The state of Utah apparently approved plans for a giant data — twice the size of Manhattan, requiring 9 GW of energy (more than the entire state of Utah and its 3.5 million residents use currently) and requiring massive quantities of water in a draught-stricken area. Studies say that the local daytime temperature may rise 2-5 degrees Fahrenheit and the nighttime temperature by 8-12. And even without this data center, Great Salt Lake is already at risk of drying out and releasing toxic dust clouds carrying arsenic, mercury and lead into Salt Lake City.
Less dire than that, but still kind of shocking: the price of Valve’s OLED Steam Deck just shot up by 45% entirely due to RAM shortages caused by the AI industry.
I wasn’t in the market for a Steam Deck, but I was looking for an external backup drive, instead of using a not particularly reliable cloud backup service who’s in the process of awkwardly changing its billing. (I discovered, after finishing my album, that none of my 2026 work got backed up because I hadn’t configured it to — but I had to be choosy about what to back up because my storage is limited.) I went with a WD_BLACK spinning platter hard drive because they’re so much cheaper than SSDs at this point…
The price of a 1TB SATA SSD went from $100 to $275 since November 2025. For a nice NVME 4TB SSD, it went from $325 to $825. Fast DDR RAM just about quadrupled. The price of spinning hard drives also went up significantly, but not as steeply — these also have some RAM for caching purposes, and are probably also used in these AI data centers, and demand has probably gone up just because SSDs are so expensive.
Meanwhile, over this same period the price of CPUs, and even GPUs, have remained fairly steady, for the moment.
The “Stochastic Parrot” paper, published in 2020, predicted 5 kinds of failure that would be caused by AI. One of its authors was fired by Google for refusing to take her name off the paper, but every prediction has come true:
- systems that mimic fluency but have zero comprehension of language or fact are by design incapable of being reliable — and yet fool users and developers into trusting them.
- a training process that assigns higher confidence to more frequent phrases is one that amplifies existing biases and inequalities — with the result that systems based on these tools discriminate against women and minorities more than an average human while people assume they are impartial. (And indeed, it was found that when a man and a woman submit the exact same financial application, the man is granted a 10x higher credit limit; the appearance of the word “women” anywhere in an Amazon job application reduces the score assigned to that application.)
- this same amplification effect works on a cultural level. Less common languages and viewpoints in the training data tend to be filtered out as part of a feedback loop when AI-generated output is used to train AI. Automatic translation of less common languages has demonstrably gotten worse over time. (We also see and feel this homogeneity in AI-generated “art.”)
- the energy cost of training an LLM on a massive data set is large, and this gets far worse with scale.
- massive data sets too large for human review inevitably contain objectionable material, defamatory and dangerous disinformation, child SA, etc. and this iworks its way into the output.
All of this so that people can ask ChatGPT questions they could have used a regular search engine for, make fake photos, put artists and customer service workers out of a job, and get extremely bad medical advice.
We live in the dumbest timeline.
With that, I wanted to point out that DuckDuckGo Search has a no-AI version: https://noai.duckduckgo.com/
(Yes, I was going to try Kagi, but that 100 free searches has me leaving it as a secondary option rather than a primary search option, so far. The Brave browser on Android, I’ve found much faster than Chrome… but its search engine is not so hot. And unfortunately you can’t add custom search URLs to the mobile Brave.)