AI link dump: AI & spreadsheets; AI & software development; AI & mining

A collection of interesting AI-related links

Large language models & spreadsheets

Back in May I went viral on Twiter for posting my astonishment at getting ChatGPT to build a discounted cash flow analysis for Apple. I wrote about the experience here, and at some point in the near future I plan to write a more detailed writeup of how I got ChatGPT to do some math. At a very high level, the way that I did this was to provide ChatGPT with a spreadsheet of Apple’s financial statements, as sourced from the SEC’s web site, and tell it to use Python to parse the data. Python, in other words, served as a wrapper around the spreadsheet data.

It never occurred to me to ask ChatGPT whether it could natively access, and make sense of, the data in a spreadsheet. But, according to a group of Microsoft researchers, today’s large language models basically can’t parse spreadsheets. Though as I indicate above, today’s LLMs can call Python to parse the data on their, and their user’s, behalf. The researchers’ paper is here, and they explain their research:

Spreadsheets are characterized by their extensive two-dimensional grids, flexible layouts, and varied formatting options, which pose significant challenges for large language models (LLMs). In response, we introduce SPREADSHEETLLM, pioneering an efficient coding method designed to unleash and optimize LLMs’ poewrful understanding and reasoning capability on spreadsheets.

You can immediately see the appeal of such technology. Instead of having to rely on Python to parse a spreadsheet, which relatively few people know, as compared to the number of people who use spreadsheets, you could just use the spreadsheet itself, and query it using natural language. If their research turns into a usable product, the implications would be quite profound: suddenly, the cognitive powers of large language models would be available to vast swathes of knowledge workers who rely on spreadsheets to perform their various work tasks.

---

Building Devin

Here’s an interesting video of Scott Wu talking about his company’s AI software building product, Devin. Wu is the CEO of Cognition AI, the company behind Devin. The basic idea is: you should be able to provide an AI a natural language explanation of some software that you want to have built, and the AI will go ahead and build it. It’s a great concept, and if Cognition AI is able to pull it off, it will go a long way to making AI more generally useful. At present the product still appears to be invite-only, so I haven’t had a chance to try it out.

---

Mining minerals with AI

KoBold Metals is a geoscience company, which makes extensive use of AI to find good mining prospects around the world. The New York Times recently reported on its efforts in Zambia. KoBold successfully found a vast new source for copper, estimated to provide 300,000 tons of copper per year once operational.

A quick Google search suggests that Chile is the world’s largest producer of copper, at around 5 million tons per year. 300,000 additional tons of copper per year sounds like a lot, but relative to world supply it does not seem like it really moves the needle that much. Nonetheless, to the extent that this proves KoBold’s method for using AI to find rich sources of metals, it augurs good things for our future.