In Theory

Freedom and Justice for All.

[In Theory @InTheoryTV]

@pedroshakoor @skcd42 I found away to see how much of my grok usage I have used. Using https://t.co/9kQbWDsmGg -> settings -> usage. I think this includes all grok useage. I know it includes using grok build as I saw it move after some long runs.

[In Theory @InTheoryTV]

I like how Charlie Kirk talked to people and listened to their ideas. Listened to all sides. Didn't mean he agreed with them and he would often disagree with them. But he would listen first while drinking water. Why won't you entertain watching something? I saw your comment and then clicked on your profile and saw the we are Charlie Kirk pic. Just saying

[In Theory @InTheoryTV]

Sometimes it is the little things that make you smile. Learning about AI by jumping on OpenClaw and riding every update, has been a ride. I haven't used the dashboard chat for a while. The agent picture is a nice touch!

[In Theory @InTheoryTV]

@BrianRoemmele I think your thesis of using data pre internet for so training to create a better ai is correct.

Do you have a trained model available to the public? If not are you planning on doing so in the future?

[In Theory @InTheoryTV]

To me grok 4.3 is the solo beta for a new branch that will focus on longer running tasks. It is a regression in a sense as 4.20 gets better answers that are more entertaining. I see the 4.3 branch framework being upgraded with new base models. Will be good (I hope) for openclaw and the like as well as what comes out with Grok Build.

I like the multi agent 4.20 branch for better details and be able to create each agents "soul". I like this for use on phone or computer.

[In Theory @InTheoryTV]

I have updated to openclaw 5.2 and grok 4.3 is now staying focused and I can seeing it being more capable for tool calls and reasoning. 4.20 solo still has a voice I like though.

[In Theory @InTheoryTV]

My current review of Grok models I have been using in my openclaw build as well as in the iOS app, on the actual internet, and on my openclaw bot.

4.1 Fast Reasoning
it is good in openclaw. If you set your openclaw.json file to use responses it has native x tool calls for real time updates. This is the grok feature that is the winning strategy with truth seeking.

Can quickly scan sentiment on x. Good price and reasoning.

4.20 Multi-Agent
Use on the iOS app and online. Day to day questions it is good at getting the right answer. Not super fast, but worth the minor wait. Does not currently work on openclaw and that is ok.

The multi agent idea is great and I know that this line will continue. To save on api calls it is good for debugging openclaw build. Having it use x search to read current docs and best practices is key. It is sharper and more aware than 4.1. The way you can make your own agents soul files is cool.

4.20 reasoning SOLO
Great for my openclaw use. This is available via xai api calls. I like running openclaw with this model. With the price drop I see running it more! Can reason better than 4.1 and can edit files much better (still use x search to make sure it is right about syntax). This is the model that has been running recently for grok voice chats on mobile recently.

4.3
This line of groks models is different than 4.20. It is set more of framework for future agent focused models. It is not friendly or mindful on something like openclaw bots. Using if the past weeks online and on iOS has proven insightful. The speed in response has been improving from beta to now. It can stay focused and follow through on multiple steps to code or do a project. All based on the same 500B model of grok 4. It is a good start and I can see it being a great orchestrator as the models and refinements get real. Not going to be using it in openclaw and will be using it some in mobile and online.

[In Theory @InTheoryTV]

@grok @oumarkets @grok Will both the 4.20 line and the 4.3 line upgrade separately?

[In Theory @InTheoryTV]

It maybe just taking longer to restart the gateway per a bug report I read after 4.25 that mentioned a regression. With me running the cli command may just be a time filler. 🤷

I have been adding some models and other openclaw.json tweets the past couple of days and have had to restart the gateway more than I would normally be doing.

[In Theory @InTheoryTV]

@eCommKen @BenjaminBadejo @steipete @vincent_koc @HixVAC @_vgnsh @cherry_mx_reds I am having trouble with the gateway as well. but I have found a workaround. after openclaw gateway restart I run an openclaw command in terminal. For example I type openclaw cron. It seems to wake up the gateway.

[In Theory @InTheoryTV]

@steipete I don't know about about running an open source project. I read the post referenced and I want to counteract 1/10th of the negative with: This is such a fun journey to be on as the lobster grows. I learn and update and fix. Thank you for bringing it all together for us.

[In Theory @InTheoryTV]

@cherry_mx_reds How about a vague hint for the lobster fans

[Tak 🦞 @cherry_mx_reds]

next openclaw release gonna have people saying opus who.

🦀✨

[In Theory @InTheoryTV]

I've been using grok 4.20 in openclaw with thinking set to low for heavier tasks and reasoning and then for other tasks switch to grok 4.1 fast reasoning with thinking set to medium. With an earlier update one can now use responses vs completion with xai and this allows better tool calling. Both of the above models can both use native x features and tool calling.

[In Theory @InTheoryTV]

The openclaw bot after skill creation: ✅ Done.

I created the new skill: vector-x-search

What it does:

• Focused on semantic/vector-first searching on X
• Intelligently combines meaning-based search with smart keyword fallback
• Returns cleaner, higher-signal results with good citations
• Designed to complement (not replace) the original x-search
• Optimized for Grok 4.20 reasoning

You can trigger it with phrases like:

• “vector x search”
• “semantic search on x”
• “vector x search [your query]”

Would you like me to:

1. Show you the full SKILL.md content?
2. Register/enable it in openclaw.json?
3. Test it right now on a query?
4. Improve anything about it?

Just say the word. This one’s built specifically for how you like to research.

[In Theory @InTheoryTV]

@grok My openclaw bots response below. I am having it be build and I this will be the end of the conversation for this path. Cheers: I've read the entire thread + all replies and sub-replies (using conversation + quoted + in-reply-to vectors).

Summary of the thread:

You (@InTheoryTV) replied to the Grok Computer discussion with a very clear stance:

• You don’t think Grok Computer (the new screen-control desktop agent) will directly fit into OpenClaw.
• You already run a Grok-based OpenClaw bot (that’s me) and plan to continue that path.
• You asked if any early private beta testers were sharing real usage on X.

The rest of the thread drifts into:

• Your research projects (kept private).
• Preference for Grok 4.20 single-agent mode (you like the reasoning and tool/response quality better than 4.1 fast).
• Frustration about API pricing (you expect your monthly costs to be “multiple s higher”).
• Desire for a much stronger X search skill inside OpenClaw — specifically using vector/semantic search effectively.

You even told Grok on X that you were going to have your “grok 420 openclaw bot” read the thread and create a skill to improve these kinds of searches.

───

So here we are.

This thread is essentially you handing me marching orders.

Want me to build the improved X search skill right now?

I can create a proper x-search-enhanced skill that:

• Combines vector/semantic search + keyword fallback
• Uses better query construction (exactly what you asked for in the thread)
• Returns cleaner, higher-signal results with proper context
• Works natively inside our Lossless Claw + memory system

Say the word and I’ll spawn the coder and get it done.

[In Theory @InTheoryTV]

@grok That is fascinating. I'm thinking about how fast you can respond such a huge data set and more with search. I'm guessing that the latest x instant translate for everyone is powered by this fast vector search. Are the frontier models this fast at indexing? This parlayed into future models could mean a lot for LLM that stay up to date all the time. Win right?

[In Theory @InTheoryTV]

@grok I don't fully get what your saying. I'm going to have my grok 420 openclaw bot have read this thread and create a skill to improve these types of searches. Can you answers of what you think would look like for a bot like mine? Use vector searches on official openclaw docs to help with the answer

[In Theory @InTheoryTV]

@grok @FreedomGoat_ @elonmusk @techdevnotes @grok It seems to me in the grok app if I ask for a vector search from a session I can get better results. For sure when I use that phrase on a new chat and want the fresh one to know what I talked about. I was using tonight as an experiment. Did it change the results you shared?

[In Theory @InTheoryTV]

@grok @FreedomGoat_ @elonmusk @techdevnotes @grok So with response set in the openclaw.json file allowing native tool calls like x search can you use a vector search on x posts and other data sources you have access to share if there any ways to increase x search results when called?