CyrilDeFi

$38 MILLION IN BITCOIN WAS JUST STOLEN FROM COLDCARD HARDWARE WALLETS. IN UNDER 25 MINUTES.
Not a phishing attack. Not a hack of the device itself. A firmware bug in the wallet's own random number generator.
Root cause: a macro that should have enabled the hardware's true random number generator was misconfigured, and a supporting library only checked whether the setting existed, not whether it was actually working. The fallback silently used MicroPython's software generator, seeded by the chip's serial number and internal clock, neither of which is secret.
Roughly 594 BTC drained from about 5

2.8 trillion parameters. 1 million token context. Native multimodal.
Kimi Delta Attention delivers up to 6.3x faster decoding at million-token context, cutting KV-cache memory by up to 75% along the way.
Attention Residuals deliver roughly 25% higher training efficiency at under 2% additional compute cost.
Built for long-horizon agentic coding and self-evolving workflows, sustaining long engineering sessions with minimal human oversight.

BTC/USD - my read on the daily
O: $65,062.78 | H: $65,337.38 | L: $64,792.00 | C: $65,218.87 (+0.28%)
I'm not excited by this candle.
We're still trading under the 200 EMA ($73,696), and that line's been sloping down since April. Until that flips, I'm treating this as a downtrend, not a dip to buy blindly.
June's breakdown from the $72K–$76K zone was the real damage — that was demand turning into supply, and it dumped us into the low $60Ks fast.
What I like: price has been carving higher lows since that flush, riding a trendline off the June bottom. That's the first real sign of buyers steppin

this is f*cking gold
KARPATHY JOINED ANTHROPIC FIVE WEEKS AGO TO TEACH CLAUDE HOW TO IMPROVE ITSELF.
Not as a researcher who uses Claude. As the person building the team that uses Claude to accelerate its own pretraining.
His own words on joining. "The next few years at the frontier of LLMs will be especially formative."
The man who taught millions of engineers how neural networks work is now inside the lab building the next version of the model.

SOME USERS ARE CALLING GROK 4.5 ONE OF THE MOST NEUTRAL, TRUTH-SEEKING MODELS AVAILABLE.
There is a real study behind that claim now.
The Neutrality Project tested 18 models from 12 labs and found Grok's family sitting closest to the political center, while every other flagship model (GPT, Gemini, Claude, Llama) clustered left of it.
That said, this is a live and disputed topic.
Grok has also been the subject of extensive reporting on direct political tuning by xAI in response to Musk's own reactions to its outputs, including incidents where those tweaks produced genuinely extremist responses

Fable 5 flips to pay‑per‑use in 48 hours. Burn the remaining runway to make sure you don’t actually leave anything behind.
The play: have Fable comb your repo and draft a complete skills library before the timer hits zero. Frame it as a retiring principal engineer doing a final handoff to the team, and it snaps into a totally different gear.
On the other side you get 10–16 skill files—debugging runbooks, change policies, the painful failure tales that ate days. Every instinctive judgment it would’ve made silently is now documented and reusable by Opus and Sonnet for pennies.
Single pass. About

This site has tons of agents loops you can use right now
Maximize your prompting using these

NOTEBOOKLM, AND OBSIDIAN TURN LEARNING INTO AN ACTUAL SECOND BRAIN SYSTEM.
NotebookLM reads and summarizes your sources, grounded only in what you actually uploaded.
Gemini expands on it, drafts from it, thinks through it with you.
Obsidian stores and connects all of it permanently, linked to everything else you have ever learned.
Three tools. One loop. Knowledge that compounds instead of disappearing the moment you close the tab.