Amjad Masad & Adam D’Angelo: How Far Are We From AGI?

Honestly, I don't understand what people are talking about. A year ago, the world was very different. Considering the progress we've made in the last year—with advancements in reasoning models, improved code generation, and video generation—it seems things are accelerating faster than ever. I don't really understand where the bearishness is coming from.

My view is that LLMs are amazing machines, but I don't believe they are equivalent to human intelligence. You can still trick LLMs, even with simple questions. For instance, I recently tweeted about asking, "How many 'R's are in this sentence?" Three out of four models failed to answer correctly. Even GPT-5, with its advanced reasoning, took 15 seconds to respond to such a question. This suggests LLMs possess a different kind of intelligence than humans. They have clear limitations that we often try to work around, whether through adjustments in the LLM itself, its training data, or the surrounding infrastructure, to make them functional.

One concern I have is the potentially harmful effect of large language models (LLMs) on the economy. For instance, LLMs could automate entry-level jobs without impacting expert roles. Consider quality assurance (QA): LLMs are highly effective, yet they still struggle with long-tail events. Consequently, many skilled QA professionals are now managing hundreds of agents, significantly boosting productivity. However, this means fewer new hires, as these agents outperform new human employees. This creates an unusual equilibrium that I believe many people are not considering.

I believe "The Sovereign Individual" offers a consistently strong set of predictions for the future. While it's not a scientific book, but rather a polemic, its ideas originated in the late 80s and early 90s. The authors, two individuals from the UK—I'm unsure if they were economists or political science majors—wrote this book to predict what would happen as computer technology matured. They argued that humanity has experienced the agricultural and industrial revolutions, and we are clearly undergoing another: an information revolution. Now we call it the intelligence revolution, or something similar. I think we won't be able to name it definitively; people in the future will. But we are certainly experiencing a significant transformation.

The terminology of sustaining versus disruptive innovation comes from *The Innovator's Dilemma*. It describes how new technology trends often begin as something akin to a toy, or a product for the lower end of the market. As the technology evolves, it moves up the power curve, eventually disrupting even established incumbents. Initially, incumbents ignore these innovations because they seem insignificant. However, these innovations ultimately disrupt and consume the entire market. This was evident with personal computers (PCs). When PCs emerged, large mainframe manufacturers disregarded them, viewing them as mere toys. Yet, PCs eventually became a massively disruptive force, even powering data centers today.

The main idea was to integrate a verifier into the loop. I recall reading a paper from Nvidia about DeepSeek, which described how they used it to write CUDA kernels. They could run DeepSeek for about 20 minutes if a verifier, such as a test runner, was included. This made me wonder what kind of verifier we could implement. While unit tests are an option, they don't fully capture an application's functionality. So, we began exploring computer vision to test applications. However, computer vision is expensive and still quite buggy. As Adam mentioned, this area needs significant improvement to unlock many applications. Ultimately, we developed our own framework, incorporating various hacks and AI research, to replace computer vision for testing models. I believe it's one of the best.

I believe this technology opens up software's potential to everyone. It's currently underhyped because the tools are still far from what a professional software engineer can achieve. However, I expect these tools to advance significantly within a few years. When they do, anyone will be able to create things that would have previously required a team of 100 professional software engineers. This will massively expand opportunities for everyone. Replit is a prime example, but this trend will also create opportunities beyond just building applications.

The question of consciousness is fundamentally not a scientific one. We've seemingly given up on making it scientific. This is a problem, as all the energy directed towards LLMs means no one is truly contemplating the nature of intelligence or consciousness. There are many core questions. For example, Roger Penrose's "Emperor's New Mind" discusses how the philosophy of mind and the broader scientific community began viewing the brain as a computer. In his book, Penrose argued it's impossible for the brain to be a computer because humans can perform tasks that Turing machines cannot, or get stuck on, such as basic logic puzzles we easily solve.