Josh Bolding

[Josh Bolding @BoldingBuilds]

I got Rodriguez but dang I want Fluffy to win at this point! So much heart!

[Josh Bolding @BoldingBuilds]

I played with Ox alpha last night and asked my usual Tianamen Square questions that Chinese models always refuse to answer, and it had no problem trash talking China and the CCP. I came away thinking it for sure wasn’t a Chinese model, but I see everyone saying it’s in GLM family.

[Josh Bolding @BoldingBuilds]

@MiaAI_lab Is it an American company?

[Josh Bolding @BoldingBuilds]

Benchmarked it this week. Same 3090, same VRAM, identical config, vs Qwen3.8-27B.

Verdict: fastest model in its class, and it still loses.

2.15x the throughput. But Qwen won all 5 agentic missions, 74.4 vs 60.7, and finished 30/30 runs to Ornith's 77%.

Trying to put the full numbers together for a post

[Josh Bolding @BoldingBuilds]

Its pretty wild...I used uncesored 27b to attack my spare router.
It deauthenticated a client device, captured the 4-way handshake when the device reconnected, and then used a word list to crack the .cap file using my 3090 rig and handed me my wifi password. Blew my mind.

We have entered a new exciting, yet terrifying time in cyber security

[Josh Bolding @BoldingBuilds]

My Obsidian is the single most powerful thing in my workflow. Set up a second brain if you havnt yet

[0xSero @0xSero]

It's so over for me.

Obsidian Pilled.

So good for agents, so good, so good.

[Josh Bolding @BoldingBuilds]

@birdabo This is how the horror movie starts 😅

[Josh Bolding @BoldingBuilds]

@JohnJenocide @JoshuaWorth @vasuman It’s impossible to ban it like this once released, but it still costs millions to train a new 27b model. The labs is who they would target

[Josh Bolding @BoldingBuilds]

@Joash0x @vasuman Shoulda done that 6 months ago 😅

[Josh Bolding @BoldingBuilds]

PSA for Claude Code users: by default it silently deletes your local session transcripts after 30 days (cleanupPeriodDays = 30). If you ever want to grep old sessions, resume them, or audit usage, add this to ~/.claude/settings.json:

{ "cleanupPeriodDays": 3650 }

Codex CLI keeps everything by default, Claude Code doesn't.

[Josh Bolding @BoldingBuilds]

@Hikari_07_jp Iv been thinking about this same thing since 3.8 came out. It definitely feels like they have figured a way to put most of the parameters towards reasoning vs knowledge

[Josh Bolding @BoldingBuilds]

Codex app makes my mouse freeze on screen, like a delay/stutter especially right after I click a thread and it’s loading in. It’s the only app I have seen do this on my pc. Iv tried to have agents fix it but no luck, I’ll just keep hoping it’ll fix itself on one of these app updates smh

[Josh Bolding @BoldingBuilds]

@sama The interesting part will be if China agrees to slow down in the name of safety as well. Something tells me I doubt they will

[Josh Bolding @BoldingBuilds]

@WearForbidden Lmaooo videos like this are my favorite use of AI

[Josh Bolding @BoldingBuilds]

Claude is having a rough Tuesday

[Josh Bolding @BoldingBuilds]

This is so real…the most exciting release this month and no one to talk with about it 😅

[Lotto @LottoLabs]

I wish anyone I knew in rl knew how big a deal qwen 27b is

It’s just me and you bros

[Josh Bolding @BoldingBuilds]

I wonder what Alibaba figured out to pull this off, and does this mean those frontier models it beat could apply it as well?

[Alok @analogalok]

Holy moly. 🤯

Qwen 3.8 27B just scored higher than:

GPT 5.6 Terra
GLM 5.2
DeepSeek V4 Pro
Muse Spark 1.2
Claude Opus 4.8

on the Artificial Analysis Agentic Index.

And you can run this on a single RTX 3090/4090.

Go show your GPU some respect. 🫡

[Josh Bolding @BoldingBuilds]

Day 2. Same instrument, first MoE.

https://t.co/hclXdNCHPd

[Josh Bolding @BoldingBuilds]

Got BoldScope working on an MoE. It did not disappoint.

The magenta wedges that flare are the ones the router actually picked. Not a random shimmer.

Nemotron Lightning 30B-A3B. 128 experts, 6 per token.

Q4 - 2× 3090 - locally

[Josh Bolding @BoldingBuilds]

Qwen 3.8 27B, locally, hood open.

Each disc is a real block. Brightness is how hard that block worked on this token versus its own baseline.

~8 tok/s because I am reading every layer. That is the cost of a real signal.

BoldScope.

[Josh Bolding @BoldingBuilds]

Yesterday this needed a PyTorch probe and ran at ~8 tok/s. That was me hooking every block in eager mode.

This clip is the GGUF, through llama.cpp. Same per block signal, plus the MoE router top-k, streamed off the graph. The collector costs about 2% versus a normal decode. That’s why this one is 44 tok/s.

Wave travel is still animation. Which experts fire is measured.

[Josh Bolding @BoldingBuilds]

Got BoldScope working on an MoE. It did not disappoint.

The magenta wedges that flare are the ones the router actually picked. Not a random shimmer.

Nemotron Lightning 30B-A3B. 128 experts, 6 per token.

Q4 - 2× 3090 - locally