Andy Wheeler

Anthropic has AI architect roles that are focused on Federal, State, and Local government. I think these are good positions -- I would like gov to invest in more in house capabilities to do these things as well. For current students and professors, the skills to get this job (listed at $240 - $345k in NYC for this particular advert), is fundamentally what my book, Large Language Models for Mortals: A Practical Guide for Analysts with Python, is about. This thread is a tech skills needed for a modern LLM stack focused AI engineer.

So I wrote this before I knew that it was reps at AWS that asked for the models to be turned off. Which just reinforces that this will ultimately be a potential solution.

Just as an FYI to anyone sad about Anthropic, Antigravity IDE has plenty of free credits (I toggled this, did a ton of work, and did not even tick off a single credit, honestly it may be a glitch at the moment).

Weekend project alpha release -- this is a dashboard to look at trends for reporting crime in the US (via the National Crime Victimization Survey) crimede-coder.com/graphs/ncvs Let me know what you think!

Anthropic should turn on Fable for AWS govcloud accounts.

For a counter-point, andrewpwheeler.com/2025/02/20/lin…

Links, Image credit from the algae bloom prediction competition I won (via Data Driven), drivendata.org/competitions/1… Daniel Neetzel was the one working on good algae, liquidtrees.org/about One commercial product to kill the algae, lgsonic.com/ultrasonic-alg… (not affiliated, just one example)

Some nerd notes on algae (inspired by the recent algae in the reflecting pool notes I see in the timeline). The algae that sit at the top of the water (cyanobacteria) are bad for multiple reasons. They tend to kill other flora and fauna in the water sources, and emit greenhouse gases. It is a problem in natural water sources as well, in particular ones around farming areas that have sources of nitrogen. So I do not know the filtration with the reflecting pool, but there are non-chemical solutions. One solution is to use ultrasonic frequency to kill the algae. There are commercial products available now for this. The other I am familiar with (which makes more sense for natural areas) are to basically seed the pool with good bacteria (diatoms) that sink and do not cause the toxic blooms. They compete and eat up the food, so the bad bacteria cannot take root.

Paper published in Cartography and GIS, tandfonline.com/doi/abs/10.108…

Example of clustering flows.

Journey to crime flow data is hard to visualize.

We know that not all crimes are reported, so I presented one way to adjust reported crime estimates given this undercounting. If you have a crime reporting rate of 50%, and you observe 100 crimes, the up-adjusted number of crimes is 200. In a paper with Alex Piquero, I extend this to show up-adjusted estimates of serious domestic violence combining NIBRS and NCVS.

@DavidBuil I don't handle D in my framework, but it is just R can vary based on individual level crimes. (I am just interested in getting an estimate of the total number of crimes, not in estimating right hand side of regression equations.)

@CrimeDecoder I’ll read the paper, thanks Andy! The ‘rcme’ R package led by Pina-Sánchez deals with a similar issue: link.springer.com/article/10.118…

This was published in Crime & Delinquency, journals.sagepub.com/doi/abs/10.117… Curious to see it applied for more crime types (pooling reporting rates in NCVS I think is a good idea given the small N), as well as sub-city (nothing stopping you from doing this at the neighborhood level).

The biggest critique of this work is it begs the question of whether the NCVS estimate can be applied to micro places based on demographics. E.g. can the nationwide sample approximate the reporting rate for any city (or even smaller neighborhood). I think that is likely yes, but it is an empirical question.