps to Reach This Level: Reaching this level would necessitate AI systems that are not only innovative and autonomous, but also capable of coordinating complex activities across multiple domains. They would have to be capable of functioning as a highly efficient and intelligent organizational entity.

Current Status and Next Steps

We are currently at the Conversational AI stage (Tier 1) and are nearing the Reasoners stage (Tier 2).

To move from Reasoners to Agents, we would need to:

• Enhance the contextual understanding and memory capabilities of our AI systems.

• Develop more sophisticated task management and autonomy in our AI systems.

• Ensure that AI can handle complex, multi-step tasks over extended periods without human intervention.

And to do those things we need to:

1. Solve the data wall // problem.

2. Solve the energy problem.

For more on these problems, see my recent article about what comes after LLMs.

I see a lot of people, both on Substack, on X, and in person, bemoaning:

“How long is this going to take?”

“This better not be Blockchain 2.0”.

“Delayed until the fall?”

First of all, Blockchain is still in its infancy, and will prove to be a consequential technology in its own right.

Second of all, you don’t want this technology to change the world over night. That’s not a good scenario. You want this technology to be iterative.

You want it to take 30 years. You want it to be the internet. And then the browser. And then the mobile phone with 4G. And then high speed broadband.

You don’t want to wake up one morning with super-intelligent AI. Society needs time to adjust.

Eventually, we will have to regulate this technology. I don’t think we’re there just yet, but the technology is here to stay.

So take a breath. Telescope out. Understand that AI is in a bubble, and that AI will change the world in unrecognizable ways.