This is a new tool which relies on existing introspection libraries like TransformerLens (which is similar in spirit to Garcon) to build an attribution graph. This graph displays intermediate computational steps the model took to sample a token.
Building Machine Learning Powered Applications walks you through building an ML application end-to-end, from product idea to a (simple) deployed version.
Sorry to hear that! Has tensorflow been your main issue? It is only used in one of the example notebooks, so you can skip that requirement without too much of an impact.
If I can help with troubleshooting, send me an email at mlpoweredapplications@gmail.com
In the meantime I'll bump up the version in requirements.txt to solve the conflict.
Yes, it has - I didn't have any other problems, but I did notice you recommending the usage of `pip install -r requirements.txt`. I am by no means an expert in Python, but it is my understanding that using pip this way isn't recommended and you should instead invoke `python -m install -r requirements.txt` where python could be python36, python38, or just python. The same with virtualenv, `python -m venv ml_editor` appears to be the new way of doing this.
Interesting. I wasn't aware of that issue. It seems like after running activate, `pip` does point to the right location (`which pip` points to the virtualenv pip), but I'll look into it.
Unfortunately the changes you did didn't improve my problem, pip insisted again that a correct tensorflow version wasn't available. I was, however, able to download tb-nightly so I do at least have a version. Quite shocked tensorflow thrashes so intensely on installation - is that normal?
I'm glad you have found a workaround. Again, this is for an example script outside of the main narrative of the book, so the impact should be minimal.
If pip struggles to find a specific version, I'll usually remove the specific version requirement, and give whichever version pip finds a try by running `pip install tensorflow`
That’s a great question, and I’m not sure. Googling around led me to an HN thread where a commenter claims that it is TheSans (lucasfonts.com/fonts/the-sans) but I do not have a good enough eye to confirm!
I answered the next comment down the chain, but I agree with you. As framed, there is no satisfactory solution to making a question "good".
Improving question quality could however be a product goal for StackOverflow, or for a company focusing on writing tools (Grammarly, Textio,...). The book describes a process for turning that vague product goal into a more tangible set of metrics, which lead to choosing an ML approach, and iterating on it.
Eventually there is a finished prototype, for which you can find a GitHub link in the PDF preview. It has definitely not solved how to evaluate if a question is "good", but aims to provide a narrower set of recommendations (the first chapter actually dives into an approach for this).
Yes, this is the approach in the book. The concept of question quality is nuanced, and does not have a clear definition. It can be easy to feel like you've solved the problem by just throwing in ML and calling it a day, but producing something useful is a real challenge.
The book covers multiple aspect of that process, from choosing an ML approach that isn't too simple or ambitious, to iterating on a model within the context of its final use case (i.e rather than only optimizing for a metric, testing how the model helps with its end goal).
In my experience, I've found that it is often those challenges that make or break the quality of an ML product, so the book focuses on tools to make complex problems more tractable, and less risky.
This is a new tool which relies on existing introspection libraries like TransformerLens (which is similar in spirit to Garcon) to build an attribution graph. This graph displays intermediate computational steps the model took to sample a token.
For more details on the method, see this paper: https://transformer-circuits.pub/2025/attribution-graphs/met....
For examples of using it to study Gemma 2, check out the linked notebooks: https://github.com/safety-research/circuit-tracer/blob/main/...)
We also document some findings on Claude 3.5 Haiku here: https://transformer-circuits.pub/2025/attribution-graphs/bio...)