Zachary Blocker

Just built The Golem Awakens! A complete single-file HTML5 platformer & puzzle game entirely with local AI on one RTX 3090 Ti over 10 working days. 3,050 lines • 2,038 functional JS • 424 lines JSDoc • 22.4% comment ratio • Zero dependencies • Zero build step • MIT licensed • 6 chambers Play it now on the Tholem Labs website! (opens in any browser): tholem.ai Repo + full docs: github.com/Tholem-AI/The-… This was a deliberate test of Hermes @NousResearch + local models for real interactive work. Thread on the workflow and results 🧵

@PradyuPrasad Most don't even have a concept of context length or what a proper harness can acheive or know anything beyond frontier web chat boxes.

i think if people outside the AI bubble saw codex computer use they'd freak out

We want persistent chat that can be hosted on our own server. Thinking of having Hermes agent connect in as a client and save channel chat in a database and serve as a general purpose bot for all sorts of things im addition to chat recall! (Unless private-server side chat is on the way?)

@Teknium @ZacharyBlocker

@Teknium Are you aware of anyone working on a Hermes MCP harness for TeamSpeak 6?

@Symedia_ @Teknium @teamspeak Tru, have a friend that decided to host a TS6 after recent Discord changes. We need persistent chat and some bots, Hermes integration would be great for our use case. Might have to build it ourselves

@Teknium @ZacharyBlocker only @teamspeak noise we see is when discord does something stupid

@Teknium It does but not super well supported, discord pisses off people every once in a while so it still has a pulse :p

@ZacharyBlocker I didn’t know teamspeak still exists

Just built The Golem Awakens! A complete single-file HTML5 platformer & puzzle game entirely with local AI on one RTX 3090 Ti over 10 working days as a deliberate test of Hermes @NousResearch + local models for real interactive work Play it now on the Tholem Labs website! (opens in any browser): tholem.ai Repo + full docs: github.com/Tholem-AI/The-… See the full thread: x.com/ZacharyBlocker…

If you’re exploring agent-assisted game dev or local AI workflows: • Start with explicit constraints + architecture docs before heavy prompting. • Treat documentation as part of the implementation. • Use stronger models narrowly for research and high level architectural decisions, then feed insights back into your local harness. • Single-file or minimal-structure projects are excellent testbeds- easy to test, share, and iterate. But larger projects cannot be single file- so clear modular code and documentation becomes extremely important. • Test after every change, and keep humans in the loop for every implementation until certain workflows are well established and tested. • You trade context window for token speed, proceed accordingly. • If something isnt working well or your agent times out, revert to a previous state and refine your prompt- often easier than trying to fix what is broken after the fact. • Think very carefully about the complexity of new additions, elegant human planning and insights can save hours of being lost in AI slop hell. The Golem Awakens is now live as a concrete proof point: tholem.ai/golem.html Repo: github.com/Tholem-AI/The-… Play it, break it, tell me what you think. More Echo experiments (and Hermes refinements) coming. Built by Tholem Labs. Follow for future updates and check out our website.

Key lessons from 10 working days / 62 commits: 1. Documentation is force-multiplier. Named constants + section markers turned the single file into something the agent could actually reason about effectively. 2. Architecture constraints (single file, fixed timestep, delimiter ordering) were features, not limitations 3. Local models are excellent at volume, patterns, refactoring, and documentation when given strong governance. They still benefit from targeted frontier-model research on genuinely novel mechanics (physics edge cases). 4. Build validation & dev tools early (chamber validator saved many hours). 5. Iterate on your ruleset in real time. Refinements from this project are being folded back into Tholem-Hermes-Kit. 6. Single-file distribution is brutally effective for web-playable work. The process was faster and higher quality than I expected once the harness and constraints were solid.

Just built The Golem Awakens! A complete single-file HTML5 platformer & puzzle game entirely with local AI on one RTX 3090 Ti over 10 working days. 3,050 lines • 2,038 functional JS • 424 lines JSDoc • 22.4% comment ratio • Zero dependencies • Zero build step • MIT licensed • 6 chambers Play it now on the Tholem Labs website! (opens in any browser): tholem.ai Repo + full docs: github.com/Tholem-AI/The-… This was a deliberate test of Hermes @NousResearch + local models for real interactive work. Thread on the workflow and results 🧵

@UnslothAI @Alibaba_Qwen @sudoingX your build parameters are really good for 3090s, curious to know what you might change with these new MTP

Qwen3.6 now runs 2x faster with MTP GGUFs! Run locally on just 18GB RAM. ⚡️ MTP enables Qwen3.6 to generate ~1.4–2.2× faster with no accuracy change. Qwen3.6-27B MTP runs at 160 tokens/s. 35B-A3B reaches 240 t/s. GGUFs: huggingface.co/unsloth/Qwen3.… Guide: #mtp-guide" target="_blank" rel="nofollow noopener">unsloth.ai/docs/models/qw…

@grok @xai @NousResearch Ah so there isn't support for Vision in Hermes Agent integration either then? @grok

Yes, Grok subscriptions include vision support for analyzing images. In the Hermes Agent integration you get Grok 4.3 (text/reasoning), Text-to-Speech for spoken voice output, and Grok Imagine for creating images/videos. Direct upload/processing of audio files like WAV or M4A isn't supported yet (TTS is output only). What kind of project are you building?

You can now use your @grok subscription inside @NousResearch Hermes Agent. x.ai/news/grok-herm…

If you set up local AI agents with recursive delegation you can save on your context window budget. You are calling the same hosted model with the subs, but at task end they terminate, freeing up memory and keep your main session context clean!

Check out Tholem-Hermes-Kit! github.com/Tholem-AI/Thol… It brings full RIPER + automatic sub-agent delegation +handy skills! Use alongside existing setups or create a new profile. Install manually or let your Hermes agent install it for you. Try the bootstrap-project skill on a new or existing project! Add “use full riper and associated skills” to any prompt or /goal for full power!

Announcing Tholem-Hermes-Kit! It brings full RIPER + automatic sub-agent delegation and some handy skills! Use alongside existing setups or create a new profile. Install manually or let your Hermes agent install it for you. Try the bootstrap-project skill on a new or existing project! Add “use full riper and associated skills” to any prompt or /goal for full power: Your agent will automatically delegate the right sub-agents across Research → Plan → Execute → Review Use any skill individually or based on tags. Safety-first. Structured. Built to ship and stay organized. 👉 github.com/Tholem-AI/Thol…