The Legend Of John Henry’s Hammer & other AI-related stuff we mostly cut from the Detox

I’m doing way too much AI-related research. It is maddening to try to grasp functional truth and applicable terms in a world of hyperbole (good and bad) and definition creep that gets tied up with every other major issue in our society. My death spiral to the root problem being capitalism is not particularly helpful (even if it’s accurate). This rambling post will traverse a bunch of stuff that we’ve mostly cut from the existing Digital Detox articles. There is no real path or pattern to this. It just felt irritating to throw a chunk of thought and writing away. Plus, I’m part of the blogging renaissance and, on the quantitative side, this post will increase my links-per-post ratio quite a bit. I also forgot that writing in the visual interface wipes out my footnote html blocks. So if you see weird sentences, pretend they are footnotes.

Water

Want to chase your tail? Try to figure out how water use plays out at a macro scale by industry in the USA. How can we have any useful conversation if we can’t contextualize water use? Is water usage from cooling tower evaporation the same as other types of water usage? Like if I put water in paint (a fairly heavy industrial use– accounting for 5% of US water withdrawals or 12% of public water), how do we think about that? Does the water return in some pure-ish form when the paint dries or is it forever tainted? How does that compare to water use in paper manufacturing? Do cooling towers and water vapor produced from evaporation impact local precipitation?¹ What’s the impact of waste heat? Should we treat data center’s on-site usage of water differently than off-site water usage for electricity generation? What do I need to know to have a decent conversation about this? I might know enough now to have more questions.

Energy

It seems overly-simple to me to compare AI queries to Google searches (in that limited space where AI use parallels a Google search) without comparing the energy impact of Google’s ad infrastructure and the result of clicking through to the web page where the answer lies. How do we even think about the energy use of the robots that crawl the internet, the servers that store all the caches, etc. etc.?

I hear people blame AI for all the energy problems but the data I see indicates data centers used 4.4% of US electricity in 2023 and may hit somewhere between 6.7% to 12% in 2028. That increase is bad but it should open up a lot of questions about the other 95.6% of electricity usage. These conversations should be as simple as possible, but not simpler. So many criticisms just feel like click-bait headlines shouted in an echo chamber. That’s both on the pro and negative side. AI is part of the problem but a smaller part than a lot of conversations seem to indicate. It is new-ish which is bad because adding energy-intensive things is bad . . . and good in that technologies tend to get better in terms of efficiency (if for no other reason than it saves money). We’ve seen some of that already. Hooray?

Replacing steel driving virtuosos

To be fair, this is referenced in week two in a less comprehensive way.

With a million tabs open across three browsers and four different profiles², Johnny Cash’s “The Legend Of John Henry’s Hammer” came on. I’ve heard it many times. It’s a great song but given I was trying to map AI to the industrial revolution at that particular moment it seemed particularly pertinent and ran quite contrary to people assuring us that AI wasn’t going to replace jobs because they were multifaceted. My opinion there isn’t driven by AI hype, but rather by beliefs around how far business will go to save money and avoid the complications of humans. Enshitification is not just for online products and services.

John Henry can hoist a jack, he can lay track, and he knows how to pick and shovel too. The man is the best hammer swinger around! The thing is he’s an exception (and the captain seems to hate him for his size, skill, better pay heck — 35 cents³ a day for driving steel) . . . and he is better than the machine. Unfortunately, he dies proving that and leaves his wife(?) to try to take his place as a cog in the industrial machine because she can hoist a jack, she can lay a track, she can pick and shovel too.

1000s of AI researchers & extinction events

Very few fields of research have experts surveyed on whether success will lead to authoritarian control, human extinction, human economic disempowerment, mass existential crises . . . and even fewer have a large percentage of experts who think that there’s a pretty good chance some of those dystopian outcome will occur.  Mix that with 70% of AI researchers feeling that society should prioritize AI safety research more (36%) or much more (34%).

“As measured by the percentage of respondents who thought a scenario constituted either a “substantial” or “extreme” concern, the scenarios worthy of most concern were: spread of false information e.g. deepfakes (86%), manipulation of large-scale public opinion trends (79%), AI letting dangerous groups make powerful tools (e.g. engineered viruses) (73%), authoritarian rulers using AI to control their populations (73%), and AI systems worsening economic inequality by disproportionately benefiting certain individuals (71%).”

source

K12 teaching

These opposing concerns are also key issues in education that we’ve been aware of for many years- the challenge of supporting personalized learning and the effectiveness of assessment in education. With teaching now seen as a low-paying, high stress, and unsafe career, it’s actually surprising that even 18% of Americans would be likely to recommend it as a career. It is not surprising that 86% of US K12 school systems reported issues hiring enough teachers in the 2023-24 school  year, that  K12 teachers have one of the highest burnout rates of all industries (44%). With wages that continue to decline when compared to the average for other US workers, it’s clear there are foundational issues that are entwined with larger cultural issues. These aren’t AI problems but when sectors have long-standing and unaddressed foundational issues, AI can be seen as the “solution.”

Pizza

Amazon’s Prime Air uses autonomous drones to deliver packages from the warehouse to homes.  Claude’s personal and computer-focused interactions show one end of the spectrum while Prime Air shows a corporate focus with AI controlling machines in the real world. There have also been a variety of examples of AI interacting with humans to do things like ordering chocolate covered strawberries, or (more relatable) ordering pizza (semi-successfully and not so successfully) and there is already a PizzaVOICE AI ordering system available to take pizza orders.

Terms

Multimodal AI – AI that can consume, process, and produce various types of media. Many models today handle documents, audio, and images in addition to text. We’ve had voice recognition and speech creation for some time (Siri, Google Assistant, and Alexa for instance) but new models are more sophisticated and powerful. The voice assistant race is in full gear.

Slop – Just like spam is associated with email, slop is the low-quality, unreviewed content associated with generative AI. It can be strangely viral (and seemingly without purpose) like Shrimp Jesus, shoddy search summaries, part of disinformation campaigns, or product reviews but it’s often just filler content created on various services to generate ad revenue.  This low-quality content has infiltrated spaces as diverse as lace-making and mushroom hunting.

Context Length – Context length is the amount of information a model can hold in working memory. This is a key aspect of more complex interactions. Contex lengths have been increasing rapidly and that expansion has positively impacted AI interactions.

Adding to the confusing terms, we are seeing more and more examples of AI integrations that leverage multiple types of AI for a single (or small set of) purpose(s). For example, LLMs are more and more frequently the bridge between human requests and other technologies (AI and otherwise). When you provide a prompt for an image generation tool the LLM provides the bridge between your words and another technology which does the actual image creation. Those other “hidden” models (TTS, image creation, etc.) make discussions of isolated capabilities more complex.

Spotify

**I didn’t link this one up yet. I still have some hope I’ll use it somewhere.

Spotify makes for an interesting example to see what AI integration starts to look like across a spectrum in a company where you wouldn’t necessarily expect it. Spotify gathers data on users’ listening habits (source data), uses AI to drive recommendations and playlists, creates ai animations, while AI-generated music proliferates, and Spotify is strongly incentivized to use artists where royalties are lower/not needed. It’s easy to see where Spotify might cut out the actual artists entirely and start creating their own virtual music that’s been built on the listening habits of individuals and created based on the music of the real artists it’s no longer paying.