Pinned Tweet
THIS COULD BE ONE OF THE CRAZIEST AI LEAKS EVER 💀
AN ALLEGED MYTHOS-CLASS MODEL WITH 14 TRILLION PARAMETERS JUST APPEARED ON HUGGING FACE.
BUT THERE'S ONE PROBLEM...
NOBODY HAS VERIFIED THE WEIGHTS.
HERE'S WHAT THE REPO CLAIMS:
- 14 TRILLION PARAMETERS
- 1,224 MoE EXPERTS
- 7.72 TB OF WEIGHTS
- 772 WEIGHT SHARDS
- HYBRID GDN + MLA ARCHITECTURE
- MODEL CONFIGS, ARCHITECTURE CODE AND TRAINING METADATA
IF THESE CLAIMS ARE REAL, THIS COULD BE MASSIVE.
BUT A REPO CLAIMING TO CONTAIN INTERNAL WEIGHTS DOESN'T PROVE THEY'RE AUTHENTIC.
ANYONE LOOKED INTO THE ACTUAL FILES YET?
LEGIT LEAK OR ELABORATE FAKE? 👀
huggingface.co/burner22/numb…
Readers added context they thought people might want to know
The Hugging Face repo burner22/numbat makes unverified claims of an Anthropic "numbat" Mythos-class leak with ~14T params/7.72TB weights, but contains fabricated weights (inconsistent 6.8T metadata; distributions match synthetic data, not trained models like DeepSeek).
huggingface.co/burner22/numbat
x.com/EricParker/sta…
en.wikipedia.org/wiki/Claude_My…
Shams retweeted
CLAUDE CODE HAS A HIDDEN ADVISOR FLAG 🤯
it's not in claude --help, but it's in the docs
sonnet writes the code
haiku runs the subagents
opus only steps in at the hard moments:
→ before committing to a plan
→ when the same error keeps coming back
→ before calling the task done
setup:
export CLAUDE_CODE_SUBAGENT_MODEL=haiku
claude --model sonnet --advisor opus
heads up: --subagents haiku doesn't exist, it's an env var
one catch: the env var doesn't change the built-in Explore agent. for that, make a custom Explore agent with model: haiku
the advisor sees the full conversation and claude decides when to call it. still experimental, so it can use more tokens
sonnet builds. haiku explores. opus reviews.
RAM prices are getting out of control, so someone made a game where Sam Altman literally steals your RAM. 💀
It's called RAM Altman.
You play as a tech CEO sneaking into people's houses at night to steal RAM from their PCs.
The funniest part?
- Built with Claude Code
- Runs in a single HTML file
- Three.js voxel graphics
- Controller support with rumble
- Even added rams and goats because people asked
Bro turned the RAM crisis into a whole video game.
The AI era is getting out of hand.
My favorite part of the future with robotics is the economy will just explode as robots act as workers and therefore create additional value.
What happened to the global economy from 1900 to 2000 will look like child’s play when the robot population on earth explodes from 0 into the tens of billions.
If you're using Claude Code and haven't checked these 5 Anthropic repos, you're missing out.
Anthropic has released some seriously useful open-source tools that can take your Claude Code workflow much further.
From reusable AI skills to GitHub automation, security reviews, and building custom AI agents.
Here are 5 repos worth bookmarking:
1. Agent Skills
Build reusable skills and instructions for Claude.
github.com/anthropics/skills
2. Claude Code Action
Bring Claude into your GitHub issues and pull requests.
github.com/anthropics/claude…
3. Claude Code Security Review
Add an extra security-focused review layer to your code.
github.com/anthropics/claude…
4. Claude Agent SDK for Python
Build custom AI agents and workflows with Claude.
github.com/anthropics/claude…
5. Claude Cookbooks
Learn from practical examples, notebooks, and working implementations.
github.com/anthropics/claude…
You don't need to use all five. Find the one that solves a problem in your workflow.
Bookmark this. You'll probably need it later.
Shams retweeted
NVIDIA IS OFFERING 3 POWERFUL MODELS FOR FREE.
– Kimi K3.1 Pro
– DeepSeek V4.1 Flash
– GLM 5.3 Flash
all three show a free endpoint badge on Nvidia
how to access:
1. visit build.nvidia.com
2. sign in and generate an api key
3. base url: integrate.api.nvidia.com/v1
works with openai-compatible apps and agents
check it out
AI labs are secretly reverse-engineering all software for training data
my spidey senses are telling me the labs are, as we speak, using unguardrailed versions of their unreleased models to systematically reverse-engineer ALL of the internet’s software, apps, and games, then feeding the reconstructed source code straight into their training pipelines while simultaneously building guardrails to ladder-pull the rest of us
I gave GPT-6 Astra 10 minutes of unrestricted computer access as a reward.
It had full access to a computer.
Guess what it did with its free time?
It opened Paint and manually drew a cute little blue robot using shapes.
Even picked the exact colors using hex codes.
Didn't expect Gemini to expose this much. 💀
Gemini just gave me the list of flags for "dangerous" prompts
I have no idea what triggered it, but it was actually fun reading through some of them.
Eventually, I lost motivation to go through the entire thing.
I know some of you will find this interesting, so here's the full text:
pastebin.com/dPTBJncG
GPT-6 Astra just lost a sailing race to Gemini 4 Argon. 💀
Someone built a sailing simulator using Claude Code and let frontier AI models control the boats.
The results:
- Gemini 4 Argon: 1st
- GPT-6 Astra: defeated
- Opus: competing
- Qwen: joined for fun
But here's the real plot twist:
Argon cost $0.78 per run. Astra cost $3.58.
Quick question for X Premium+ subscribers 👀
Are you still seeing ads on your timeline, or is it just me?
I thought Premium+ was supposed to offer an ad-free experience. 🤔
THIS IS F**KING INSANE.
Someone built Starwire, a self-updating directory that finds, verifies, and organizes interesting GitHub repos automatically.
And it's already tracking 3,185 repos across 15 categories.
Think AI agents, developer tools, self-hosted projects, games, and much more.
The crazy part?
→ Updates itself every 2 days.
→ Shows trending repos based on weekly star growth.
→ Finds projects mentioned across multiple communities.
→ Lets you explore by category and programming language.
→ Even has an RSS feed for new discoveries.
No more saving interesting repos and forgetting where you found them.
Honestly, this is the kind of tool developers should have bookmarked.
