As we head into the long weekend, I’m excited to dive into a few pet data projects. What’s making this ridiculously easy is Sphinx AI + the new Google Colab extension. In <1 min, I can spin up free GPUs for our best-in-class copilot. See how below! Added bonus: it's all free 💸

No compute? No problem! You can now use Sphinx with Google Colab's VS Code extension to build complex models and analyses in your IDE, all totally free and with no need for local resources! See Sphinx build a CNN model on MNIST on a Colab GPU:

Sphinx is now even faster! And, we've improved the knobs for how autonomously Sphinx acts as it works alongside you. https://t.co/sjK6Ay6kX6

See the full deep dive and more details on our blog!

Even over a wide range of random scatterplots, frontier models only have a crude understanding of data! This is why Sphinx’s representation learning is critical to ensuring that data science gets done correctly.

Here’s why we think this happens – even if an agent naively looks at code and plots, AI is not good at building data intuition in the same way as a human. For example, we gave GPT-5 the following plot and asked it for the correlation, it estimated 0% (correct answer: -48%)

Unfortunately, house prices don’t start to fall after 3,200 sq ft! This model doesn't really make sense. When we toss this task to Sphinx, it excises the bad data and arrives at a much more reasonable solution

The bad values are not accounted for at all, so the regression line barely fits the data. When we ask Hex's AI to fix and improve this model, we get something even worse:

Consider asking Hex’s AI agent to model house prices using square footage. We allow around 30% of the square footages to be invalid numbers close to 0. Here’s what it comes up with

Not all data agents are created equal. LLMs are fundamentally bad at understanding data, and without a good representation of data they make mistakes that are obvious to a human. This is what separates Sphinx from solutions where AI is bolted on to a notebook (1/N)

Keep your notebooks clean🧹 Sphinx knows when cells are mostly just for exploration, and you can now ask it to keep these hidden automatically

Access Snowflake or Databricks data from anywhere with the new global connector configurations in the Sphinx Dashboard -- and enforce RBAC through familiar systems to ensure control

Sphinx CLI in action -- you can seamlessly jump back into interactively working with the same notebook with Sphinx in your IDE

We're back to shipping 🚢 Sphinx 0.6.3 makes data science more enjoyable than ever with major improvements that let Sphinx seamlessly connect to your data and analyses. We also are releasing Sphinx AI's core intelligence in a CLI, now in public beta! https://t.co/mPT4C1diRn

AI agents navigate by dead reckoning ... they have vague headings and take reckless steps. This results in logical errors and conclusions no better than hallucinations. Sphinx AI always finds its bearings. See how control and introspection lets us deliver accurate results:

MCP servers are a huge mess. Many servers return data, not context, and AI needs to account for that. When using MCP, Sphinx AI now disambiguates between data and context, and treats data as data. And yes ... we tried Linear for a week, and went back to our giant whiteboard.

Working with data superficially looks like software engineering , but the workflows are distinct in important ways. Watch Sphinx go head-to-head vs. Cursor on a simple data science task and see how these differences lead to significant divergences. https://t.co/TlGiOiA6H1

Playing around with @getsphinx's new batch embedding today to categorize some famous books ... not sure I'd agree that Harry Potter deserves its own cluster, but who am I to argue with the data 🧙‍♂️

Full Changelog: