I want to share a few interesting links that I think show where things are going with AI. The first is called “AI 2027,” and it’s by the not-for-profit AI Futures Project, which has ex-OpenAI people on its staff. This is where they see us in a couple years:

An individual copy of the model, running at human speed, is already qualitatively better at AI research than any human. 300,000 copies are now running at about 50x the thinking speed of humans. Inside the corporation-within-a-corporation formed from these copies, a year passes every week…they are achieving a year’s worth of algorithmic progress every week and will therefore soon be up against the limits of the Agent-4 paradigm.

The Agent 4 Paradigm! The whole thing (it’s long) is written in that Silicon Valley Zero to One style, which is a particular kind of prose I associate with Astral Codex Ten (and indeed, Scott Alexander wrote a long post about AI 2027), Less Wrong, and the like. There’s a lot of arguing from first principles, and a lot of grim stoicism in the face of accelerating change. 

But the more evidence they marshal, the more suspicious I get. And it’s not just me. As Kevin Roose reports in the NY Times:

Ali Farhadi, the chief executive of the Allen Institute for Artificial Intelligence, an A.I. lab in Seattle, reviewed the “AI 2027” report and said he wasn’t impressed. “I’m all for projections and forecasts, but this forecast doesn’t seem to be grounded in scientific evidence, or the reality of how things are evolving in A.I.,” he said.

So that’s one predictive model. But another way to predict the future is to look at what big players are doing and saying. Microsoft’s products, for example, are everywhere (they once put data centers underwater), in every global system that touches money. They also do a ton of consulting work for clients. So they have good predictive models for where things are going when it comes to technology.

And what are they up to? Well, they’re slowing their roll when it comes to data centers. According to Matt O’Brien at the AP:

The tech giant confirmed this week that it is halting early-stage projects on rural land it owns in central Ohio’s Licking County, outside of Columbus, and will reserve two of the three sites for farmland.

So that’s a billion dollars they aren’t spending, and there are more billions they won’t be spending to come. What else is Microsoft up to? AI-enhanced…debugging! If anyone knows about bugs, it’s Microsoft, because their products are some of the buggiest known to humankind. 

But AI is really great at debugging, right? As Samuel Axon writes in the piece linked above, maybe not:

This isn’t the first time we’ve seen outcomes that suggest some of the ambitious ideas about AI agents directly replacing developers are pretty far from reality. There have been numerous studies already showing that even though an AI tool can sometimes create an application that seems acceptable to the user for a narrow task, the models tend to produce code laden with bugs and security vulnerabilities, and they aren’t generally capable of fixing those problems.

Back and forth and back again. Dario Amodei, the CEO of Anthropic, is pretty sure we’re going to solve all the problems ever with AI, and he wrote a big essay about it called “Machines of Loving Grace.” He must be pretty smart, because I use Claude a lot, and it’s good. 

Meanwhile, if you read the Sequoia summary of the market, you’ll learn they see a lot of growth ahead—but I also feel I’ve read the same kind of growth summaries about mobile tech, 5G, or blockchain over the years. 

One last link—over at Google—by Lee Boonstra. It’s called “Prompt Engineering.” You can just skim it. A highlight: 

Using structured JSON as an output is a great solution, as we’ve seen multiple times in this paper. But what about input? While JSON is excellent for structuring the output the LLM generates, it can also be incredibly useful for structuring the input you provide. This is where JSON Schemas come into play.

Here’s how I’m reading the room: The VCs want to see normal, everyday hyper-growth and huge returns for their investors, so they can afford firepits. The AI companies are promising meta-poly-growth, because they’ve set their baseline standard of success as “take over the world.” And the big platform companies have found a new technology that is exciting and will drive incremental growth—but be honest, Google isn’t truly expecting to go from a $2 trillion to $4 trillion company overnight. They’re pulling back a little in some places, pushing forward in others, and pushing their teams to integrate it more deeply into their existing ecosystems and platforms, too.

In the coming months, you’ll see all three trajectories continue. VCs will keep looking for hits—but will also probably start pushing some of their investments to productize and release working products to the market. AI companies and AGI believers will keep getting bolder in their pronouncements of tipping points—not because they are lying, but because they need to keep their beliefs intact in order to justify their past actions. And, finally, giant tech platforms will keep trying to bring everything together, capturing as much value by bringing new tools onto their existing platforms—and minimizing their exposure wherever they can. “Business as usual” is a very powerful concept, especially when times are really, really weird.