10 AI Predictions for 2030

Predictions are hard. Especially about the future.

- Yogi Berra

I first drafted this essay about four months ago. A lot has happened since then. Here is a list of new AI models and products that have been released in the last 120 days alone:

• DeepSeek r1
• OpenAI o3 and o4-mini
• GPT-4.1, GPT-4.1 Mini, and Nano
• Gemini 2.5 Pro and Flash
• Claude 3.7 Sonnet
• Deep Research (from OpenAI, Perplexity, and Google)

The models are consistently getting better and cheaper, achieving ever higher scores on benchmarks and enabling new and exciting use cases.

This progress is fueled by a deluge of investment both at the infrastructure layer, with huge projects like Stargate, as well as the application layer, where venture capital is pouring money into AI startups (like mine!) and minting new unicorns each week. The pace of innovation and the frenzy of investment is unlike anything I've ever seen before.

Some of the developments of the past few months would seem to refute the predictions I make in this essay. But I’ve (mostly) resisted the temptation to revise them. When we forecast the future, we tend to overweight recent events, and I’m trying to resist that recency bias.

Yet the last few months have served as a clear reminder that no one can predict the future of AI. The field is moving so quickly that even experts - perhaps especially experts - cannot confidently say what the world will look like five years from now. No one can predict the future.

But that won’t stop me from trying.

Here are my ten predictions for AI over the next five years.

1. Gains in Intelligence will be Hard Fought

Many have predicted that LLM intelligence will increase exponentially, much like the size of transistors has decreased exponentially under Moore’s law.

Leopold Ashchenbrenner is a notable proponent of this view, having caught the attention of White House insiders like Ivanka Trump. In his now famous essay, Situational Awareness, he writes:

AGI by 2027 is strikingly plausible. GPT-2 to GPT-4 took us from ~preschooler to ~smart high-schooler abilities in 4 years. [...] We should expect another preschooler-to-high-schooler-sized qualitative jump by 2027.

I’m not so optimistic.

LLMs will no doubt get smarter each year, and they will achieve progressively better scores on benchmarks like Arc AGI and Humanity’s Last Exam.

But the notion that the progress will be as exponential as the increases in processing power that came from Moore’s law seems unlikely to me. I expect that it will become more and more challenging to eke out gains in intelligence. I also expect that progress on contrived benchmarks will not always translate to improved performance on real world tasks.

If you use LLMs often, you may already have begun to see this divergence. The reasoning models that score best on various benchmarks, like o3, are often not as practically useful as “dumber” models like Claude Sonnet. A recent paper titled The Leaderboard Illusion describes this phenomenon well.

Gains in intelligence will be hard fought for a simple reason: It’s easy to train a model on what’s already been done. It’s much harder to train a model on things that haven’t been done before.

So far it has been fairly straightforward to create models that are as smart as humans because there is abundant training data created by humans.

But making models smarter than the smartest humans will be much more difficult. What training data can we feed them? Who would produce it?

My prediction here is that models will continue to improve - they will not necessarily "hit a wall” - but that the gains will be increasingly hard fought, and that the most optimistic predictions, like Aschenbrenner’s, will turn out to be wrong.

2. Gains in Speed & Efficiency Will Come Easily

I am much more optimistic about improvements in speed and cost efficiency. I predict that the cost of intelligence will fall rapidly over the next five years. I expect LLMs to follow a trajectory like Moore’s law in this respect.

This trend has already been happening since 2022, and I expect it to continue unabated:

(Note that this is a logarithmic scale!)

I predict that this trend will continue apace, thanks to…

• Improvements in chip design, like Nvidia’s Blackwell GPU architecture
• Algorithmic improvements, like mixture-of-experts and more efficient, sub-quadratic attention mechanisms
• More data centers and power plants, leading to increased energy efficiency and market competition

Just as the cost of compute has been going to zero since Gordon Moore established Moore’s law in 1964, so will the cost of intelligence drop precipitously, enabling all manner of exciting new applications that would be unimaginably costly today.

3. Neither Open Source Nor Closed Will “Win”

There is an ongoing contest between closed AI - represented (ironically) by OpenAI and Anthropic - and open source AI - championed by Meta, Mistral, and more recently DeepSeek.

I predict that the only real winner in this contest will be consumers, who will have a more diverse, competitive ecosystem to choose from. Open-source and properietary models will coexist and specialize in different applications.

This is quite a lot like the OS wars of the 80s and 90s, when Mac, Windows, and Linux all competed to be the dominant operating system of the day. No one really won - all three operating systems are still used today in different contexts.

MacOS attracted consumers who enjoyed the tight integration of OS and hardware. Windows won consumers and businesses who wanted the freedom to choose whatever hardware they wanted. And Linux became the dominant OS for servers (and neckbearded hackers).

I expect LLMs will follow a similar path. For consumer applications, no one is going to run their own open source LLM, no matter how much they care about privacy. They’ll simply consult ChatGPT, Claude, or Gemini.

Businesses will choose between open source and proprietary models depending on their needs. If privacy is paramount, they might run open source models on-prem. For businesses that don’t care or don’t have sufficient scale to justify owning their own compute, proprietary models will make more sense.

What I hope and predict is that both models are allowed to coexist and compete with one another and that neither dominates. This will be better for consumers in the long run.

4. LLMs are Not the Final Architecture

So far we’ve talked about LLMs as if they are the only model architecture - and for now, they might as well be. But I expect that the biggest breakthrough of the next decade will be a new model architecture that combines the linguistic intelligence of LLMs with visual-spatial intelligence.

I don’t think we can reach AGI - whatever that means - with LLMs alone. Even if LLMs surpass the intelligence of the smartest humans, they’ll still be stuck in a box, unaware of the physical world and unable to interact with it.

The next crucial breakthrough, I predict, will be a new model architecture that can do both. I suspect this development might come from a new research lab like Fei Fei Li’s World Labs, whose mission is “to lift AI models from the 2D plane of pixels to full 3D worlds - both virtual and real - endowing them with spatial intelligence as rich as our own.” Or it might very well come from Tesla, whose self-driving unit already leads the world in spatial AI.

I can’t say for certain that this will happen by 2030. It might, but it also wouldn’t surprise me if it took much, much longer. What I do feel strongly about, though, is that when it does happen, it will be the most important breakthrough in model architecture since the transformer.

5. Google’s Search Monopoly Gradually Erodes

The sun has passed high noon on Google’s Search Monopoly. (Is that a phrase? I'm not sure. It sounds like it should be.) I predict Google will eventually lose market share to ChatGPT and other LLM assistants like Perplexity and Claude.

I base this prediction chiefly on my own habits as a consumer. I rarely use Google search anymore. Roughly 75% of my search queries go directly to ChatGPT or Claude. For another 15% where up-to-date information is crucial, I go to Perplexity (although less and less, now that ChatGPT has search built in). The remaining 10% go to Google, mostly for location-based queries, sports scores, or local movie times.

OpenAI seems laser focused on becoming a consumer tech company that can go toe-to-toe with Google, and so far they are succeeding. ChatGPT usage is currently growing 30% quarter-over-quarter, and they are pouring resources into new features like improved memory, image generation, and deep research. Sam Altman recently said that Open AI aspires to be "people's core AI subscription".

Google, on the other hand, confronts the quintessential innovator’s dilemma. To compete in the new chatbot market, they risk cannibalizing their existing search ads business. Even though Google’s LLMs are excellent and very much on par with Open AI's, Gemini has a small fraction of ChatGPT’s user base.

But Google’s dominance will not disappear overnight. They could coast on the momentum of their search business - the most profitable business unit in history - for many years and be just fine. In fact that’s probably what they will do.

But over time, their monopoly will erode. In 20 years, the search advertising industry will be a small fraction of what it is today, and Google will need to invest in new products if they wish to maintain their dominant status.

6. No More Foundation Model Companies

I predict that there will be no new foundation model companies of any significance founded between now and 2030. The capital requirements are simply too great, and the market can only support a handful of them.

The key players that exist today, along with their capital partners - OpenAI/Microsoft, Anthropic/Amazon, Google/DeepMind, Meta, DeepSeek, and Mistral - will continue to duke it out for at least the next five years, probably much longer. I would not be surprised to see more consolidation (hard to imagine Mistral surviving for 10 more years without being acquired).

If I were a venture capitalist I would not be eager to invest in new foundation model companies. Instead, I would look for returns in the application layer, which leads us to our next point.

7. Cambrian Explosion of Application Layer Companies

While the race in the model layer is mostly over, the race in the application layer is only just getting started. I predict that the second half of the 2020s will see a Cambrian explosion of new application layer companies that leverage AI in exciting new ways.

I predict that most of the action here will be in the B2B space, not the B2C space. The B2C space might see some exciting new applications - like AI girlfriends or AI personal trainers - but they will look relatively small compared to the B2B applications.

The knowledge economy is teeming with opportunities to enhance productivity and automate workflows using AI. This opportunity will be seized, I predict, primarily by new companies that don’t exist yet or are only now in their infancy.

The next Salesforce, Adobe, Shopify, Intuit, and Hubspot - these are all waiting to be built. (My company, Quotient, will be one of them. Just wait and see.)

8. The US Will Lead, with China Close Behind

In January, DeepSeek R1, a powerful and shockingly cheap reasoning model from a Chinese lab, had many people wondering if China was catching up to the US in AI dominance. Around the same time, President Trump signed an executive order titled “Removing Barriers to American Leadership in Artificial Intelligence”, which stated “it is the policy of the United States to sustain and enhance America's global AI dominance in order to promote human flourishing, economic competitiveness, and national security”.

The competition has begun, and the two most powerful nations on earth both understand the importance of AI supremacy to their economic and military stature.

I predict that America will win this competition, but I’m betting it will be close. Xi Jinping has long lamented China’s status as a laggard in the semiconductor industry, and he has been trying desperately to catch up by subsidizing domestic firms like Huawei, stealing US IP, and pushing for supply chain self-sufficiency.

Xi won’t let China fall behind again. To China’s credit, breakthroughs like DeepSeek R1 prove that China has the talent to go toe to toe with America in this new arms race. We shouldn’t underestimate them.

And yet I think that America will ultimately win. Our ecosystem is simply more competitive, our talent pool is deeper, our culture is more innovative, and our markets are freer, less bogged down by state control. I also predict that China will be consistently hamstrung by export controls that constrict its access to GPUs.

I predict that in 2030, American AI companies will still be the most important and most successful on the planet, and that we will lead the world through yet another technological revolution.

The rest of the world, I expect, will be mostly irrelevant in this competition. The EU is too overregulated and lacks the political will to do anything. Everywhere else lacks the capital and infrastructure.

9. The Economy Will Change Only Slowly

Bill Gates famously said “Most people overestimate what they can do in one year and underestimate what they can do in ten years.”

Similarly, most people overestimate how much the world can change in one year, but underestimate how much it can change in twenty.

This will be true of AI. I expect that by 2030, we will look back and feel somewhat underwhelmed by the degree to which AI has changed the world economy.

In a recent interview, Dwarkesh Patel asked Tyler Cowen, “Why won't we have explosive economic growth, 20% plus, because of AI?”

Cowen’s response summarizes my more measured view well:

It's very hard to get explosive economic growth for any reason, AI or not. [...] That will take, say, 30 years. So you'll have some sectors of the economy, less regulated, where it happens very quickly. But that only gets you a modest boost in growth rates, not anything like the whole economy grows 40% a year.

Even a technology as miraculous as AI takes a long time to permeate the economy. Even the most forward thinking companies move at glacial pace. What about more risk averse, regulated industries? It might take a decade before doctors are allowed to use LLMs in their daily work (which is unfortunate, and will result in many preventable deaths).

So don’t hold your breath. By 2030 I predict that AI adoption will have grown considerably, and that it will modestly boost economic growth. But for the most part we will feel disappointed by the pace of change. Many of us will wish that we could use more AI in our jobs, and we’ll often ask “Shouldn’t AI be able to do this for me?”

10. AI Will Benefit the Smartest and Richest Most

Consider two possibilities:

1. AI makes cognitive ability economically irrelevant, because everyone has access to the same superhuman intelligence. In the same way that physical strength used to be an economic asset when most people worked in the fields, cognitive ability will cease to be an economic asset, rendered obsolete by AI.

2. AI proves most valuable to the smartest people who know how to wield it best. It acts as a multiplier on human intelligence. The smartest, most productive people are made smarter and more productive. Their cognitive economic advantages are magnified.

Which scenario seems more likely to you?

I’m betting the farm on scenario #2.

I’m betting that AI is not a great economic equalizer but rather an accelerant to inequality.

As is well known, inequality has been increasing steadily in America since 1960. Between 1960 and 2020, the Gini coefficients for wealth and income inequality increased by 42% and 20% respectively.

Experts argue about the cause of rising inequality - some say it’s that tax policies favor the rich and big corporations, others say it’s that globalization shipped all of the good middle class jobs overseas.

I’ve always felt that the real reason is that we live increasingly in a world where the smartest and most capable people can use computers and the internet to automate a great deal of labor that once provided gainful employment for a thriving middle class. The more we automate, the less work there is for anyone to do but the very smartest among us - the cognitive elite.

Unfortunately I expect AI will only exacerbate this trend. Think of all of the lower level white collar workers who can be put out of a job by AI. Call center employees, data analysts, paralegals. Anyone who does repetitive, rote cognitive work.

Fortunately, there is a glimmer of hope. After all, at one point the vast majority of humans worked on farms. As technology automated farming, humans learned different skills and found employment in other fields of endeavor. Most of us today have jobs that would have been unimaginable to our ancestors. (And, importantly, we also have living standards that are unimaginably superior to anything our ancestors could have imagined.)

In the short term, I predict some degree of job loss and a corresponding increase in inequality. But in the long term, I remain mostly optimistic, as I believe we will find new economic roles for humans and we will achieve ever greater standards of living.