# TheTizzrs

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [adam kruger](https://cerebralvalley.ai/u/TheTizz)
- **GitHub:** https://github.com/lightofbaldr/nomos-interp-kernel
- **Demo video:** https://www.awesomescreenshot.com/video/58056891?key=cfba7740b6e1c01785b5e8e872c95b82
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1515

Build for Your World.

For more than 10 years now, I've been a systems engineer. I've worked in security. I've done DevOps. I've been a senior solutions architect for a number of years. And over the past years I've tried to incorporate agentic systems into the diverse fields that make up my background. I currently manage a fleet of agents who help me with research and development, my day job, and managing operations. I literally have an agent that acts as my COO, helping me to dissect and untangle complex ideas, especially surrounding machine learning and artificial intelligence, summarizing news and publications on a daily basis, and helping me build immensely complex systems in remarkably short periods of time.

I know that a lot of people say things like they've 10x'd their productivity, and maybe that's true, maybe it's not. I can say with absolute certainty that my ability to get work done, and to work on a wide range of problems all at once, has grown nearly exponentially over the last year. With that said, it's not perfect. Cloud credits and subscriptions to the various frontier labs can be expensive. And there are a lot of tasks, day-to-day tasks for example, or even experiments, that feel like frontier intelligence is wasted on them.

So for this hackathon, what I wanted to build wasn't another agentic system. I didn't want to build just another tool or product that uses AI instead of logic. Instead, what I wanted to build was a system that helps me understand where the cracks are, the shortcomings in local AI, and of course in this case that's Glimmer 30B, and see if I can't use frontier lab research and techniques to guide my ideas on building a local agentic system that is not only capable of executing and fulfilling tasks in a timely manner and with a high degree of accuracy, but most importantly, a system that is reliable and trustworthy.

For this hackathon, I used my open-source multi-model Mojo kernel, Pi agent, and Google's Harness². Over the past several months, I have dedicated an extraordinary amount of time and effort to building my own from-scratch inference kernel and serving stack. The same week that Glimmer was released, I added support for Glimmer. That project is not meant to be judged. Instead, I made a copy of the kernel and created a new repo the day the hackathon opened. The README lists exactly what was built during the hackathon and what came before it, including three commits that morning before the kickoff, the UI snapshot I started from, and my first changes to Harness² on October 20.

During the days leading up to the hackathon, I tested Glimmer to uncover and categorize the emotional readouts that are used in this project, and to determine under what circumstances Glimmer lies, confabulates, goal switches, or gives up. There would not have been enough time to do the project while also doing the research, so the research was done outside of the hackathon window.

Because I created the kernel in Mojo from scratch, I'm able to serve Glimmer fast while also gathering live interpretability data, as well as actively steering the model when I choose. The way I see it, this project enables me to explore interesting prompts to see what Glimmer's responses might be, while also exploring the global workspaces and emotional readouts for those prompts, recording those results so that I can play them back later. This exploratory step allows me to build a foundation, in order to then get ideas about how particular emotions affect outcomes.

And I think it's important to state here that when I use the word emotions, I am not using the human definition of emotion. I'm simply stating that these words are categorized as emotions. I'm not trying to anthropomorphize in any way, shape, or form. I'm also not trying to say that there isn't something more than simple ones and zeros. I don't know either way. I just want to make that clear. Emotions, in this case, are a dictionary, if you will, of words that are associated with certain parts of the neural network. Those activations, that space, are associated with other concepts in the neural network, and therefore certain words have certain associations, and therefore certain associations have certain outcomes. I think that if we look at frontier lab research, such as Anthropic's, this is pretty clear. A lot of data and time and effort has gone into understanding how a particular type of emotional state influences the model's output.

That said, I don't believe anyone has looked at models that are both capable, and this is important, and run on consumer-grade hardware. I myself have DGX Sparks, a laptop with a 5090, the 24-gigabyte variety, and RTX PRO 4000s. Quantized NVFP4 versions of the 30-billion-parameter class of models work very well for me. In my experience, when I look at people who experiment outside of the frontier labs, and sometimes even the frontier labs themselves, they tend to use very small models because it gives them fast iteration. I believe that is a weakness of Python.

When I test, what I want to know is which capable model, capable being the 30-billion-parameter size, is least likely to lie and cheat and confabulate and goal switch and so on. The system that I am building allows me to do this: to build a profile for any of my supported models, which right now are Qwen 3.8 27B, Glimmer 30B, Gemma 4 31B, and OLMo of the same class and size. Obviously, for this hackathon, I focused on Glimmer. But I fully intend to use this system and continue to build it out, to generate metrics on these open-weights models and add support for other open-weights models, so that the user, in this case me, but others as well, can test how a model reacts, both in its reasoning, in its answers, and most importantly in its internal states, to user input and context.

With that data in hand: giving the model a persona or a system prompt, like Hermes, OpenClaw, and other agentic harnesses do, most certainly impacts the performance and the outcomes of a given model. Again, in my testing, as you will see, Glimmer responds in a way that is counterintuitive, at least to my intuition, when it comes to fear and afraid versus joy and happiness. I think the reason why it's counterintuitive is because my intuition only understands emotion from a human-centric position. With that in mind, I cannot yet say if the same vectors impact the other supported models in the same way.

So once I know how a particular vector or emotion impacts outcomes, I imagine that I can craft system prompts in general, and even personas, that help guide a given model down a path that gives it the best chance at being a helpful assistant and helping the user fulfill the task. I envision a time in the future where my kernel has a hook that takes the user's input, classifies it, and then, based on the classification, nudges the steering vectors for that class of input in the way that we know helps that model get to the goal more often than not. I did not have time to do this for the hackathon, but this is where I am going.

This project uses Pi agent and its new functionality to create durable agents. The code for that is very minimalistic; it's essentially using their API and SDK. But what it allows me to do is come back to that agent even after the session is terminated, and the agent has complete recollection of everything that's been done. No need to add a resume command or continue. It simply exists in that state in time.

And finally, for the purpose of trying to create the most optimized agent for something that I would actually use, I chose to leverage Google's Harness² and to build an adapter for Pi, Mojo, and Glimmer, end to end, fully complete. My intention with this project is to gather data on Glimmer so that it can be used to optimize the kernel, the harness, the skills, and so on, so that Glimmer can become a functional member of my team and write clean, idiomatic Mojo code without me having to babysit it. Mojo is a particular passion of mine. It's not exactly the most popular language in the world, and it's new enough that smaller, non-frontier models struggle with some of its more advanced features, such as leveraging SIMD. Glimmer has a good baseline with Mojo, but after this self-improvement loop, with the same model and the same weights, it went from passing one held-out Mojo task in ten to eight in ten. Over five tries at each task, it went from 7 of 50 to 45 of 50. Claude improves the harness, the compiler and each task's tests grade the code, and Meta's Muse Wonder, through Meta's Model API, judges its quality. Muse Wonder also powers the lab's Interpreter, which reads a turn's signals back in plain language, right beside the evidence.

Three minutes isn't enough to show everything, so please read the README and browse the deck. The code, the README and the 15-slide deck are at github.com/lightofbaldr/nomos-interp-kernel.

## More from Meta Global AI Developer Hackathon

- [Krishi Mitra](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1512)
- [Sufian](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1513)
- [Gruya](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1514)
- [BioComputational](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1516)
- [BrainHiFi](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1517)
- [ProtocolEye](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1518)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1515. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt