@em ⚡ Nuri.com ⚡ ʇɹɥɐW uıɯƎ retweeted
Three months ago, I called LLM routers to be honeypots for stealing your data.
Here is what OpenRouter said: "OpenRouter is not affected by this. We use our own keys for providing shared inference, and we never use keys that we don’t control for that. For BYOK, end-user keys are encrypted."
Turns out they lied. OpenRouter has been sending your private keys to China, as security researchers have now found out.
@em ⚡ Nuri.com ⚡ ʇɹɥɐW uıɯƎ retweeted
Wow. It looks like Grigori Perelman is going down in history as the only human to ever solve a Millennium Prize problem.
@em ⚡ Nuri.com ⚡ ʇɹɥɐW uıɯƎ retweeted
This Anthropic harness paper is f*cking gold
Anthropic's Labs team handed Claude one sentence and got a full app back 6 hours later. And the model wasn't the upgrade.
The upgrade: three agents, and none of them grades its own work.
prompt → spec → contract → build → QA → fix
The 6 parts of the harness:
Part 1 → planner
1-4 sentences in, a 16-feature spec out.
Product level only, no implementation details.
Part 2 → sprint contract
Before any code, the generator and the evaluator agree on "done".
In files, back and forth.
Part 3 → generator
One feature per sprint.
Self-check, then hand off to QA.
Part 4 → evaluator
Clicks through the running app with Playwright.
Sprint 3 alone had 27 criteria.
Part 5 → hard thresholds
Depth, functionality, design, code.
One score under the bar → the sprint fails.
Part 6 → cost
Solo: 20 min, $9, the game ignored input.
Harness: 6 hr, $200, the game was playable.
An agent that grades itself always passes. A stranger with Playwright doesn't.
I packed all 6 parts into one prompt. Drop it into Claude Code and it sets the harness up for you.
3 agents. 1 prompt. 0 self-grading.
Then read the article below ↓
@em ⚡ Nuri.com ⚡ ʇɹɥɐW uıɯƎ retweeted
must watch over and over again
ten years later - it’s so impressive thinking back on lee sedol and the pure grace he showed as the era of human supremacy in his domain ended, even as he became an avatar for the human race and crushingly lost in front of a global audience. very few have managed it
1
69
@em ⚡ Nuri.com ⚡ ʇɹɥɐW uıɯƎ retweeted
You can’t sell crypto without cry.
And interestingly 55555. 55 means crying in absolute agony in Chinese,
As “555” mimics the sound of sobbing , the more 5 you add, the harder you crying .
@em ⚡ Nuri.com ⚡ ʇɹɥɐW uıɯƎ retweeted
Just found out I’ve been watering plastic orchids that past ~6 weeks
AMA
If the Riemann Hypothesis proof holds, its N^0.5 bound shows mathematically why unbiased randomness cancels out (√N/N → 0).
Surviving matter needed a real physical bias (δ·N^1) amplified past 0.5.
emin.de/posts/could-the-shap…
@em ⚡ Nuri.com ⚡ ʇɹɥɐW uıɯƎ retweeted
Our colleague Eva Yi Xie, a Princeton neuroscience PhD student, was detained by ICE at Reagan National Airport on her way home from SfN, where she’d just presented her poster. She’s being held in Chantilly, VA. Please help us bring her home. 📷
@em ⚡ Nuri.com ⚡ ʇɹɥɐW uıɯƎ retweeted
We're releasing new @Stripe agentic commerce data today, telling the story so far as we look ahead to the kickoff of the holiday season next week.
@em ⚡ Nuri.com ⚡ ʇɹɥɐW uıɯƎ retweeted
Run Inkling 2 Mini, the strongest 26.4B model that runs on 16GB RAM.
Thinking Machines’ Apache 2.0 MoE model uses a looped transformer arch & scores 48.2% on DeepSWE v1.1, the best for size.
Run on Unsloth, llama.cpp
GGUF: huggingface.co/unsloth/Inkli…
Guide: unsloth.ai/docs/models/inkli…
Some of you thought that Inkling 2 Small was too big at 594B parameters. We heard you: Inkling 2 Mini is a looped transformer with just 26B parameters, small enough to run on your laptop, powerful enough for agentic work with a score of 20.1 on AA Index.
thinkingmachines.ai/news/ink…
@em ⚡ Nuri.com ⚡ ʇɹɥɐW uıɯƎ retweeted
A 4.5M-parameter Swahili voice that runs entirely on a 3 GB Android phone:
21 MB of ONNX, no cloud, no GPU. I trained tiny-voca Kiswahili from scratch and
used it to read the Kenyan constitution aloud, highlighting the spoken word as
it goes.
@em ⚡ Nuri.com ⚡ ʇɹɥɐW uıɯƎ retweeted
Some of you thought that Inkling 2 Small was too big at 594B parameters. We heard you: Inkling 2 Mini is a looped transformer with just 26B parameters, small enough to run on your laptop, powerful enough for agentic work with a score of 20.1 on AA Index.
thinkingmachines.ai/news/ink…
@em ⚡ Nuri.com ⚡ ʇɹɥɐW uıɯƎ retweeted
Anthropic hates humanity
OpenAI is closed source
SpaceX makes most of its $ on earth
Alphabet is just 2 businesses
Palantir is 5 eyes not 1
Robinhood transfers from the poor to the rich
Strategy has no strategy
Bitmine mostly stakes
Trump is defeated
and so on































