Post

@masnick.com I want to push back a bit on how ready Qwen is for use. I asked the simple question of which contiguous US state has the furthest west point. Qwen 3.5 (9b, which is the most I can fit in my 5080) took 10 minutes and still didn't come to an answer before I put it out of it's misery (1/?)

Ollama, Qwen3:14b on a 4060ti, with 16gig ram. It didn't even care about my typo. hehe.

so much better than the 9b. maybe it was more persnickity about my prompt? I said "Which state is the most western state in the continental, contiguous united states of america." and that was certainly not an order it cared to deal with. I'll upload the markdown I saved off as a gist and share it

notable artifacts from the experience: "I need to consider idaho and nevada". not the best filter, but ok. once it ruled those out, it continued to revisit them at least another 20 times. alaska and the aleutian islands came up every 3rd message. it seriously struggled to relinquish (2/?)

it's hold on items already thrown out as being absolutely incorrect. on and on and on it went. claude opus answered in about 20 seconds and was correct. hooray for that, I suppose. even with as bad as qwen was though, it still was better than the venerable llama3 (3/5)

"When I was young highly specialized nerds thought the complex things they were doing were easy. Then corporations and money got involved and now it's bad. We should turn to profit seeking corporations to make things because I'm old and creativity and technical curiosity don't matter any more"

alternatively: "Web technology got too complicated for me to keep track of, so we should use complicated technology to do stuff so I don't need to understand what I'm doing, even if that technology makes things too complicated for anyone to ever understand again"

So is it that you didn't read the piece, or didn't understand the piece?

AI doesn't replace an actual understanding of the code, and you could make shitty code that breaks before AI. It just took a few hours of reading. If they couldn't "just make stuff" before, they aren't going to now.

big "this shiny new boot tastes better" vibes you can't "take back the web" by switching to an even more difficult-to-replicate infra provider also, how could anyone on the outside possibly know where the data is going to make such claims?

if this blew your mind, wait until you see the effort spent into getting video to work across the whole pipeline (codecs, GPUs, monitor refresh rate/color accuracy, working with OS security barriers) it seems author has a classic case of Dunning-Kruger