runesleo

I’ve been working with AI and vibe coding for so long, but today was the first time I seriously looked at my Codex usage statistics:
Total tokens: 103.82 billion
Daily peak: 4.59 billion
Longest consecutive usage: 119 days
The numbers are much more staggering than I imagined.
But what’s really worth taking stock of isn’t how many tokens were burned, but what those tokens ultimately became.
Recently, I’m preparing to go through the AI automation systems I currently have up and running in full:
Research, monitoring, trading assistance, content production, remote execution across multiple devices

There’s a new model called Jev in the AI space these past few days, and I think it deserves a closer look.
The most counterintuitive thing about it isn’t how much higher its benchmark scores are, but that it doesn’t want to be an LLM that is simply “better at chatting.”
@typesafeai’s approach is:
Many software tasks don’t need the model to “generate a piece of text” at all.
A large number of AI workflows today actually look like this:
state → LLM → generate text / JSON → parse → validate → retry → the program decides what to do next
But what software really needs AI to answer is often just:
Is

Did you claim those $100 from the Grok bot?
I originally just wanted to connect the X paid API to Grok Bot for searches / timelines / mentions.
I added $25 to try it, but total showed approximately $123.
Only after breaking it down did I understand:
• prepaid ≈ $23.7 → the remainder of what I added
• free ≈ $99 → the starting credits granted when connecting (expires in about 1 year
Official statement on 8/29: Paid Grok Bot users get free X API credits to start
The official amount wasn’t fixed; the image above shows my actual test. Check free before connecting, and don’t be tricked by total int

Ever since I got MCPX, I’ve genuinely had this feeling lately:
It’s like I’ve unlocked “unlimited usage” for ChatGPT.
Of course, it isn’t actually unlimited.
The Chat mode itself already has a pretty generous quota, and after installing and configuring MCPX, many things that originally required me to switch software, find tools, or run commands myself can now be done directly in Chat.
And I think the best thing about MCPX isn’t that it “adds another new tool.”
It’s precisely that once it’s installed, you can even forget MCPX exists.
You can keep talking to ChatGPT normally, just like before:
H

Lately, everyone on X has been joining the Thin Muscle Club
I’m checking in too

I recently ran into a very strange problem:
I was clearly using a regular Chat, and after connecting an MCP tool I had been using for a long time, the same conversation would somehow turn into Work midway through.
It wasn't a new Work session, nor was “Work” part of the name of one of my own background tasks. The original conversation really did gain a “Work” label in ChatGPT's sidebar.
At first, I suspected that mechanisms such as remote execution, asynchronous tasks, or long-running tasks had triggered some kind of automatic upgrade in ChatGPT. So I went ahead and broke the problem down, ope

When a robot is about to fail, humans temporarily take over to correct a few steps. Are those few steps really worth teaching the model?
@axisrobotics A single-task simulation experiment report published on August 19 included the following results.
The task was to place a piece of tofu on a plate. The project team tested on the same batch of 480 new starting states:
• Original model: 40.0% success rate
• Directly learning the complete human correction process: 36.7%
• Training on filtered short correction segments: 48.3% (average across three training seeds)
At least in this comparison, direct

I turned my X bookmarks into an Agent inbox that processes them automatically.
From now on, when I see something worth saving, all I need to do is bookmark it.
The Agent handles the rest:
1. Periodically read new bookmarks
2. Remove duplicates and determine whether they have value
3. Assign them to research, content, product, tasks, or no action needed
4. Have downstream workflows actually produce results
5. Record an acceptance receipt
6. Once explicitly enabled, automatically remove bookmarks that have been fully processed
7. Reopen the original post to confirm that the bookmark was actually