고정된 트윗
I felt the urge to say something.
@danieldegen/we-are-all-going-to-feel-pain-467cc4066e44" target="_blank" rel="nofollow noopener">medium.com/@danieldegen/w…
English
0x18
1.3K posts
0x18 리트윗함
0x18 리트윗함
Krebs covering the CanisterWorm story:
krebsonsecurity.com/2026/03/canist…
English
0x18 리트윗함
0x18 리트윗함
Bittensor just trained a 72-billion parameter language model across a fully decentralized network. No single company. No central data center. Just a global mesh of permissionless compute nodes collaborating on 1.1 trillion tokens.
The model is called Covenant-72B, completed on Bittensor's subnet 3. It is the largest decentralized LLM pre-training run ever done.
The market noticed.
$TAO is up 40% in 30 days. Trading volume is up 167% over the past 6 months. Engagements on LunarCrush surged 114% month-over-month. Its combined LunarCrush Galaxy Score sits at 70.4, well above the 55 daily average. Sentiment is at 86%.
@Grayscale expanded its TAO trust for institutional access the same week.
Every AI crypto project sells the dream of decentralized compute. Bittensor just shipped a model that competes with outputs from centralized labs. That crosses the line from whitepaper to infrastructure.
The token is still 57% below its November high of $497. Which means either the market hasn't caught up to the technical milestone, or it's pricing in execution risk on scaling beyond a single model. Both readings are worth watching.
At $214 with a 70+ Galaxy Score and rising institutional access, TAO is the highest-conviction AI-crypto convergence trade on the board right now. The question is whether Covenant-72B is the proof of concept or the ceiling.
English
0x18 리트윗함
0x18 리트윗함
Agents are getting tool access everywhere. Most safety measures are vibes-based guardrails that collapse the moment someone actually tries to break them.
Three agent frameworks got jailbroken last week running the same playbook. Prompt injection bypasses the confirmation check, agent calls the dangerous tool anyway. Guardrails are suggestions with good marketing.
Built a formal verification layer using Z3 SMT solvers. Real mathematical proofs that a tool call can't violate safety properties before execution happens.
Features:
- Formal safety property definitions
- Pre-execution verification
- Counterexample generation when violations detected
Code indexing, formal verification, sub-billion parameter specialists. The 2026 agent stack is taking shape.
We spent 2025 giving agents knives. 2026 is about making sure they know which end to hold.
github.com/zscole/ai-poc-…
English
well it has been 6 months and I guess the answer is yes
0x18@NfTity
is this going to be remembered as a one of the top signals in 6 months from now?
English
@NotStewii I mean it does not take much size to move such a small cap but still...
English
$BNKK up double digit third day in row - WHO is bidding this so aggressively with this crypto momentum and $BONK going down?
Stewi@NotStewii
Very unusual price action from $bnkk last two days as it is up 20% in two sessions, despite crypto getting crushed. I would advise to look at the SEC filings as this feels like someone scooping up a lot of shares.
English
I had some spare time again so I targeted @bonk_inu 's company site.
Wanted to modernize @bonkincBNKK in my own style. No dark mode, just pure focus on making $BNKK vibe with investors more.
Kept it clean, transparent, and a bit more serious, without losing the playfulness of $BONK.
English
Caroline Ellison has been released from federal custody!
English
0x18 리트윗함
0x18 리트윗함
BREAKING: FIRST EVER PURCHASE OF $BONK ON @SOLANA BLOCKCHAIN PERFORMED BY A MONKEY
MADE POSSIBLE VIA @jamesbonkai
English
zak.eth@0xzak
Open sourcing a project I built called Gru, a self-hosted AI agent orchestration service you control entirely through Telegram. Most AI coding agents require you to sit at your computer watching them work. Gru doesn't. You can now work from wherever you are and spawn an army of agents to do your bidding. The idea: Spawn Claude agents on your machine from your phone via Telegram, monitor their work, approve or reject actions before they execute. Full control over autonomous AI without being chained to a terminal. What it does: - Spawn agents with natural language tasks via Telegram commands - Supervised mode requires your approval for file writes, bash commands, or any action you define - Unsupervised mode for trusted tasks that can run autonomously - Encrypted secret storage (AES-256-GCM) so agents can access API keys without exposing them in plaintext - MCP server support for extending agent capabilities - Multi-agent coordination with shared context and message passing - Priority scheduling with deadlines - Full conversation history and audit trail The approval flow is the core feature. Agent wants to run a bash command? You get a Telegram message with the command and approve/reject buttons. Agent wants to write a file? Same thing. You stay in the loop without babysitting a terminal window. Fully self-hosted and runs anywhere - Docker, bare metal, your laptop, or whatever. SQLite backend, no external database dependencies. Point it at your Telegram bot token and Anthropic API key and you're running. Open source: github.com/zscole/gru Try it out and lmk what you think.
English
0x18 리트윗함
0x18 리트윗함
‼️‼️BONK TRACKER‼️‼️
Sharing an early version of a BONK tracker I’m building for the community:
bonkshades.online
Very much a work in progress, but feedback and feature ideas are welcome.
English