I first heard of Windsurf about 6 months ago, from
, on a Bluesky thread where he said it was better than Cursor for “workflowy” things.I was using Cursor to build my dlt and sqlmesh project that I used for my Advent series in December, and it just couldn’t deal with a problem I had. Reading what Nico said, I thought why not give it a try?
I had just paid for Cursor in order to get more credits and access to a more powerful model, but it still hadn’t helped. When I tried Windsurf, it solved the problem relatively quickly. I asked for a refund from Cursor and provided them with some feedback and where the puck I was going.
dlt windsurfing
Yesterday, I mentioned that I wanted to focus on getting some data from the Bluesky API endpoints using dlt - which involves learning how dlt works. You can follow along with my public repo here.
I thought Windsurf was great, but six months ago, I never would have guessed that they would have a $3 billion exit, acquired by OpenAI, and especially so soon. Could a fork of VSCode be worth so much1? In the meantime, Cursor became the fastest-growing company of all time.
Rumours circulated that OpenAI would have preferred to acquire Cursor, but the price was too high, especially with a comparable, and in my opinion, superior product available. Of the big LLM providers, only GCP could realistically afford to acquire Cursor now.
There are tremendous synergies available from this acquisition. Not only do OpenAI already offer some great models for coding, like o3 and o4 mini, but imagine what they could offer if they built a specialised LLM/s2 for use in Windsurf? To date, the AI IDE companies and providers have had to anneal their tools to state-of-the-art LLMs. Some had the right context window, some had the right level of ability, some were better at coding…. they had to adapt their software to what was already state of the art because it’s not reasonable for them to pre-train their own LLMs - not any more.
I believe Windsurf is already better than Cursor, but the gap at the time I assessed it was not insurmountable. If Windsurf is now able to use LLMs, which are specialised to be used with it, and Cursor does not, then this becomes a killer advantage. I can’t see how Cursor can catch up in this scenario. However, there is no reason that Cursor couldn’t build some kind of strategic partnership with Anthropic, given that they are too large to be acquired by Anthropic (although they could merge - unlikely though). Cursor could also be acquired by GCP and have specialised Gemini Pro models built for it, which would level the playing field. This does seem to be the year of acquisitions.
GCP just released Firebase Studio, which I have yet to try. So it may be unlikely3 for GCP to have any interest in Cursor for some time, at least until they have validated whether Firebase Studio is a success.
How is all of this relevant to data folks? Well, if you use an IDE, you should probably think about trying something like Windsurf or Cursor, if you haven’t already. I know you might say, “I don’t need it, I already know how to code.” What I would say in response, from my experience building whole projects4 in SQLMesh with Windsurf, is that it’s great that you know how to code as you can see where these IDEs make mistakes, but they are already much faster than us puny humans and will only get faster and better from here. We are not going to become an order of magnitude better at coding as a species on our own - we’re about as good as we’re ever going to get. We’ve made language servers, great IDEs (not considering AI ones), coding bootcamps and tutorials galore, more accessible languages, languages better for memory safety… none of these things have given us anywhere near the boost we’ve had from AI in the last two years.
Mon ami Christophe Blefari, alias
, l'ingénieur data le plus célèbre de France, has recently co-founded a startup, with Claire Gouze, providing a new AI IDE for data folks, called ‘nao’, also a VSCode fork. I will try it soon and share what I think. It will be a tall order for me to say I prefer it to Windsurf. I haven’t really looked at it in any depth yet, but I expect things like native SQL explorer support, where I can explore a database schema and run SQL queries against it without leaving my IDE, in order for there to be data specific advantages over Windsurf and Cursor.May the forks be with you.
Codeium actually produces other plugins that allow you to have the Windsurf experience in other IDES as well. Of course, the Windsurf IDE is much more than just a fork of VSCode. Starting from VSCode offers several advantages compared to building from scratch. Many engineers are already familiar with VSCode, making it easier for them to migrate. Additionally, you benefit from the extension ecosystem, and it’s already a very mature IDE.
Ancillary features and integrations custom-made for Windsurf may be as crucial as specially pre-trained models.
I built some example projects in their entirety in one shot. No matter how great you are at coding, you can’t make 20 models in 2 minutes.
Also to mention that Cursor might be fundraising $900m at $9b valuation. The funny thing is, that Cursor and Windsurf are not "forks" companies, the fork is just a vehicle of an experience—they’re actually model companies. What makes them better than Copilot is the UX, but more than anything else the way their own models provide fast and "magical" answer. From the ground up they bet on training their own models.
Thanks for the mention! Regarding nao, would love to get your feedback on the native SQL explorer experience! It's working great atm and we have so many ideas to improve it and I think you have too.
Already, Cursor IDE is utilizing some of their homemade LLMs. Their CEO stated that when they shifted from generic LLMs to more tailored LLMs, such as those used for editing code, the performance increased dramatically.