I wonder how will this behaviour evolve over time? Right now waymo is definitely prioritising safety, but as the tech matures (and competition grows) will the systems start to prioritise speed and so little-by-little start cutting the margins they give to pedestrians? As with any digital platform this degradation wouldn't be explicitly chosen, but just the consequence of many little A/B tests designed to optimise some other metric
I wrote this post setting out how to work with cloud data in Polars. Although I've been using Polars for almost 3 years I still learned from writing this. I cover topics like how a CSV scan is different from a Parquet scan, authentication and query optimisations. Interested to hear your feedback!
There are some big questions around OpenAI's Advanced Data Analysis. Is it cutting edge data science or just bad Medium articles all the way down? Can you trust it? And will it let Larry from Sales take your job?
tldr: the exploratory analysis was nice, the text summaries are oft wrong, the choice of ML models was poor IMO and it gains a lot from expert guidance
Many of the physical models are of similar complexity to the model in this paper - the sort of analytical model that you run in a few minutes on your laptop, not on a super-computer. It's not a question of complexity it's about whether a statistical model that is not constrained by well-understood constraints is a high-value model
I've got a background in this field and I am very surprised to see this published in Nature. The model presented is purely statistical with no representation of the underlying physics. When we are dealing with a phenomenon that is driven by well-understood physical laws (e.g. geophysical fluid dynamics, radiation physics etc) then these physical models are the most reliable basis for prediction.
When I say physical models here by the way I'm referring to physically-based mathematical models as well as numerical models.
It seems that the authors have done a good job in developing their model. My issue is with Nature deciding to publish it. If this paper was not published in Nature it would receive little attention within or without climate science - in fact many such statistical models are published each year without much comment. However, Nature have published a paper that I think many ocean scientists would feel draws dramatic conclusions from a weak basis but will now inevitably draw much more attention than more insightful papers.
MIT professor Carl Wunsch accused Nature in 2010 of near-tabloid science with a tendency towards sensational papers built on weak foundations. However, I've felt that Nature's choice of publications on climate in recent years has been high quality. This paper feels like a big step-down from that standard.
Some thoughts about gathering trends in data science - in particular the area of data processing that's been dominated by pandas for the last 10 years. I'd love to hear what you think!
I've been making these comparisons between the new dataframe libraries in python so a couple of comments:
"Reading the full CSV without datetime parsing is in line in terms of speed though."
This sentence was a bit ambiguous, but is important: if you read this file in pandas with engine='pyarrow' but don't convert the date/time column to a pandas datetime dtype you get the same ~100 ms read time as calling PyArrow directly. So basically the entire time is spent converting the strings to dates. If this datetime thing isn't an issue for you then you can just use the engine='pyarrow' argument to read CSVs with Pandas.
In my own tests with various datesets Polars has always been much faster than duckdb/pyarrow. For this relatively small dataset it's about 2x faster, which is about the smallest margin I've found. Polars is also much easier to write, as its query optimization is so effective - you don't need to know all the tricks that Pandas requires (and is still 3x-10x faster than Pandas even when you apply all the tricks).