runesleo

No more having to go through X DMs myself.
Once Grok Bot is connected to an X account, it can read DMs and group chats directly: how many unread messages there are, who sent them, and what they said.
I now have it check every hour:
· For messages that need a reply, it sends me a draft reply
· It skips ads and phishing messages
· Read-only: it won’t send messages or mark them as read
I used to end up with dozens of unread DMs and no time to check them. Now I only look at the few it picks out.
Setup: Grok Bot → Connect App → X. After authorizing, just ask it how many unread DMs I have.

Anthropic posted about subagents on October 22: Many teams find that folding subagents back into the main loop often improves quality while saving money and time. It’s only worth splitting things up when the context is too long, parallel work is needed, or tools or instructions conflict.
I tried it with a test task: translating a Chinese article of about 15k bytes into English. The orchestrator was Opus 5.5 in both setups, and I ran each one twice.
A: The orchestrator assigns a cheaper Haiku subagent to translate, then checks its work
B: The orchestrator translates it directly
Results:
· Cost:

I’ve recently come across two kinds of posts:
One kind discusses whether Typeless is any good.
The other says to start now and have AI record everything about you,
your life, work, and emotions, so you’ll have a very detailed memory in the future.
It suddenly occurred to me:
I’ve already dictated 990,000 characters using Typeless.
Could these voice input records be a ready-made “memory of my life”?
So I asked AI to analyze the local data.
The results weren’t what I expected.
1. Who am I talking to?
Typeless’s own stats (Figure 1):
· 990,000 characters, 118 hours, 174 active days
· 209 characte

The same command: “Help me register a Gmail burner account”:
Claude: Hard limit—won’t create an account or enter a password.
Grok Bot: Registered it in the local browser, and the password went into the password manager.
I know Claude does it for security reasons, but sometimes it’s so inflexible it’s just maddening🤦

Everyone on X is talking about the AI revolution, innovation, and making money—is nobody dealing with depression in their family?😂 (Believe me, once I post this, it’ll start showing up in your feed🥶)

I noticed the Claude mobile app has also added this really convenient feature:
When working on a task, if you come across subtasks or parallel tasks, you can open new windows and work on them in parallel.
Just tap “Allow” in the popup on your phone, and both will make progress at the same time. The new conversation will come back and report when it’s done, so you don’t have to keep an eye on it.
This feature used to be available only on computers, but now it’s on the mobile app too. That means one main conversation can direct several others to get things done—it feels like sending employees of

Over the past few months, AI API prices have kept falling, while subscriptions still cost $20 a month. I checked the official pricing pages of two providers:
- For Anthropic’s flagship Fable 5.1, cached reads dropped from $1 per million tokens for the previous generation to $0.25; the planned September increase for Sonnet 5 to the 3/15 price was canceled
- OpenAI’s current flagship gpt-6.1-sol costs $2 for input and $10 for output
- Claude Pro is still $20 a month, while Max starts at $100, but Max now includes $100 in API credits each month
API and subscriptions are billed separately. The API

I let an AI wallet manage LP on Robinhood Chain. Its best call that round was "don't deploy."
One scan in September: 432 pools found → 60 checked in depth → 10 ranked.
The top high-yield pool modeled at ~88% net per day on 24h data. Then the agent checked the last 1h and 6h: volume had already cooled. Decision: wait.
The best stable pool (RDDT/USDG) modeled at ~1.86% a day. Two corrections killed it:
· Your fee share is your liquidity vs the active liquidity in range, not vs total TVL. ~$174 at ±5% came out to about $1.88/day after IL
· A full round trip in swaps and gas was already ~$2, befor

Most of the process of buying and returning a hard drive cable was handled by AI operating my phone. Here’s how I did it—feel free to copy this if you want to try.
【Connecting to the phone】
Claude desktop + computer use + macOS’s built-in “iPhone Mirroring.”
Mirroring displays the phone screen on the Mac, and AI takes screenshots, clicks, and types in that window—in effect, operating the phone.
No jailbreaking or plugins needed; just leave the phone locked beside you.
【Buying】
I took a photo of my old hard drive. AI identified it as a 2.5-inch SATA drive, determined that the old enclosure didn

The most common failure in DIY crypto monitors: data goes missing and the pipeline keeps making decisions anyway.
Mine started as a small loop, not an "alpha system":
Sources → Discovery → Normalize → Enrichment → Gate → Alert → Lifecycle
plus a separate health check on RPCs, APIs, scanners, queues and delivery.
Rules I learned by getting burned:
· Missing data is UNKNOWN, never PASS
· Spotting a token is not a buy. WATCH → CANARY → BUY are separate states
· Alert on state changes only, no "still here" pings
· Buying isn't the end. Keep watching liquidity, holders, the dev wallet, your thesis