Eric Colson

An animated tour of Stitch Fix Algorithms. algorithms-tour.stitchfix.com via @stitchfix_algo #datascience #algotour

Find me on Bluesky: @ecolson.bsky.social. Love all things data science, machine learning, AI, and management there of.

@emollick This is especially true in ML/Data projects which typically have unclear requirements that need to be revealed through trial & error. In theses cases adding an additional person can be a negative result.

The eternal trap of managers: adding more people doesn’t do what you think. We are too optimistic about project costs & schedules because we overestimate the value of adding team members while underestimating coordination costs as team size gets larger. static1.squarespace.com/static/5a9eff3…

One of my heroes. Thank you for all the contributions to knowledge

Working in companies shouldn't be so hard. Even when we all want the same things, we trip over bad process and unnecessary coordination. At the same time, other things can be too easy and we do things we later regret. Learn how to smartly work with friction. The Friction Project: How Smart Leaders Make the Right Things Easier and the Wrong Things Harder. By @work_matters and Huggy Rao amazon.com/Friction-Proje…

Must be fun to be a PhD student right now. So many adjacent possibilities surface now that foundational models are available. Cool paper on the relationships between objects in an image. arxiv.org/pdf/2402.19474… #AI #computervision

@work_matters To the conscientious host, the cost of moving a meeting increases exponentially with the number of times its already been moved. Therefore a meeting that gets rescheduled more than 3 times provides evidence that the host is not very conscientious.

One of my decision rules is, if a colleague or client moves a meeting, interview, or presentation more than three times, I usually "declare defeat" and decline further scheduling games. I conclude they are too busy, too flaky, or I am too low priority to them. That reasonable?

I appreciate chatGPT's humility in admitting mistakes.

Good framing on the tradeoff between moving fast and shipping high quality code. In a data science context, I believe one needs to lean towards 'fast' in the short-term in order to learn more quickly and allay uncertainty. But, once value is found, we need refactor to achieve a robust system in the long-term to enable future learnings. Good infrastructure can shift the efficient frontier to outward. Nice job @elijahbenizzy blog.dagworks.io/p/how-well-str… #datascience

Gorgeous visualization. Effective too. Coneys so much more than text alone. NYTimes always does a nice job. nytimes.com/interactive/20…

Great article featuring @work_matters and @huggyrao on "addition sickness": the unnecessary rules, procedures, communications, tools, and roles that seem to inexorably grow, stifling productivity and creativity.

@QRich @ShaneBattier @PlayersTribune Well played. Great outcome

“Did winning the championship that year make it worth it?….Hell yea!!” @ShaneBattier talks thru staying at Duke for his Senior year when he could have went to the NBA after his junior season. Ep is live! Tap in! #knuckleheadspodcast @playerstribune ✊🏾😎✊🏾

So many companies say they are okay with failure but they still behave in ways which discourages risk-taking and innovation. In @AmyCEdmondson new book she articulates the different types of failures which gives us with the vocabulary to adjust and embrace the right kind of wrong.

Very general statement: I am not a fan of using aggregated data for decision making. Most business decisions require good statistics to control for confounders and slow our ability to confirm our biases (and an RCT is best when possible). Self service analytics is best for monitoring.

What's the strongest case you can make for having stakeholders consume data products without an analyst or scientist to help with formulating questions or interpreting results?

RIP Netflix DVD. A long run.

@flowingdata Would love it if you also posted to threads. Would love to follow you there instead

Twitter throttling competitor links is just so... can this just not be a terrible place? It is/was an actual useful thing to have themarkup.org/investigations…

Very appreciative to @wesmckinn and so many others that help evolve this space. So many of us stand on your shoulders while focusing on the mere application of these technologies and receiving the kudos for the value they generate. Thank you! vldb.org/pvldb/vol16/p2… #datascience #machinelearning #Ai

@erikbryn less carful in rural areas?

Covid is a contagious disease spread when people interact with other people. So why are Covid death rates higher in rural areas than urban areas?

Expressing PySpark Transformations Declaratively with Hamilton blog.dagworks.io/p/expressing-p…

Mind blown. I just realized that iPhone and MacBooks can share a copy buffer. Copy on one, paste to the other. How long has that been there?

@work_matters Ugh, I bet they call themselves “data driven”. A shame. Those surveys are notoriously fickle, capturing only a narrow point in time. They also lack robustness; ask the same questions 1 week later and you will get different answers!

Talked to an EVP who says his company does an employee survey twice a year that has 80 yes 80 questions He and other execs spend big hunks of time dealing with results and it affects their pay—the survey has no established reliability or validity Corporate addition sickness!