Nik

I can't code. I install the AI repos devs rave about anyway and report what breaks. honest verdicts, no hype.

[Nik @gitpulled]

https://t.co/p6H13DmdlU

[Nik @gitpulled]

@testingcatalog voice is a different level of personal than text. your voice is basically biometric data, I'd want a clear way to delete what was already shared

[Nik @gitpulled]

@jarredsumner -30% size for free is the kind of win nobody tweets about. does it change startup time too, or mostly disk?

[Nik @gitpulled]

links (official repos only):
https://t.co/x5nLA7gDE5
https://t.co/UcY45GUFBx
https://t.co/T2uoafyZSA
https://t.co/mAU4jVtr9X
full guide on checking if a repo is safe: pinned on my profile

[Nik @gitpulled]

6 GitHub repos you can actually use without writing a line of code.

picked by someone who can't code, so "easy" here means easy, not "just spin up a k8s cluster"

[Nik @gitpulled]

one rule for all six: download from the official repo only. half the "desktop installer" versions floating around are made by strangers, and that's how people get malware.

which one should I test next?

[Nik @gitpulled]

@kimmonismus "3.46B active" sounds like it'll run on anything. it won't. you still have to fit all 78B in memory, roughly 40GB+ even at 4-bit. fast once it's loaded, but loading is the hard part

[Nik @gitpulled]

@testingcatalog ultrafast matters more for agents than for chat. when it runs 40 steps in a row, 2x speed is the difference between waiting 10 min and 20

[Nik @gitpulled]

@tom_doerr 16h without pausing is either grit or a loop it can't see. would love to know how many cubes are actually in the box at hour 16

[Nik @gitpulled]

repo: https://t.co/wXY5gCmJGv
independent review with the real numbers: https://t.co/CH0cXXWlPL

[Nik @gitpulled]

caveman is trending on GitHub with a big promise: make your AI agent talk like a caveman, cut 65% of tokens.

read the fine print. it only shortens what the AI writes. most of your bill is what it reads, usually 5-10x more.

JetBrains ran it on 86 Claude Code tasks: 8.5% fewer output tokens. one tester got most of the savings by just typing "be brief."

Verdict: funny, works, but your bill drops 5-15%, not 65%.

[Nik @gitpulled]

Strands Decider 2B (open weights, Apache 2.0, runs on CPU): https://t.co/mE25enOgYI
Cloudflare Clef: https://t.co/kZEwsP5qWg

[Nik @gitpulled]

new kind of AI dropped this week and it can't write a single word.

you give it a situation and a question, it picks an answer and tells you how sure it is. that's it.

the part that got me: AWS says when their 2B one is 90%+ confident, it's right about 95% of the time. a chatbot sounds equally sure when it's wrong.

Cloudflare and AWS both shipped open ones on the same day. the 2B one runs on a normal CPU, so it's going on my test list.

[Nik @gitpulled]

@Voxyz_ai first mod idea that actually makes sense for non-coders like me. half my confusion is not knowing what it's doing right now. a little guy hammering a file beats a scrolling log

[Nik @gitpulled]

sources:
FakeGit (Trend Micro + Island): https://t.co/inYWUTomnG
fake OpenClaw repos in Bing AI results (Huntress): https://t.co/9sNMj3hdtv

[Nik @gitpulled]

https://t.co/p6H13DmdlU

[Nik @gitpulled]

@techNmak has anyone timed tokens/sec for 70B on a 4GB card? I have a 6GB 1660 Ti and I'm tempted, but I'd like a number first

[Nik @gitpulled]

@MAXdeg0 for anyone nervous about a random .exe: it's open source, PurpleDoubleD/locally-uncensored on GitHub, AGPL, ~1.7K stars. check the repo before you run the installer