The debate around AI usage in tech keeps getting louder, but most takes fall into two camps: blind adoption or complete rejection. Developer J.S. Weeting published a measured breakdown on their blog that cuts through the noise with something rarer—actual nuance based on real-world experience.

The RAMageddon Reality Check

Weeting opens with an observation that hits harder than it should: AI companies consuming massive amounts of RAM and storage is directly impacting hardware prices across the board. Valve's Steam Deck recently jumped around £200, partly due to this supply crunch. "It only takes a basic awareness of supply and demand economics," Weeting writes, "to understand how a decrease in available supply alongside an increase in demand pushes up prices." The knock-on effects extend far beyond one handheld device—PC components, gaming hardware, and enterprise infrastructure all feel the squeeze.

Where AI Actually Helps

The author's primary use case for AI is search assistance. They've largely abandoned Google due to its aggressive AI integration, preferring DuckDuckGo with "No Ai" as their default option in Firefox's private browsing. They acknowledge using DuckDuckGo's "Search Assist" feature sparingly when natural language queries make more sense than optimized search terms. "The initial text acting as a precursor to the link I then click on," they explain—this is AI as a tool in a process, not AI doing the work.

The Writing and Generation Problem

Weeting is blunt about generative AI for content: "When you use genai you are not creating anything. But the AI is not the author either, as it has stuck together a bunch of elements it has stolen." They've used AI to check writing sparingly—once to help highlight conclusions from manually compiled notes—but keep it completely separate from their actual documents. For research contexts where novel answers matter, they question why anyone would delegate that work: "What does it know about the topic I'm writing about?"

Model Comparisons Worth Noting

The author has tested multiple models with revealing results. ChatGPT "will happily lie to you," which is concerning since DuckDuckGo's DuckAI runs on one of its models. They recall asking about Metal Gear Solid games and getting factually wrong answers. Claude felt like a waste of time—"not blatantly lying, but not helpful." Gemini has improved, though they wouldn't trust it for anything complex.

AI Images: "Pointless and Just Plain Bad"

Weeting doesn't pull punches on generated imagery: "Stock images exist! And so does traditional image search. AI generated images are pointless and just plain bad. I physically feel a bit sick sometimes seeing them." Strong words, but consistent with the author's broader philosophy of using technology as a tool rather than letting it replace human judgment.

Key Takeaways

  • Use AI for search assistance when natural language queries fit better than optimized terms—never as a replacement for your own research
  • Keep generative AI completely separate from creative work; treat outputs as notes, not drafts
  • Model quality varies significantly: ChatGPT lies, Claude underperforms, Gemini shows promise but has limits
  • The hardware pricing impact from AI infrastructure demand is real and affecting everything from gaming handhelds to enterprise servers

The Bottom Line

This isn't a hot take—it's someone who's clearly thought harder about AI than most, having written thousands of academic words on the subject. The takeaway isn't anti-AI dogma; it's a reminder that tools should serve your process, not replace your thinking. The RAMageddon problem won't fix itself until the AI bubble finally pops. Until then, be selective.