6.5 KiB
| date | post_meta | tags | title | type | url | thumbnail | images | |||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 2023-10-08T13:41:04.117044 |
|
|
Weeknote Week 40 2023 | posts | /posts/2023/10/08/weeknote-week-40-20231696772464 | http://brainsteam.co.uk/media/2023/10/08/1696769396_ |
|
This week work has been incredibly busy again - a running theme this time of year.
On Monday we went to CostCo and spent lots of money bulk-buying things. Although I'm always horrified by the bill at the checkout, I love to come home and calculate how much I've saved versus buying the same product from my local supermarket in small quantities. This time it was about £60 which is pretty good!
On Tuesday I went up to London for an event that our firm put on for our clients and potential clients. I had planned to stay over but found out on the day about a train strike that would have left me stuck in London on Wednesday so I had to change my plans last-minute and come home.
It's been pretty darn chaotic all week (as it usually is at this time of year) and I've got quite a few plates and projects to juggle.
I've been spending more time working with Airbyte, trying to build ways to bulk export data from our client environments in a safe but efficient way. I've ended up building a little lightweight config format that sits on top of Airbyte's very verbose, yaml format and applies sensible defaults and a little diff tool for effectively helping a user to "transpile" my config format into the proper one and see what differences will be applied to the sync process. I was happy to find oyaml which allows you to serialize yaml files in the same order that you read them in python - this makes diffing much, much easier!
I've been helping some of my colleagues evaluate Microsoft's Retrieval-Augmented-Generation solution for searching data in our product. Overall we've found that it gives pretty mixed results. The problem is that it gets things right consistently until it suddenly doesn't and just makes stuff up. I can't say I'd recommend that people use these systems to do factual question answering unless the model is fully extractive (i.e. rather than generating the output, it simply points to a snippet of text in a document that answers the question). I've also done a lightweight due diligence exercise on another tech firm's tech + processes which I need to write up properly this coming week.
On the home front, when I've had time to stop and relax, Mrs R and I have been watching the last ever series of Sex Education and on Thursday we also went to the cinema to see The Creator which I enjoyed a lot. Over the weekend I got my flu shot (the nurse said "you're clearly under 65, do you have an underlying health condition?" - "yes, Asthma!"). I also managed to get a few chores done around the house and donated some board games that we haven't played for a while to a charity shop.
I've got a Delonghi coffee machine that I've had for about 10 years which started really struggling to pump water through the espresso basket. I read online that this is a common fault, found some videos of others dismantling the machine to rectify the fault and purchased a replacement for component that commonly breaks in this machine. I was able to dismantle the machine and replace the part but when I rebuilt it, I found that the machine only produces cold water (although the steam wand still makes steam so the heater must still be working). I'm at the edge of my knowledge now so I might take the machine to the local repair cafe to see if they can help me figure out what to do next.
I've also been rushing around trying to make sure that my tuxedo still fits and find a suitable shirt and bow tie for an event next week.
I read some interesting articles:
- This paper which I've only skim-read suggests that allowing language models to "pause" by emitting characters that don't matter (I guess like ellipses "...") can improve their performance at some tasks -
- I enjoyed this article about how our perception of what we read can be altered by the medium upon which we read. The author speculates that e-readers feel less special than paper because they don't weather or acquire scars (...urm mate, you should see my kindle...) and you can't make notes in the back pages (writing in books is a pet hate of mine).
- This article talks about a productivity tool called resistance lists where you create a to-do list just for things you really don't want to do and force yourself to do one thing from it every day. A nice idea but I can already imagine the ways that I'd cheat this system so I probably won't try it.
Next week is already looking to be a busy one. I'm in London for 3 days, attending a Snowflake conference on Tuesday and giving a guest talk at a University of Essex event on Wednesday. On Thursday I'll be polishing off that tux for the South Coast Technology awards where my company is shortlisted.