Couldn't agree more. It's wild how much is realy happening in AI right now, even after the holidays. Seriously impressive you're keeping up and breaking it all down for us after those epic Thanksgiving meals! Makes my weekend reading list look chill.
Thanks for this great read, your insights into the current AI landscape are super valuable. That "Santa came early this year" metaphor for all the AI labs dumping updates is just spot on, it makes you wonder what kind of incredible sprints and breakthoughs are happening behind closed doors to lead to such a flury of relases.
Anecdotally, the labs aren't hiding anything. My friends at these companies are releasing things within a few months of having them in house because it's such a race to be the best
The TPU to GCP migration angle is pretty wild. You're right that once you commit to TPUs, you're basically locked into GCP for everything else because managing multi-cloud is a nightmare. Google's basically using chip superiority to force people into their whole ecosystem. It's the same playbook Nvidia ran with CUDA, just in reverse.
Thanks, this was a good read, and sparked a couple of thoughts:
* Claude vs Cursor vs Gemini, at least for programmers, I'm guessing is platform lock-in wrt MCP specifically, which non-programmers care less about. Between that and GCU vs TCU it's lovely to have the platform wars back. So exciting for the news.
* AI should be the perfect product for Google since it's the no-UX UX, it's all tech. I wish I were still at Google to see the interplay between Jeff Dean and Demis Hassabis. I suspect part of Gemini's lead is because they're taking reasoning pretty seriously.
* Ads is not the only monetization strategy that works, there's also pure momentum / market share. This is a la mode in politics, where more voters means winner takes all. Unfortunately this operates via back room deals not competition. Even more unfortunately and ominously, its largely the same set of people now pursuing this horrible mode in gov and biz, which I'm deeply worried is a long term realignment.
AI definitely feels like it lands in Google's sweet spot in terms of core competency. The same is true of self driving cars -- the business execution doesn't really matter, the demand is there. All of the difficulty is in building the tech.
RE the agents, I don't think mcp creates lockin. It's the opposite in fact, mcp is at this point pretty cross agent. But the big source of lock in is the various configuration options. I mostly use Claude because I have it really dialed in for my use cases (if you're interested, you can check it out here! https://github.com/tilework-tech/nori-profiles)
Interesting, I'd tried MCP on some other models (qwen IIRC) and its type inference on arguments was much less capable than promised. Should have tried the famous ones.
Really glad I stumbled upon your substack a few weeks ago. Enjoying the grounded writing, and the insider takes. So much noise out there at the moment so am glad I found some signal!
In the beginning on the Gemini performance, this was "yeah, yeah - we all read Zvi," but it really got interesting when you dove into the specific advantages TPU's have, and how this is essentially a structural advantage on Google's part.
One thing I've wondered, doesn't TSMC make all the TPU's too? So all USA AI is still bottlenecked if a political situation shut them down? I had a conversation (with Gemini) about that, and it thought they're still made there because of packaging and HBM requirements.
Re tsmc, yes I believe so. I think the US government has recognized this is a pretty big strategic risk and has started trying to get fabs online in the US. Of course, this admin is stupid so it is undoing much of the positive work of the previous admin just to redo it again but worse
If you care about the general AI landscape across all the companies, reading him is nearly mandatory - nobody else goes into such detail and catches everything important going on.
> has started trying to get fabs online in the US
I thought the fabs in the US could only do "two generations back" chip fabbing? Even the ones literally built / run by TSMC, for strategic reasons on their part? It's not going to alleviate the frontier chip bottleneck, in other words - but maybe that changed.
Couldn't agree more. It's wild how much is realy happening in AI right now, even after the holidays. Seriously impressive you're keeping up and breaking it all down for us after those epic Thanksgiving meals! Makes my weekend reading list look chill.
Thanks for this great read, your insights into the current AI landscape are super valuable. That "Santa came early this year" metaphor for all the AI labs dumping updates is just spot on, it makes you wonder what kind of incredible sprints and breakthoughs are happening behind closed doors to lead to such a flury of relases.
Anecdotally, the labs aren't hiding anything. My friends at these companies are releasing things within a few months of having them in house because it's such a race to be the best
The TPU to GCP migration angle is pretty wild. You're right that once you commit to TPUs, you're basically locked into GCP for everything else because managing multi-cloud is a nightmare. Google's basically using chip superiority to force people into their whole ecosystem. It's the same playbook Nvidia ran with CUDA, just in reverse.
Thanks, this was a good read, and sparked a couple of thoughts:
* Claude vs Cursor vs Gemini, at least for programmers, I'm guessing is platform lock-in wrt MCP specifically, which non-programmers care less about. Between that and GCU vs TCU it's lovely to have the platform wars back. So exciting for the news.
* AI should be the perfect product for Google since it's the no-UX UX, it's all tech. I wish I were still at Google to see the interplay between Jeff Dean and Demis Hassabis. I suspect part of Gemini's lead is because they're taking reasoning pretty seriously.
* Ads is not the only monetization strategy that works, there's also pure momentum / market share. This is a la mode in politics, where more voters means winner takes all. Unfortunately this operates via back room deals not competition. Even more unfortunately and ominously, its largely the same set of people now pursuing this horrible mode in gov and biz, which I'm deeply worried is a long term realignment.
AI definitely feels like it lands in Google's sweet spot in terms of core competency. The same is true of self driving cars -- the business execution doesn't really matter, the demand is there. All of the difficulty is in building the tech.
RE the agents, I don't think mcp creates lockin. It's the opposite in fact, mcp is at this point pretty cross agent. But the big source of lock in is the various configuration options. I mostly use Claude because I have it really dialed in for my use cases (if you're interested, you can check it out here! https://github.com/tilework-tech/nori-profiles)
Interesting, I'd tried MCP on some other models (qwen IIRC) and its type inference on arguments was much less capable than promised. Should have tried the famous ones.
Really glad I stumbled upon your substack a few weeks ago. Enjoying the grounded writing, and the insider takes. So much noise out there at the moment so am glad I found some signal!
thanks! tell yo friends!
In the beginning on the Gemini performance, this was "yeah, yeah - we all read Zvi," but it really got interesting when you dove into the specific advantages TPU's have, and how this is essentially a structural advantage on Google's part.
One thing I've wondered, doesn't TSMC make all the TPU's too? So all USA AI is still bottlenecked if a political situation shut them down? I had a conversation (with Gemini) about that, and it thought they're still made there because of packaging and HBM requirements.
I should probably read more zvi.
Re tsmc, yes I believe so. I think the US government has recognized this is a pretty big strategic risk and has started trying to get fabs online in the US. Of course, this admin is stupid so it is undoing much of the positive work of the previous admin just to redo it again but worse
> I should probably read more zvi.
If you care about the general AI landscape across all the companies, reading him is nearly mandatory - nobody else goes into such detail and catches everything important going on.
> has started trying to get fabs online in the US
I thought the fabs in the US could only do "two generations back" chip fabbing? Even the ones literally built / run by TSMC, for strategic reasons on their part? It's not going to alleviate the frontier chip bottleneck, in other words - but maybe that changed.
I believe the Arizona plant (https://en.wikipedia.org/wiki/TSMC_Arizona?wprov=sfla1) will eventually be able to do the more advanced stuff