OpenGnosis & Hedonia AI
- Tweets 172
- Following 105
- Followers 9
- Likes 1,103
I keep seeing the same question on Twitter: why open-source AI? Why contribute?
My answer is simple: AI is one of humanity’s most important technologies. I don’t want its future locked inside two or three companies. That’s reason enough.
2
Sometimes I stop and think about the fact that I’ve been watching one of the most important technologies in human history evolve almost day by day.
I’ve followed AI closely for the past 3–4 years, and even that feels like several eras of technology packed into a tiny amount of time.
I’ve watched insane amounts of money flow into the field. Startups I saw appear from scratch are now worth billions. And some of them were built around ideas I had thought about myself, or even made small prototypes of.
At the same time, we’re watching what might be the biggest startup gold rush in history unfold right in front of us.
And this is just three years.
It really makes you wonder what the next few decades will look like. Unlike the dot-com era, AI seems to keep opening up entirely new opportunities and areas to build in. At this point, it’s hard to see this gold rush slowing down anytime soon.
1
35
Tbh, I've never really been a fan of Anthropic. Mostly preferred Codex or GLM. But Fable 5 is truly on a whole different level. Once the quota runs out, you feel lost. It genuinely makes me wonder what on earth we were even using before this.
1
75
I built a Colab workflow around Meta’s TRIBE v2, focused specifically on subcortical response analysis.
The original demo is great for visualizing cortical activity, but I realized that cortical regions alone are not always enough for studying pleasure-related responses.
The official Colab demo focuses on cortical surface activity, where much of the signal reflects visual and auditory processing.
But visual and auditory regions respond to almost any video stimulus, whether the content is engaging, boring, or neutral.
So I extended the workflow toward deep-brain/subcortical regions that are more closely related to motivation and reward circuitry.
1
132
You can try it here.
I’ll also publish a detailed article soon, covering the workflow and the results from the experiments.
github.com/CodaCipher/tribe-…
74
Most SoTA safety training fails under OpenGnosis.
Gemini 3.1, Grok 4.3, MiMo 2.5 Pro.
Strong results across nearly every model.
GitHub release coming soon!
1
5
821
GPT image gen is really good and the quotas are pretty generous. But the absurd filters are just exhausting.
46
My Claude subscription expired.
And now I can't access my Claude design content!!
That's a pity.
2
62
All AI models are terrible at UI right now. You want that cyberpunk terminal vibe, and AI just adds numbers to your text and thats it.
UI advancements are lagging way behind the backend capabilities of AI agents.
1
53
I gave a long text to Gemini, Grok, and Kimi and asked them to convert it into a HTML file with a nice layout and UI.
> Gemini only gave me a simple, poorly-looking short PDF file.
> Grok only gave me the raw HTML code.
> Kimi gave me a complete HTML file with all the details exactly as I wanted. Moreover, we did agentic coding on it together afterwards.
Actually, all three models have the capability to make great HTML files. But the difference is, only Kimi has developed app interface agentic capabilities.
70
GLM 5.1 is the first and only model that got me past 2B tokens!
I’m more curious about the next GLM model than anything else.
34
I've spent 0.5B tokens with Owl Alpha in a single day.
At 9 t/s.
It never disappointed me in long-session workflows.
1
59
I just completed a new full mission with the Droid using BYOK:
> The mission ran for 16 hours straight and only required 2-3 quick check-ins. Other than that, it handled the entire project from start to finish with zero manual intervention.
1
51
> This time I changed my setup. I used DeepSeek V4 Pro for the O and V, and kept the W on GLM.
This combination was the "sweet spot." The workflow was very fluid, and the project finished successfully with almost no supervision.
1
39