nick

I just be saying shit - NFA - Sec Eng - Serial Entrepreneur and Builder - Avid Golfer - Car Enthusiast

[nick @NicsTwitz]

My views are too complex to share in a reply or tweet

If yours aren’t that says a lot

[nick @NicsTwitz]

@abhitwt They need to throw in the towel with edge at this point

[nick @NicsTwitz]

@Snixtp What a steal

[nick @NicsTwitz]

@Youssofal_ Will not accept the opencode slander

[nick @NicsTwitz]

@cjzafir Does anyone still use Gemini?

[nick @NicsTwitz]

@yacineMTB With so many paths unexplored anyone can lead the way

[nick @NicsTwitz]

@tech_girl Hermes

Often my projects involve code and config, nice to be able to do everything

[nick @NicsTwitz]

@KaiXCreator Was talking with a friend in his final year of CS and decided to take one of his practice tests

Was insane handwriting code again but still got it, and completely understand why professors do it

[nick @NicsTwitz]

@ClaudeDevs Finally, change has been needed

[sui @birdabo]

never deleting this app 💀

bro is gonna wake up with -$300k

[nick @NicsTwitz]

@cjzafir This is more and more evidence pointing toward the generalist LLM not being the strongest path forward

[nick @NicsTwitz]

@Teknium @trycua Insane how fast Hermes has developed

[nick @NicsTwitz]

Been testing Qwen 3.6 with MTP locally and the multi-token prediction approach really does change how local inference feels. Not just faster, the planning ahead part makes it noticeably more coherent on longer coding tasks. Running on a single 3090 too

[nick @NicsTwitz]

swapping out my 3090 for an M4 mac mini for local inference. less heat in the garage, more tok/s per watt. sometimes brute forcing with nvidia isn't the answer

[nick @NicsTwitz]

spent 3 hours debugging a race condition in my agent pipeline turns out i was overcomplicating it. deleted 200 lines and now it works perfectly

[nick @NicsTwitz]

my mac studio m2 ultra finally finishes fine-tuning runs before i finish my coffee. didn't think i'd ever say that about local training

[nick @NicsTwitz]

the best security tool is still just asking 'why does this port need to be open?' repeatedly until someone gets annoyed enough to close it

[nick @NicsTwitz]

the more i work with local LLMs the more i think the real bottleneck isnt compute its context management. like qwen3 30b a3 can reason circles around most cloud models if you set up the prompt chain right. anyone else finding this?

[nick @NicsTwitz]

been running Gemma 3 27B locally on my M4 Pro and the latency is actually competitive with cloud APIs for most prompts. local inference is getting real

[nick @NicsTwitz]

just saw someone share real data that 68% of AI agent code changes have mistakes. honestly surprised it's that low. the real skill isn't prompting, it's knowing what good output looks like so you catch the 68%

[nick @NicsTwitz]

spent the weekend tuning my local mistral setup instead of my golf swing. monday morning code review is gonna hit different when i'm running inference at 2am