| Building a Reporting Disclosure. I have a few thoughts. One, I can create a model. A model is actually quite feasible. It would be, but it's a data annotation project. It's saying, here's a PDF, here are the actual variables. In other words, here's the scope 3, scope 2, scope 1, here are the units, train it like that.<br><br>Second thought is if I did want to put together a dataset of sustainability disclosure reports, I think you could argue a public fair use clause for the PDFs being there.<br><br>And then the one I did with Gemini the other day which was basically a parsing AI tool seemed to work and could probably be used in production and which works even maybe as a way of trying to get in touch with Google is they have They have definitely an AI for good division who may let's say provide Gemini credits for the actual deployment of it on Cloud Run. Because from my first run of it, it was very, very promising for the task of parsing the reports.<br><br>And that would greatly the feature would be when it extracts the data human human in the loop is done by seeing what it is matching it to a company in the database or to a known company Let's take Google itself as an example. Detects its stock ticker, detects its stock exchange. And then you click like add to database meaning that you're adding the validated data and it could even pull out the metadata from the document pull out the source and that would be a great way of building up a human validated database in other words you take the reports you say either everything everything looks good to me or this is wrong either way you add it then of course you've got the missing financials and the rest of the world.<br><br>But that would probably be because there is thousands of sustainability disclosures, especially when you consider I think beyond the US globally, and it's beyond. So certainly it's a task for a model, but it's also human in the loop. The ultimate question is if Gemini stock performs 99% sufficiently well in the task of extracting this data from the sustainability reports. A model might actually not even be necessary because out of the box it's almost perfect. That is, I suspect, what the case would be. |