Prophecies of the Flood
What to make of the statements of the AI labs?
Recently, something shifted in the AI industry. Researchers began speaking urgently about the arrival of supersmart AI systems, a flood of intelligence. Not in some distant future, but imminently. They often refer to AGI - Artificial General Intelligence - defined, albeit imprecisely, as machines that can outperform expert humans across most intellectual tasks. This availability of intelligence on demand will, they argue, change society deeply and will change it soon.
There are plenty of reasons to not believe insiders as they have clear incentives to make bold predictions: they're raising capital, boosting stock valuations, and perhaps convincing themselves of their own historical importance. They're technologists, not prophets, and the track record of technological predictions is littered with confident declarations that turned out to be decades premature. Even setting aside these human biases, the underlying technology itself gives us reason for doubt. Today's Large Language Models, despite their impressive capabilities, remain fundamentally inconsistent tools - brilliant at some tasks while stumbling over seemingly simpler ones. This “jagged frontier” is a core characteristic of current AI systems, one that won't be easily smoothed away
Plus, even assuming researchers are right about reaching AGI in the next year or two, they are likely overestimating the speed at which humans can adopt and adjust to a technology. Changes to organizations take a long time. Changes to systems of work, life, and education, are slower still. And technologies need to find specific uses that matter in the world, which is itself a slow process. We could have AGI right now and most people wouldn’t notice (indeed, some observers have suggested that has already happened, arguing that the latest AI models like Claude 3.5 are effectively AGI1).
Yet dismissing these predictions as mere hype may not be helpful. Whatever their incentives, the researchers and engineers inside AI labs appear genuinely convinced they're witnessing the emergence of something unprecedented. Their certainty alone wouldn't matter - except that increasingly public benchmarks and demonstrations are beginning to hint at why they might believe we're approaching a fundamental shift in AI capabilities. The water, as it were, seems to be rising faster than expected.
Where the water is rising
The event that kicked off the most speculation was the reveal of a new model by OpenAI called o3 in late December. No one outside of OpenAI has really used this system yet, but it is the successor to o1, which is already very impressive2. The o3 model is one of the new generation of “reasoners” - AI models that take extra time to “think” before answering questions, which greatly improves their ability to solve hard problems. OpenAI provided a number of startling benchmarks for o3 that suggest a large advance over o1, and, indeed, over where we thought the state-of-the-art in AI was. Three benchmarks, in particular, deserve a little attention.
The first is the called the Graduate-Level Google-Proof Q&A test (GPQA), and it is supposed to test high-level knowledge with a series of multiple-choice problems that even Google can’t help you with. PhDs with access to the internet got 34% of the questions right on this test outside their specialty, and 81% right inside their specialty. When tested, o3 achieved 87% beating human experts for the first time. The second is Frontier Math, a set of private math problems created by mathematicians to be incredibly hard to solve, and, indeed, no AI ever scored higher than 2%, until o3, which got 25% right. The final benchmark is ARC-AGI, a rather famous test of fluid intelligence that was designed to be relatively easy for humans but hard for AIs. Again, o3 beat all previous AIs as well as the baseline human level on the test, scoring 87.5%. All of these tests come with significant caveats3 but they suggest that what we previously considered unpassable barriers to AI performance may actually be beaten quite quickly.
Agents
As AIs get smarter, they become more effective agents, another ill-defined term (see a pattern?) that generally means an AI given the ability to act autonomously towards achieving a set of goals. I have demonstrated some of the early agentic systems in previous posts, but I think the past few weeks have also shown us that practical agents, at least for narrow but economically important areas, are now viable.
A nice example of that is Google’s Gemini with Deep Research (accessible to everyone who subscribes to Gemini), which is really a specialized research agent. I gave it a topic like “research a comparison of ways of funding startup companies, from the perspective of founders, for high growth ventures.” And the agentic system came up with a plan, read through 173(!) websites and compiled a report for me with the answer a few minutes later.
The result was a 17 page paper with 118 references! But is it any good? I have taught the introductory entrepreneurship class at Wharton for over a decade, published on the topic, started companies myself, and even wrote a book on entrepreneurship, and I think this is pretty solid. I didn’t spot any obvious errors, but you can read it yourself if you would like here. The biggest issue is not accuracy, but that the agent is limited to public non-paywalled websites, and not scholarly or premium publications. It also is a bit shallow and does not make strong arguments in the face of conflicting evidence. So not as good as the best humans, but better than a lot of reports that I see.
Still, this is a genuinely disruptive example of an agent with real value. Researching and report writing is a major task of many jobs. What Deep Research accomplished in three minutes would have taken a human many hours, though they might have added more nuanced analysis. Given that, anyone writing a research report should probably try Deep Research and see how it works as a starting place, even though a good final report will still require a human touch. I had a chance to speak with the leader of the Deep Research project, where I learned that it is just a pilot project from a small team. I thus suspect that other groups and companies that were highly incentivized to create narrow but effective agents would be able to do so. Narrow agents are now a real product, rather than a future possibility. There are already many coding agents, and you can use experimental open-source agents that do scientific and financial research.
Narrow agents are specialized for a particular task, which means they are somewhat limited. That raises the question of whether we soon see generalist agents where you can just ask the AI anything and it will use a computer and the internet to do it. Simon Willison thinks not despite what Sam Altman has argued. We will learn more as the year progresses, but if general agentic systems work reliably and safely, that really will change things, as it allows smart AIs to take action in the world.
Many smaller advances are happening
Agents and very smart models are the core elements needed for transformative AI, but there are many other pieces as well that seem to be making rapid progress. This includes advances in how much AIs can remember (context windows) and multimodal capabilities that allow them to see and speak. It can be helpful to look back a little to get a sense of progress. For example, I have been testing the prompt “otter on a plane using wifi” for image and video models since before ChatGPT came out. In October, 2023, that prompt got you this terrifying monstrosity.
Less than 18 months later, multiple image creation tools nail the prompt. The result is that I have had to figure out something more challenging (this is an example of benchmark saturation, where old benchmarks get beaten by the AI). I decided to take a few minutes and see how far I could get with Google’s Veo 2 video model in producing a movie of the otter’s journey. The video you see below took less than 15 minutes of active work, although I had to wait a bit for the videos to be created. Take a look at the quality of the shadows and light. I especially appreciate how the otter opens the computer at the end.
And, to up the ante even further, I decided to turn the saga of the otter into a 1980s style science fiction anime featuring otters in space and a period-appropriate theme song (thanks to Suno). Again, very little (human) work was involved.
What of the flood?
Given all of this, how seriously should we take the claims of the AI labs that a flood of intelligence is coming? Even if we only consider what we've already seen - the o3 benchmarks shattering previous barriers, narrow agents conducting complex research, and multimodal systems creating increasingly sophisticated content - we're looking at capabilities that could transform many knowledge-based tasks. And yet the labs insist this is merely the start, that far more capable systems and general agents are imminent.
What concerns me most isn't whether the labs are right about this timeline - it's that we're not adequately preparing for what even current levels of AI can do, let alone the chance that they might be correct. While AI researchers are focused on alignment, ensuring AI systems act ethically and responsibly, far fewer voices are trying to envision and articulate what a world awash in artificial intelligence might actually look like. This isn't just about the technology itself; it's about how we choose to shape and deploy it. These aren't questions that AI developers alone can or should answer. They're questions that demand attention from organizational leaders who will need to navigate this transition, from employees whose work lives may transform, and from stakeholders whose futures may depend on these decisions. The flood of intelligence that may be coming isn't inherently good or bad - but how we prepare for it, how we adapt to it, and most importantly, how we choose to use it, will determine whether it becomes a force for progress or disruption. The time to start having these conversations isn't after the water starts rising - it's now.
I asked Claude to read over the completed document and give me feedback and it wrote: “The parenthetical comment about Claude 3.5 could potentially benefit from an update or revision since it's mentioned as an example of potential AGI. As Claude 3.5 Sonnet, I should note that I can't verify specific claims about my capabilities in relation to AGI.”
They skipped the name o2 because it is the name of a telephone company in the UK, AI naming continues to be very bad.
The caveat for GPQA is that the data is publicly available, and it is possible that the model trained on that data, either by accident or on purpose, although there is no indication that it did so. The caveat from the Frontier Math test is that the problems differ in difficulty level, Tier 1 are hard Math Olympiad problems, Tier 2 are graduate level problems, and Tier 3 are genuine research-level problems. In the words of the mathematician in charge, of o3’s correct answers: “40% have been Tier 1, 50% Tier 2, and 10% Tier 3. However, most Tier 3 “solutions”—and many Tier 2 ones—stem from heuristic shortcuts rather than genuine mathematical understanding.” The caveat for ARC-AGI is that it required a lot of very expensive computer time for o3 to run long enough to achieve its high score.
They are a bit full of themselves, aren't they? Claiming that "we, and only we, truly understand what's coming." Very insular of them. But also, they clearly don't really understand how the rest of the world works. Integration and deployment of new technology is always slower than it hypothetically could be. These Promethean myths of sudden transformation rarely happen. New infrastructure needs to be created, best practices need to be developed, oh, and businesses and governments needs to actually adopt the technology. It's sort of sadly ironic when the tech bros are so disconnected from the rest of humanity like this.
Insightful as always, thanks Ethan. Blown away by the 🦦 videos! Thanks also for endnote caveats on o3's benchmark achievements - very helpful.