“I used AI to combine the data from two excel lists and then send emails to people who were on one list but not another. Saved me so much time.”
My brother in academia, you just fucking discovered mail merge. Welcome to early nineties computing.
Worse often its -
Non-coding prompt = instant response
Basic coding prompt = 20% 5 hourly usage due to the amount of "thinking" tokens
Today's free newsletter is about the dangers of treating AI as being more than it is - a website or app built on expensive technology, with dubious reliability and inherently unsustainable economics.
The rhetoric from AI labs has to change.
www.wheresyoured.at/i-will-never-respect-a-website/
A better way to expose limits of AI is comparing it to things done without AI and how much it bullshits even when giving it prompts with info that should lead it towards the right information.
The current "AI literacy" approach of "prompt AI then verify" suits AI companies "use responsibly" messaging that legitimates AI use and the push towards automatically turning to AI first for anything and everything.
That's also putting aside how the housing book was rewrite of articles published in 1990 which were based on fieldwork done in the 1980s.
Image of first lecture overview including "The private-home market and the state - The Barre commission on housing"
Good way to expose genAI limits: Do a task yourself then try redoing it with genAI.
Working on presentation re importance of housing across Bourdieu's work.
GenAI claimed my abstract was overly bold as Bourdieu's lectures on the state predates his book on housing.
Image - On the State - lecture 1
It, again, truly doesn't matter whether vibecoding actually caused the outages or not - when you start bragging about how claude is making every engineer on staff more productive and the site goes through months of persistent downtime, that's a Statement.
OK, it gets worse. Digital Economy and Society Index for % of 16 - 24 year olds with "above basic" "digital skills" in 2023:
EU average - 38.07%
Sweden - 39.67%
...
...
Finland - 70.58%
Not seeing much evidence to support the strength of Sweden's previous education policy.
Also those concerned about kids missing out on "AI skills", what skills are those exactly?
Most expressing such concerns don't give specifics, but their vague statements hints towards it being how to write prompts and being critical of AI outputs. You can learn that in a day.
A lot of articles mention Sweden's "high level of digital skills" without explaining how that's measured.
The bar for "basic digital skills" is depressing low. EU average 73%-ish & Sweden 89.5% isn't practically meaningful. Most of these could learn in few days.
ec.europa.eu/eurostat/dat...
Re Sweden: Genuinely interested to know what "digital skills" people think kids are learning by giving them tablets & using gimmicky edtech apps.
Not a fan of "no tech" policies, but there is some 1990s "digital natives" style nonsense in assuming giving kids tech teaches useful digital skills.
Holding the small book Neurodiversity: A Very Short Introduction by Robert Chapman and Sue Fletcher-Watson.
New arrival by @drrobertchapman.bsky.social & @suereviews.bsky.social, a brief authoritative text. As far as we know, 1st general academic introduction to neurodiversity studies as distinct field w/ its own body of scholarship. Recommended!
Full disclosure: I reviewed it, applying my due diligence.
Jon Hartley ® @Jon_Hartley_ X.com * Another update to our Generative AI US adoption time series results from our paper "The Labor Market Effects of Generative Artificial Intelligence": we find LLM adoption at work in the US fell over the past quarter (while still up substantially from a couple years ago). 100% Fraction of U.S. Labor Force Using Generative AI At Work 90% 80% 70% 60% (%) 50% 40% 30% 20% 10% 0% May-22 Dec-22 -Pew Survey (ChatGPT use) -I-Bick, Blandin, Deming Gen Al Survey -Hartley, Jolevski, Melo, Moore Gen Al Survey Public Release of ChatGPT (First Public Large Language Model) Jun-23 Jan-24 Jul-24 Feb-25 Aug-25 Mar-26 Oct-26
I like how every study that tries to prove AI is being adopted at scale is like “jobs that AI might be able to do a small amount of are sort of affected” and every other study on AI use is “adoption is low” and “it doesn’t really work reliably or in a way with measurable outcomes”
"Bipolar disorder thrives on inconsistency. While the one-hour change may seem insignificant to those without the disorder, it can disrupt the circadian rhythm of individuals with BD."
ibpf.org/spring-forwa...
A cause I'd be willing to die for: abolishing Daylight Savings Time.
University logic: 10 years on contracts that made me ineligible for early career development programme, yet now on contract where I am eligible for early career development programme I am ineligible for funding for early career staff as been at the university for more than 10 years.
Unfortunately also too old now to rely on the strategies that got me through my degrees - pulling 16+ hour writing stints, smoking 60+ cigarettes, eating weight in chocolate hobnobs, and drinking enough Red Bull to kill an elephant.
Dyslexic horror story: now at stage where averaging an hour per 30-50 words that made it into final draft of a document that still reads like a drunk child wrote it. Worse, without any prep I could verbally explain what want to say in under 15 minutes, yet it's taken 36 hours to write so far.
Pattern noticed in programming debates on LLMs:
LLM critic: "Here's LLM generated slop code and what's wrong with it." + sometimes "Here's also the prompting/scaffolding used."
LLM advocate: "You're prompting it wrong!!! Here's how to prompt it right / scaffolding to use" with **no code** shared.
"From a technical perspective, this [lossy] compression process is much like what happens inside AI models [...]. They ingest text and images, and output text and images that approximate those inputs."
www.theatlantic.com/technology/2...
Terrible situation other ways too. You increasingly get professional award winners who know the right people and the tricks for how to repeatedly get funding; they create feudal academic fifedoms that demand loyalty; they then produce dull research due to how totalitarian and nepotistic they are.
Linux is good now. The author had enough of Windows and Microsoft shenanigans. I'm brave enough to say it: Linux is good now, and if you want to feel like you actually own your PC, make 2026 the year of Linux on (your) desktop
www.pcgamer.com/software/lin...
Almost forgot the obligatory album grid, though some of these are more "which albums I listened to had the most tracks" than ones actually listened to most.
Artist Map and Music By Decade for 2025
My last.fm Playback summary for 2025.
OpenAI didn't think through thier "Your Year With ChatGPT".
In the "Your Chat Style" section my top example was "Stop bullshitting" and number 4 was "That's not what I asked".
Summary sentence ends with "... often demanding thoroughness and evidence over vague generalities".
Most genAI videos in general rely on "techno-magic tricks" that show a LLM quickly generating various things that at a glance look impressive, especially in time it takes, but if spend some time scrutinising the outputs or trying to iterate on them further the plethora of issues become obvious.
Can further combine with polybar whereby:
- Display name of currently active task in topbar, if no active task display name of most urgent task relevant to current calendar event.
- Various task counts (due today, tasks relevant for current event, etc) that can click on to view list of the tasks.
Able to then extend with additional features per event as needed, such as if current event is an "Org" timeblock then also check events for upcoming meetings. If there is no associated meeting note in Obsidian for any meeting, add task to do so to TaskWarrior.