# Copilot and ChatGPT revealed distinct AI breakthroughs

Markdown copy of https://realalpha.app/s/copilot-and-chatgpt-revealed-distinct-ai-breakthroughs/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

- Speaker: [Satya Nadella](https://realalpha.app/builders/satya-nadella/) · Satya Nadella is CEO of Microsoft. · Led Microsoft's cloud infrastructure and services transformation, and has been CEO since February 2014.
- Host: [Mathias Döpfner](https://realalpha.app/people/mathias-dopfner/)
- Show: MD MEETS
- Episode: [Microsoft CEO Satya Nadella: AI Boom, Energy Battle & His Complicated Alliance with Sam Altman](https://realalpha.app/v/JapZJVcA1B4/), Nov 29, 2025
- Moment: 0:29 to 4:10 of the episode (4 min), playing on https://realalpha.app/s/copilot-and-chatgpt-revealed-distinct-ai-breakthroughs/; the episode from there: https://realalpha.app/v/JapZJVcA1B4/?t=29

## Quote

> “the lesson is that compute or **log of compute is intelligence**.”
> — [Satya Nadella](https://realalpha.app/builders/satya-nadella/) at [3:32](https://realalpha.app/v/JapZJVcA1B4/?t=212) · MD MEETS

## Takeaways

1. Microsoft Research spent decades on speech and natural language.
2. Copilot’s code completion was an early breakthrough for Nadella.
3. ChatGPT showed how instruction following could make a model conversational.

## About this segment

Satya Nadella traces the Copilot code-completion moment to ChatGPT’s instruction-following breakthrough. He says scaling text models changed what Microsoft believed was possible.

From “Microsoft CEO Satya Nadella: AI Boom, Energy Battle & His Complicated Alliance with Sam Altman”, Nov 29, 2025. [Watch the full episode](https://realalpha.app/v/JapZJVcA1B4/)

## Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

[0:42](https://realalpha.app/v/JapZJVcA1B4/?t=42) that surprised you most about artificial intelligence in the very recent past? Yeah, I think, you know, it's interesting you asked the question the way you asked it, Because in some sense, it is a little bit of, when I even look at my journey in this new generative AI, how we even got to OpenAI, the partnership, how we sponsored them when it was obviously not conventional wisdom, taking the big bets and all of that.

[1:08](https://realalpha.app/v/JapZJVcA1B4/?t=68) It's really, Matthias, having a prepared mind, right? So Bill started Microsoft Research in 1995. You know, his idea was that, look, Xerox PARC was the place that inspired a lot of the companies, whether it was Microsoft or Apple, and we wouldn't have existed. The PC industry wouldn't have existed. So he wanted to create a fundamental research organization. And in fact, the first group I think we had was a speech group.

[1:37](https://realalpha.app/v/JapZJVcA1B4/?t=97) And, you know, we've been at it. And in fact, we've been obsessed about natural language, right? After all, Microsoft is a productivity company. We think about text. We think about natural language, whether it's in speech or in text form. And so, interestingly enough, we have taken many shots on goal on that, and we continuously kept at it. And then when Sam and team, you know, came and said, hey, they're going to apply these transformers to text.

[2:05](https://realalpha.app/v/JapZJVcA1B4/?t=125) In fact, interestingly enough, in 2016, they were much more RL focused, although lately reinforcement learning has come back. Their original focus was more, you know, not as much text. And so in 2019 is when they said, let's really go after, you know, scaling laws of transformer, applied to transformers, applied to natural language. And that was magic, right? I mean, if you sort of say what has played out, and we have added to it a few other techniques over the years.

[2:38](https://realalpha.app/v/JapZJVcA1B4/?t=158) And now I think incrementally even new things that are coming together. But no one had thought of scaling laws applied to basically text and then with instruction following. One of the magical moments, there are two products, GitHub Copilot and ChatGPT, right? GitHub Copilot actually happened first with a codex-like model where we could see next, next, next, right? So, tab, tab, tab, and you could have code completions.

[3:04](https://realalpha.app/v/JapZJVcA1B4/?t=184) And when I saw that, that is when I really felt like, wow, this is crazy, right? This is kind of completing your thoughts, right? right? It's not just spell correction type things, but it is more completing your thought and doing it at speed. And so that was the first breakthrough. Then given the instruction following chat, right? So you could take a model and just teach it a little on how to sort of think about conversations. And it was able to hold a conversation. That was the GPT-3-5 chat GPT moment. And then the rest obviously has followed since then. But I think that's it, right? Which is, if I had to take away, the lesson is that compute or log of compute is intelligence. We obviously need to get a lot more efficient because log is not a fun thing, right? It's not the best way to spend compute to get intelligence. But that said, there are these scaling.