How many times, in how many contexts, from how many internal and external researchers, or from how many CEO's are people going to have to receive this message before they believe it: "Hallucinations" are an inherent part of the large language model architecture. www.nature.com/articles/d41...

I carried out the same exercise in metric, too: the number of hours during which the temperature (in C) is the latitude. Cooler equatorial oceans -- and a long span of northern Africa and south America. danyelf.github.io/degree-degree/

A friend posted a photo, sunbathing from Scandanavia a little bit ago, captioned "60 degrees at 60 degrees!" That made me wonder how often that really happens -- where on earth is the temperature the same as the latitude, and how often? I answered that at danyel.github.io/degree-degree

The other day, I watched a (small local) LLM melt itself down. I was testing out Ollama with the 8b Deepseek model. I gave it a moderately-challenging technical question -- how to make a Javascript and Python package talk to each other -- and hit "go". Three hours later...

Strunk & White stole all their ideas from that spider

This is hilarious - ChatGPT hallucinates repeatedly about a nonexistent feature on another website (confusing people), eventually leading the site's owners to implement the feature. www.holovaty.com/writing/chat...

Given that these grotesque abuses are being conducted for our supposed benefit, what do you imagine they'll do when Jews are not sufficiently grateful? What will be the punishment? And which Jews and Jewish orgs will cooperate? May God give us soft hearts and strong spines.

Oh, my God, it's the HUD sucker proxy.

The White House is pretty much issuing revised official photographs that no longer show out of favor politburo members.

If something reduces death by 30%, then removing that thing increases death by 50%. There are about 40,000 deaths by flu annually, placing it just behind liver disease and self-harm as a leading cause of death. With no vaccine, I bet we could make it up to 60k and hop 3 steps in the rankings!

Are you a data person and wondering how you can combat these dark times? (H/t @micheletobias.bsky.social) www.muckrock.com/news/archive...

I suspect that the person at OpenAI who saves 5% on a power cost gets a much smaller promotion than the person who shows how to get a 5% increase on a loss function. Cost isn’t the issue, so cost isn’t an issue.

I seem to have two different blog entries -- or, perhaps, rants -- half-written. Help pull me off the fence before I go away for vacation! Which would you rather see? Option 1: Visualization Is About Comparison (and anyone who disagrees is trying to sell you a gauge)

Finally got a chance to read these comments on box plots. I think it’s really important to realize that the visualization forms of the mid 20th century were doing really different work than 2”we do today.

For #30DayChartChallenge - day 26 - ai I wanted to make a chart that showed generative AI generates expected seeming results consistent with the kind of corpus associated with the prompt, with no understanding of truth, facts, logic, or causality. So I had chatGPT supply the data for the graph.

It’s surreal. Vaccination is one of the easiest-to-quantify, mind-boggling successful health discoveries ever. A solid “w” ticked on the human column.

We tried. It’s weirdly unavailable to stream right now.

Hi new website! I'm Alper and I work on the Power BI Core Visuals team as an software engineer at Microsoft. I try to bridge the gap between research and practice in driving effective visualization authoring and consumption experiences for people of varying data and visual fluency.

I've been blogging for a big about the concept of "genre" in data visualization -- and this final entry helped me clarify why I think it's a useful tool. Does this breakdown make sense to you? danyelfisher.info/blog/2024/10...

I've been thinking a lot about how to interview for UX well recently. Smarter people than me have written books on this, but I've run into a few recurrent issues. Here's a few big ones: danyelfisher.info/blog/2024/2/...

I never understood survivorship bias and I turned out okay.

I feel like new frontiers in data visualization can be explored with this approach. (Or did Tamara Munzner already cover this? graphics.stanford.edu/papers/webvi... )

It's time for the IEEE Vis conference. I'm not in Oz this year, but the paper listing is online ieeevis.org/year/2023/in.... Authors are pretty good about putting their papers on arxiv ; I've been downloading and following along.

I got to use this data-first philosophy was my work at Honeycomb on BubbleUp. We figured out the visualization first, and the way it would fit into the interface much later. danyelfisher.info/blog/2023/7/...

When I first learned about interface prototyping, we distinguished between high- and low-fidelity prototypes: cardboard and wireframes vs elegant, near-functional images. I've been chewing on how (I think) data visualization breaks that model. danyelfisher.info/blog/2023/7/...