The path to Artificial General Intelligence (AGI) is likely to differ significantly from common expectations, based on my experience working with large language models over the past four years. Most people I've talked to anticipate that a company like OpenAI or Anthropic will produce a model (perhaps GPT-6 or Claude 4) that achieves AGI-level intelligence. However, I have strong conviction that we'll soon see AGI emerge in specific domains in the application layer rather than the foundational layer.
Let's first define AGI using Sam Altman's example: a future version of ChatGPT that can successfully execute a command like "make me a million dollars." While this definition is widespread, it's important to recognize that it'll probably be many years before we have a foundation model that can achieve the complex, multi-step reasoning process required to go out and autonomously make a million dollars. Darn.
But what I want you to consider is that it's actually possible to build an agent that can do this using only today's foundation models and really good prompt engineering. to illustrate, here's a prototype of an autonomous startup agent that i whipped together a few months back with claude 3 and gpt-4:
A fully-autonomous startup founder could be accomplished by training an agent that specializes in the strategy, design, engineering, and marketing of internet businesses. Given the command to make a million dollars, this agent could work backwards to identify promising markets and opportunities, validate them using landing pages and ads, and then design, build, and scale a product automatically. It's basically a team of autonomous agents: a developer, a marketer, a designer, and a business strategist all working together towards a common goal.
However, the current limitation lies in the high error rate and inability to handle edge cases. We see this challenge even with existing software engineering agents like Devin that can write and execute code. Despite software engineering being a lot easier to solve than building a successful business, Devin and its compatriots still struggle with many edge cases users present.
Two main factors limit our current ability to build AGI: the intelligence of underlying foundational models and the sophistication of prompt engineering techniques. Prompt engineering is more than just conversing with AI; it's about concept engineering – finding unique strings or tokens that evoke expert-level responses from the AI.
For instance, when asking an AI to write a poem about an elm tree without additional context, the result is pretty...generic and average. However, by including a few-shot prompt with a diverse corpus of well-written poems, the AI's output improves tenfold. This improvement occurs because the language model picks up on tacit concepts in the expert prompt, leading to more expert output.
Unfortunately, most AI engineers I've met rely solely on instruction prompting rather than few-shot prompting, often resulting in mediocre or inconsistent outcomes. Advancing prompt engineering involves not only understanding which concepts evoke expert responses but also learning how to distill and transfer expert decision-making processes into a format consumable by language models.
Consider the challenge of building an AGI in medicine – one that goes beyond summarizing papers to simulating new experiments and generating novel hypotheses. Simply waiting for more advanced models like GPT-5 or Claude 4 is unlikely to solve this problem due to the vast amount of context and specialized tools required. Instead, partnering with multiple doctors to understand their decision-making processes and translating this into a detailed agent architecture would be a more effective approach.
This perspective suggests that AGI will not emerge suddenly as a single, all-encompassing model. Rather, we're likely to see pockets of AGI form in specific areas as the next generation of language models (GPT-5 or Claude 4 level) emerges and more companies build intelligent applications around them. Startups focusing on creating businesses or solving medical issues, if designed and implemented correctly, may achieve domain-specific AGI before major AI research companies produce a general AGI.
In conclusion, the path to AGI will likely be gradual and domain-specific, driven by advances in both foundational models and sophisticated prompt engineering techniques, rather than a sudden breakthrough in general artificial intelligence.