AI Reality Check? Steve Streit and Other Experts Weigh in on What’s Next

Global investment in artificial intelligence surged above $140 billion in 2023 and looks likely to exceed that mark in 2024, according to Edge Delta.

That’s an incredible increase from a few years ago, when AI was experiencing one of its occasional “winters” and funding had all but dried up. Much of the credit goes to large language models like ChatGPT, which broke into the public consciousness in late 2022, but other factors are at play as well.

Investors like Steve Streit, whose venture capital firm backs a number of AI-enabled startups and later-stage ventures, have been tracking the space for years. They agree that “this time is different,” as the saying goes, but caution others to be vigilant — and ready to separate hype from reality. 

Here’s what they think is coming around the corner for artificial intelligence technology and the businesses that use it.

Smaller, More Capable AI Models

Large language models like ChatGPT train on enormous volumes of information, in some cases approaching the scope of the public Internet. 

That’s a mind-boggling realization that both reinforces how amazing the data-processing ability of these models is and emphasizes their real-world limitations. At the moment, model size correlates inversely with output usefulness — that is, models that train on smaller but more focused bodies of data tend to be more effective at doing what they’re asked.

The hope is that as we approach something like artificial general intelligence (AGI), larger models will become more capable. In the meantime, the forefront of AI innovation could happen at smaller scales.

A Chip Efficiency Revolution?

If the term “power usage effectiveness” means nothing to you, don’t feel bad. It means nothing to the vast majority of LLM users right now. But it should, because as an expression of the efficiency of data center operations, PUE defines an important physical limitation of AI capabilities. The closer PUE gets to 1, the more efficient the chips running AI models become.

Right now, we’re making incremental progress at lowering PUEs, but hope remains that we’ll achieve a step change in efficiency that allows AI models to use power far more effectively. That would have huge ramifications for learning and processing speeds, and could accelerate progress toward AGI.

More Purpose-Driven AI Use Cases

The trend toward smaller AI models supports a broadening and diversification of AI use cases that should help achieve incremental improvements in the technology’s perceived utility. The importance of this can’t be overstated, as the vast majority (77%) of people who use AI for work report that it makes their jobs less productive — clearly, an unacceptable situation.

How fast this happens and how noticeable the improvements it enables are the multi-billion dollar questions. But it’s likely we’ll see broadly useful AI sooner rather than later.

Practical AI Deployments That Broaden Acceptance and Uptake

Broadly useful AI, in turn, will support wider adoption of the technology. Yes, even among people who “don’t have the time” or express generalized skepticism of its capabilities (or, perhaps, its implications for their own jobs). AI model developers that iteratively solve for real, widely felt pain points will gain a competitive advantage over those focused on sweeping capability improvements.

Significant Progress Toward AGI

Some of the brightest minds in the AI industry believe AGI is right around the corner. Dario Amodei, founder of Anthropic and perhaps the second-best-known AI leader after OpenAI’s Sam Altman, believes that we could achieve human-capable artificial intelligence by the late 2020s. Altman, for his part, has a similar, if slightly less ambitious, view.

Amodei has a “gentler” vision for AI that’s rooted in safety and societal benefit — “that’s been our general aim in the world and part of our theory of change,” he said recently. So his optimism about AGI is notable, and perhaps sobering.

Renewed Focus on AI Safety

Thinkers like Nick Bostrom, whose book Superintelligence sparked wider public awareness of the potential dangers of superintelligent AI models when it hit shelves about a decade ago, believe that the most important thing AI leaders can do right now is ensure their creations don’t threaten human civilization. As we approach something that looks like AGI, this obligation will become impossible to ignore and may heighten tension between the “move fast and break things” types and the “get it right” types.

Should You Believe the AI Hype?

There’s little question that AI is transforming the world as we speak. But reasonable people can disagree over exactly how transformative it will be, how soon its effects will be felt, and how disruptive it will prove in the long run.

“Eventually AI will know us better than we know ourselves,” says AI expert Kai-Fu Lee, who believes AI will radically transform society by the early 2040s. “This will have profound consequences for everything from how we work and play, to how we communicate and learn.”

Lee is an AI bull, but there are plenty of others like him. Even those who aren’t so evangelistic believe AI is among the most important tech innovations since the development of the Internet.

Maybe the best approach is to believe the “big picture” hype while viewing specific claims through a more skeptical lens.