My Feel the AGI Moments

The point of this post is to track pivotal moments in my grasping the potential of AI:

Timeline

  • Jul 2022: GPT-3
    This is the earliest date that I can find evidence I used OpenAI’s GPT-3 in their playground. GPT-3 was mindblowing. I had, at most, only vaguely heard of GPT-2 so I went from never using LLMs to using GPT-3. Using “base” models with no chat fine-tuning still felt extremely powerful. You could get it to output decent Python code and I used it as an alternative to Googling for Python syntax.

  • Circa Aug–Sep 2022: Stable Diffusion / Midjourney
    I played around with both of these tools. Midjourney’s artistic nature was wild. This was also right around the same time a Midjourney piece won an art competition. People at this point were breaking out their typical luddite takes such as “AI art has no soul.” I remained unconvinced—to me, there was a real skill in leveraging these tools, and I saw how others were far better at creating compelling art than I could. It was more akin to the development of photography. At this point, I became convinced of AI-driven job loss in the creative field long-term. I knew the models needed to be more steerable, but I viewed this as something to be solved in time.

  • Nov 30 2022: ChatGPT
    Like literally everyone in the tech world, I was amazed by ChatGPT. I used it a ton. It was functional enough to help me code at work for basic lookups. I was doing my master’s at the time and definitely used it to aid me on some homework. I remember thinking after graduating in Spring 2023 that I was the last cohort before AI would completely change education. I was right and we now see widespread adoption of AI in education. For now, it seems to be a net negative as kids use it to skip homework or cheat on essays. But I found ChatGPT extremely powerful for learning and I think AI can be a net positive for education. The education sector is just slow moving so it’ll take a while.

  • Mar 14 2023: GPT-4
    The intelligence jump was astounding. To me, this was the first real time I experienced scaling. I’d read scaling papers before and “believed” in the concept, but I never experienced it until then. The famous “big model” smell originates with GPT-4. Sydney in Bing was a complete acid trip.

  • September 2024: o1-Preview
    Man, this was a big moment for me. GPT-4 felt baseline human at best. I remember asking it to try to optimize something for CUDA and it flat out told me “this is beyond my capabilities” and refused to try. My friend does Capture the Flag cybersecurity challenges. He gave me a challenge and o1-Preview did a crazy good job at identifying the vulnerabilities. In certain domains, o1-Preview was highly intelligent. As opposed to GPT-4, which was competent but not intelligent.

  • April 2025: o3

    What a fascinating model. Fascinatingly Dull. o3 by itself in raw chat doesn’t feel particularly special. Just naive scaling improvements. But inside of ChatGPT? Fascinatingly Thrilling. You can give it a task of throwing in a bunch of data, ask it to figure stuff out, and it’ll go search the web, peruse the files, and it’ll actually just work.