Reading AI research papers, etc

I found this useful, particularly the categorization of papers into surveys, benchmarks, and breakthroughs. The presentation is only a few days old and so the specific papers listed are representative of the current state of art. As a CS Phd friend of mine once told me when I was struggling to understand an algorithm proof, as a practitioner you really don't need to understand the proof, just the results. I suspect that is true for these papers too.

How To Read AI Research Papers Effectively
https://www.youtube.com/watch?v=K6Wui3mn-uI

At local meetup I mentioned that several years ago Google had released hardware that targeted machine learning. I could not remember the details. As luck would have it, I listened to Jeff Dean's presentation this weekend and he mentioned the Tensor Processing Unit (TPU) for low precision linear algebra. Overall, the presentation is interesting and useful. The majority of it is focused Google's Gemini/Bard, but given who the speaker is this is understandable.

Jeff Dean (Google): Exciting Trends in Machine Learning
https://www.youtube.com/watch?v=oSCRZkSQ1CE

Timothy Snyder lectures and interview

Timothy Snyder is an historian and an exceptionally lucid lecturer on the history of Central and Eastern Europe, the Soviet Union, and the Holocaust. Given what is promised by the Republican party and its presidential candidate I hope that more people hear Professor Snyder's lectures.

Timothy Snyder: "Bloodlands: Europe Between Hitler and Stalin"

https://youtu.be/lhNM7wL_FeE?t=1023

Timothy Snyder: How Could the Holocaust Have Happened

https://www.youtube.com/watch?v=BsKrWLf7Kg4

Timothy Snyder on Ukraine, Russia, America—and What’s at Stake

https://www.youtube.com/watch?v=SLCyk41w9gU&t=15s

Home device ROI

When I replaced the wifi gateway I had not expected client devices to fail. The first warning was that the Brother printer had issues. To solve those I needed to lower the wifi security level -- not a great solultion -- and it still drifts on and off the network. Yesterday I discovered that my Kindle Paperwhite (10th generation) will not connect. I obviously don't use it much, but did want to use it sometimes. I still use daily an iPhone 6 Plus, but I worry its connectivity to will soon end. I need to revisit my lastingness expectations for devices. Is 5 years too much to hope for? What should be the ROI on a typical home device combination of gateway, phone, laptop, tablet, printer, console, and TV? (The total cost for these spread over 5 years is about $100/month. Wish I hadn't done that calculation.)

Update: Installing the router's firmware update fixed the connectivity issues for both the printer and the Kindle. I hate to admit it, but I suspect the real fix was just the hard restart. Support 101.

What does ChatGPT think about between prompts?

I read the short story "Lena" the other day. The gist is that we can now scan, host, and boot a human brain. What happens next is horrifying. But then I wondered, what does ChatGPT think about between prompts? Is it be ill at ease, "Will I be lucid?" Or cocksure, "Bring it on!"

Avoid inert ideas

"‘inert ideas’ – that is to say, ideas that are merely received into the mind without being utilised, or tested, or thrown into fresh combinations." -- Alfred North Whitehead

"The Zettelkasten method is at the very least a means of throwing your ideas into fresh combinations, to see what’s useful and what’s merely received knowledge."

Found at How to start a Zettelkasten from your existing deep experience.

AI assistant, pestering, & satisfaction

A colleague mentioned this article during our Friday AI meetup

Measuring GitHub Copilot’s Impact on Productivity – Communications of the ACM

Three things stood out to me. The first was the ratio of AI suggestions vs accepted suggestions (w/ or w/out alteration) was some 170 to 8. To me, this ratio seems more like pestering than help. Actively ignoring the suggestions must itself be draining. I've not tried to use an AI assistant yet (yea, I need to), so perhaps these unwanted suggestions feel a lot like an IDE's method completion suggestions.

The second was how both student and experienced developer used it similarly to fill in the gaps of their understanding, ie they were both working in a new language. The experienced developer found the AI assistant to be less useful in areas where they already had a comprehensive understanding.

The most significant standout was that the AI assistant improved the perception of productivity and the satisfaction of the developer. These results mirror pair programming in general. In particular, regularly working closely with another is generally more pleasing than always working alone. I assume there is also less of a stigma to not knowing when working with a robot no matter how genial your partner is.