Extraordinary claims require extraordinary evidence.

Thank you for sharing.

I also tried vibe coding with Qwen3.8 (just give it a few lines of what you want, and step away). and you’re right, it does… something, but not at all how you want it. (obvious in hindsight)

Your Step 4 is how I use AI now. I use Deepseek Harness. I don’t use a dedicated plan mode, but I do tell it ‘let’s plan and write a TODO .md for a <task>’, and we discuss ways of implementing something, what libraries we can use, pit-falls, future-proofing etc., and it will write whatever we decide on in a TODO file. Only after each step has been discussed, I tell it to review the TODO and ask me if it has any more questions. Once it’s satisfied, I verify what it wrote in the TODO one last time, and when I’m satisfied I tell it to implement the TODO we just planned.

TODOs work quite well if you make and review them with the model before it implements it.

Not scientific, but I sometimes get Gemma4 check Qwen 3.8’s work because I noticed when a model knows its not their work, they’re a lot more critical. Might be helpful to get another model to review tests.

She’s a North Korean defector propped up by the West (in this picture, she’s on Joe Rogan’s podcast). She says a lot anecdotal things about North Korea that cannot be verified, and people (like Joe) believe her without question. This meme has her saying real things that are happening in the US right now, and that’s exactly how she talks about North Korea. It’s ironic, because in this meme’s case, while what she is saying is just as crazy, we in the West can actually verify what she’s saying is true.

Edit (for context):

Girl: Yeonmi Park.

Person shot in the face: many, probably ‘Renée Nicole Good’ or ‘Alex Jeffrey Pretti’, but without papers? ‘Silverio Villegas González’ maybe?

I’ve used pi.dev’s default compaction (this was a few months ago, it may have improved) and it lobotomized the model, so I kept if off, and made my model do a handoff for the next session. However, Deepseek harness’s compaction is very good. Unfortunately, if you use under 264K context, you have to modify its settings because its point of compaction is TOTAL - max_tokens - 16k reserve, and max_tokens is 64k. I had to modify the settings down to 16K (using Qwen 3.8, because there’s a bug in how DSH reads settings and it wasn’t working when I did it), and I still compact at 73% (88K) on 120K context (ie: 120 - 16k - 16K). But it’s really good. I’ve gone to 2M output tokens (100M cached) until I thought it was getting too dumb for the task and started a new session.

So yes, compaction can work well. Make sure your prefill speed is 1000tk/s+ or else it will be painful.