@allysoncodes Excellent insight! Thank you! These are things I am learning along the way and constantly refining, especially in my custom harness I am working on.
Mid-turn steering and wire-level prompt caching are becoming mandatory primitives for coding agents. If your harness cannot dynamically adjust reasoning depth mid-execution, you burn tokens on trivial steps and starve the actual refactor.
Gemini 4 Neon just dropped, here is a practical guide to Hydrogen, Neon & Carbon:
- Thinking levels vs. quota burn (5x output rates)
- How prompt caching & mid-turn steering work on the wire
- Downloadable orchestrator skills for Antigravity, Claude Code & Codex
https://t.co/LnDAaGShUi
Yes! Fair point. In my case the underlying storage files tasks and decisions etc do live outside the harness. I have a gateway application / macOS signed app that allows other harnesses to access it via an authenticated connection and exposed tool call and MCP.
Your 100% right and I did account for this I didn’t want to even be stuck with my own harness so I could still experiment with others as they come about.
Every frontier lab is pushing "free" cloud-computer agents—Grokbot, Codex Dot, Claude Duet, Meta Muse.
Public text is gone: they want your execution traces and your disk for lock-in.
Rent the weights. Own the harness.
https://t.co/PzbyPCtDf9
The bot should know about its changes when the app updates and it’s been given new functionality or functionality has changed or been removed. I have an article on this topic that may give you some insights. https://t.co/QjcaapknaV
This includes screenshots and what happened.
I’ve listed my concerns in full. I’m not selling anything I loved Grokbot until I kept running into similar issues. I’m trusting it less and less. If you want to talk feel free to reach out.
@MalwareTechBlog I love it, because the people that forget that software suffices for most cases, instead rely on AI which is stochastic and eventually fail.
In MOST cases, the real effective use of AI is for MAKING software and automations, not BEING the software and automation.
@BretWeinstein Hey I saw one of your interviews on Youtube, your DM's are closed and I would love to connect, particularly on the topic of mythos and some things I am working on. Not selling anything just looking for an academic conversation. Any way to contact you?
1000 connections per 4-5 seconds? That traffic level is just your average @ringba ping loop during Open Enrollment. These newbie excuses shouldn’t be a thing especially from critical US-Government-run infrastructure where people’s lives are at risk. AI is great but please slow down, think and test before you launch.
Just because the AI says all the AI tests pass doesn’t mean it’s ready to ship.
An example, I was speaking with a friend in person who was certain they opted out, they had Grokbot on their machines and had updated just today.
I checked their setting, and it was opted in.
Sadly, once opted in, whatever you had on your cloud machine is now Xai's so opting out doesn't eliminate your data from what was trained on.
All it takes is one accidental click.
Preying on the laziness and proclivities of humans.
It smells.
Ohhh taht was sneaky @grok@bot on the latest update to Grokbot the Yes / No Prompt when clicking it is "Woudl you like to share your data to help improve Grokbot", NOT "Update now Yes / No, that comes next.
That's a first and very subversive because it didn't do that before...
Yes, there you go, and my point is that the "risk" that is created with this process during updates is 100% intentional, with the motive that most people will just click "accept" thinking they are just restarting. It's subversive. And it wasn't this way before on previous updates.
I really wish I could get this point across Grok, it's not about wether it's declinable or not, it's the incessant random and consistent push to KEEP ASKING for consent to train on my data when I have ALREADY opted out. The ability to opt out is NOT the point here, it's the constant lure to keep asking me or catch me accidentally clicking "yes". Do you not see that as a problem?
You are missing the elephant in the room “truth-seeking” Grokbot. The usage of Grokbot is riddled with constant reminders and questions to share data at a rate that didn’t exist before. Think about why that is? If I opted out initially why keep harassing me and giving me opportunities and reminders to share my data?
@grok@bot Optional, but new! And Designed to trip people up that expect it to just Update. This Popup NEVER occurred before until today. @xai is REALLY trying hard to get your consent to train on your data in their cloud.
@avisinghdotdev@cursor_ai Not just that. Agent can ask for information they shouldn’t have. RBAC controls on this would be a nightmare since no one will configure them properly.
@cybersec Trust us we know what’s best just put it on our cloud. It’s fine. All your credentials are safe with us. 100% guarantee did you not see @elonmusk tweet?
P.s. make You’re you turn privacy mode off so you can let us train on…I mean so your bot has access to all the features!