- Kevin Feasel
- Mala Mahadevan
Notes: Questions and Topics
Azure Data Studio and SSMS 19.1
Our first topic was an update on SQL Server tools. Mala and I talked about our experiences using Azure Data Studio, and I mentioned a three-part series from Erin Stellato on the most recent release for SQL Server Management Studio (part 1, part 2, part 3).
Azure Data Studio has made some great strides, especially in how it displays execution plans. If T-SQL development is your primary use case, give Azure Data Studio a try. If you’re mostly a database administrator, you can also give Azure Data Studio a try, though SQL Server Management Studio is likely to have more of what you need to do your job.
AI Feedback Loops and Costs of Training
Anders pitched me a softball with this article about the AI feedback loop. The idea is that, with the proliferation of Large Language Models (LLMs), they’re generating a large amount of the data which will be used to train the next generation of LLMs, and we tend to see model quality drop with auto-generated gibberish, leading to even more auto-generated gibberish.
I also covered an article on diminishing returns to Deep Learning techniques, indicating that it takes about k^9 computational power to cut errors down to 1/k. For example, if we want to cut training errors in half, that’s k=2, and we need about 500x as much computational power to do that. For this reason, there are some slowly-changing boundaries around how much neural network model improvement we can see with additional hardware generations.
A Case Study in Insider Risk Management
The last topic was an interesting story I learned about because of a training I helped deliver. The case study involved a woman who stole a formula while working at Coca-Cola and tried to use it to bootstrap a new company in China. The whole story is quite long, but goes into a good amount of detail on the topic and was pretty interesting, especially as it details how she was able to circumvent Coca-Cola’s Data Loss Prevention software that they were using at the time.