Added support for Gemma 4 12B to "Little Control Room" via @ollama . Using it now for project management activities. Running 12 tok/s on my M2 MacBook Pro. Might try later for simple coding as well. https://t.co/szCZTyTOxb
@ivanfioravanti Naive memory system will take a note and turn it into a hard fact. A single piece of memory should be built progressively... reinforced across multiple sessions, and in many cases probably just forgotten.
@orionintx@petergostev ...and I think we're getting close with code, because top models make so few mistakes and can write a lot of test code to catch them anyway (I've only been using GPT, not sure about Opus these days)
A reminder that most people are not very bright or not very motivated, and they'll find a way to be mediocre no matter what tool you put in their hands. I'm cool with that.
@melamedp I guess production is already saturated... but could keep 25% of the programmers + top AI model. Maybe they don't have 25% of programmers that are effective at AI.
I can definitely do now jobs I could never have done before, but I invested considerable time to adapt.
@AdolfoUsier@jetbrains one would be crazy not to... no downsides in getting a smart LLM to perform a review. I didn't see the interview, but I guess the Zig team probably does, even if they don't admit it.
@levelsio > in Aman Tokyo the hallway reeked of weed
Weed is very uncommon in Japan, so in a sense it's still an experience away from the general population 8)
@melamedp Most people out there are NPCs. They can't use AI meaningfully because they are a cheap form of AI themselves. This includes modern programmers, those that think that JavaScript, Rest and all that shovelware is intellectual work.
Short on Codex usage, I'm doing real work with alternative models:
- DeepSeek V4 Pro: very cheap, but 6/10 engineer.
- Xiaomi MiMo V2.5 Pro: very cheap, but 7/10 engineer.
- Kimi K2.6: 8/10 engineer, but could easily spend $5-10/day
@melamedp Refers to the news: "JUST IN: Uberโs COO says heavy AI spending is getting harder to justify, as higher token usage fails to show a clear payoff in consumer features."