One of my favorite scientific figures is this one of the entropy levels of 100 world cities by the orientation of streets. The cities with most ordered streets: Chicago, Miami, & Minneapolis. Most disordered: Charlotte, Sao Paulo, Rome & Singapore. Paper: appliednetsci.springeropen.com/articles/10.10…
To illustrate, here is how the maps look for Minneapolis (most ordered) & Charlotte (most disordered), rendered with this amazing tool for showing maps of roads by @anvaka: anvaka.github.io/city-roads/?q=…
Here's the cities organized by alphabetical order from the paper (by @gboeing )
On a related note, here is the alignment of runways around the world, along with a rather amazing global zoomable map 👇 of every runway & how it is oriented.
To lower your city’s entropy level, use this map extending the streets & avenues of New York’s grid to cover the world. 10 Downing Street is really 3,709th Street & East 10,894th Avenue. All roads meet in Ishkuduk, Uzbekistan, the antipodean anti-New York. extendny.com/#S22.St.4.Ave/9
Also on the topic of roads: 2,000 years later, the paths of Roman roads explain economic development patterns in Europe, but not so much in North Africa, where the general use of camels after 100BCE led to less need for wheeled vehicles. (Cool subway-style maps by @sasha_trub)
• • •
Missing some Tweet in this thread? You can try to
force a refresh
Microsoft keeps launching Copilot tools that seem interesting but which I can't ever seem to locate. Can't find them in my institution's enterprise account, nor my personal account, nor the many Copilot apps or copilots to apps or Agents for copilots
Each has their own UIs. 🤷♂️
For a while in 2023, Microsoft, with its GPT-4-powered Bing, was the absolute leader in making LLMs accessible and easy to use.
Even Amazon made Nova accessible through a simple URL.
Make your products easy to experiment with and people will discover use cases. Make them impossible without some sort of elaborate IT intervention and nobody will notice and they will just go back to ChatGPT or Gemini.
As someone who has spent a lot of time thinking and building in AI education, and sees huge potential, I have been shown this headline a lot
I am sure Alpha School is doing interesting things, but there is no deployed AI tutor yet that drives up test scores like this implies.
I am not doubting their test results, but I would want to learn more about the role AI is playing, and what they mean by AI tutor, before attributing their success to AI as opposed to the other dials they are turning.
Google has been doing a lot of work on fine-tuning Gemini for learning, and you can see a good overview of the issues and approaches in their paper (which also tests some of our work on tutor prompts). arxiv.org/abs/2412.16429
I suspect that a lot of "AI training" in companies and schools has become obsolete in the last few months
As models get larger, the prompting tricks that used to be useful are no longer good; reasoners don't play well with Chain-of-Thought; hallucination rates have dropped, etc.
I think caution is warranted when teaching prompting approaches for individual use or if training is trying to define clear lines about tasks where AI is bad/good. Those areas are changing very rapidly.
None of this is the fault of trainers - I have taught my students how to do Chain-of-thought, etc. But we need to start to think about how to teach people to use AI in a world that is changing quite rapidly. Focusing on exploration and use, rather than a set of defined rules.
“GPT-4.5, Give me a secret history ala Borges. Tie together the steel at Scapa Flow, the return of Napoleon from exile, betamax versus VHS, and the fact that Kafka wanted his manuscripts burned. There should be deep meanings and connections”
“Make it better” a few times…
It should have integrated the scuttling of the High Seas Fleet better but it knocked the Betamax thing out of the park
🚨Our Generative AI Lab at Wharton is releasing its first Prompt Engineering Report, empirically testing prompting approaches. This time we find: 1) Prompting “tricks” like saying “please” do not help consistently or predictably 2) How you measure against benchmarks matters a lot
Using social science methodologies for measuring prompting results helped give us some useful insights, I think. Here’s the report, the first of hopefully many to come. papers.ssrn.com/sol3/papers.cf…
This is what complicates things. Making a polite request ("please") had huge positive effects in some cases and negative ones in others. Similarly being rude ("I order you") helped in some cases and not others.
There was no clear way to predict in advance which would work when.