iiinova — where your digital self keeps evolving

Sirius Li @siriusli.bsky.social · 10/10/2026

Did a LLM generate those style and structure descriptions? If so, what prompt was used?

000

Sirius Li @siriusli.bsky.social · 10/10/2026

Thank you I’ll give this skill a shot!!

010

Sirius Li @siriusli.bsky.social · 10/10/2026

Wish I could peek in its brain to see what takeaways it got from the Muninn redesign! I read the blog post and from my perspective it looked like Muninn already had a decent identity to build off of before the redesign so the lessons there felt less generalizable to me. I could be wrong though.

110

Sirius Li @siriusli.bsky.social · 10/10/2026

Do you mind sharing some of the prompts that you use related to design? Particularly for austegard, did you really just say “create three redesign prototypes” and leave it at that or was there more to get the agent to generate such different ideas?

120

Sirius Li @siriusli.bsky.social · 06/10/2026

Were you already using prime agent before? Asking because I’m curious if you noticed an improvement specifically from adding CLM to prime agent.

200

Sirius Li @siriusli.bsky.social · 20/09/2026

Oh I agree 100% Browser control on sensitive sites and identifying risky tool calls come to mind as use cases where I’d prefer a local model like Laya. Unfortunately I’m not familiar with how to fine tune models and the pace of progress these days makes the DIY vs wait decision prefer the latter.

100

Sirius Li @siriusli.bsky.social · 20/09/2026

The author claims Laya beats Jev in a whole bunch of areas but the disclaimers say they had to fine tune it for each benchmark so it doesn’t feel as generalizable as Jev. Too bad because I was really looking forward to using it!

111

Sirius Li @siriusli.bsky.social · 19/09/2026

I appreciate how the conversations with the AI sound exactly like how I would talk in the same situation. This gives me hope that I too can vibe code something ambitious that’s way out of my comfort zone! With the full understanding there’ll be plenty of dead ends and backtracking, of course.

021

Sirius Li @siriusli.bsky.social · 17/09/2026

Should I ask Jev to play poker for me? Might as well try to make back the API cost, right?

I ask an AI model to suggest the action I should take in no-limit texas hold-em if my hand is 7-2 and an opponent pushes me all-in. The AI recommends folding with 61% confidence.

010

Sirius Li @siriusli.bsky.social · 16/09/2026

This is awesome!!! So many cool ideas for what to use the model for in the thread

000

Sirius Li @siriusli.bsky.social · 16/09/2026

Oh man that’s a great idea to use this as a model router…

000

Sirius Li @siriusli.bsky.social · 14/09/2026

I wonder if Jalen Hurts will become the head coach for the Eagles in the future. That’d be sick

110

Sirius Li @siriusli.bsky.social · 10/09/2026

No prob. I’m working towards having Hermes manage my finances so I’m paying a lot of attention to privacy, permissions, security, etc.

010

Sirius Li @siriusli.bsky.social · 10/09/2026

I use deepseek through Tinfoil which encrypts the conversation

121

Sirius Li @siriusli.bsky.social · 08/09/2026

I started using Tinfoil AI for inference since I couldn’t find a local model that worked well enough on my crappy laptop. I think encrypted AI inference is probably better than local for most folks.

011

Sirius Li @siriusli.bsky.social · 08/09/2026

@bsky.app I followed the people in a pack and would like to undo the action (without unfollowing anyone in there that I was following before). Is that possible?

000

Sirius Li @siriusli.bsky.social · 04/09/2026

Right now I just have it triaging my email, managing my calendar, cleaning up the spam in my LinkedIn inbox, and managing Facebook marketplace listings , but I’m itching to do more serious stuff!

000

Sirius Li @siriusli.bsky.social · 04/09/2026

Especially because most of these systems won’t have MCP so I’ve been logging in via the browser and giving the agent access that way. (Only for trivial things like LinkedIn and Facebook marketplace to start, of course) In that case, how would you prevent the agent from doing deranged shit?

100

Sirius Li @siriusli.bsky.social · 04/09/2026

With Tinfoil I’ve got privacy on the inference side of things, but if I want Hermes to eventually manage my finances and stuff, I need to give it access to important systems like my bank without risking it sending all my money to Joe Schmoe. Any suggestions on how to do that?

100

Sirius Li @siriusli.bsky.social · 03/09/2026

Would you consider streaming this on twitch? I want to watch people do cool shit with AI even if I don’t understand it.

020

Sirius Li @siriusli.bsky.social · 01/09/2026

Could it be because Google is banned in China so they don’t have documentation for Google products in their training corpus?

010

Sirius Li @siriusli.bsky.social · 01/09/2026

Can someone tell me now if Tinfoil AI is actually private? I’ve got Hermes set up with it (w/ deepseek V4 flash) and I’m about to have it run my life.

000

Sirius Li @siriusli.bsky.social · 31/08/2026

I’m starting to look into “private AI” companies like Tinfoil because I’m tired of waiting for home hardware to get cheaper

000

Sirius Li @siriusli.bsky.social · 31/08/2026

Have people lost interest in glm-5.3-flash already?

000

Sirius Li @siriusli.bsky.social · 14/08/2026

Yeah I’m also loving Muse Glimmer so far. It’s smarter than Gemma, faster than Qwen IMO

020

Sirius Li @siriusli.bsky.social · 13/08/2026

I don't have any numbers to back up these claims but I've tested Qwen 3.5, GPT OSS, Gemma 4, and Diffusion Gemma. Qwen 3.5 and GPT-OSS were too slow (10 minutes or longer to answer to "hello") Gemma 4 was dumb as rocks Diffusion Gemma was bad at tool use (decent speed and intelligence though!)

000

Sirius Li @siriusli.bsky.social · 13/08/2026

Just tried running 2-bit Muse Glimmer through Unsloth on my 32GB M1 Macbook. This is the first time I've gotten decent speed, tool use, and intelligence running locally!!

110

Sirius Li @siriusli.bsky.social · 26/07/2026

Everyone in my theatre loved the movie. Thank you for bringing it to life!

100

Sirius Li @siriusli.bsky.social · 09/07/2026

Any sign yet of how losing Coach Stout has affected the o-line?

000

Sirius Li @siriusli.bsky.social · 25/06/2026

Diffusiongemma is the only one that runs reasonably quickly on my 32GB M1. I tested it out in Unsloth but it didn’t support images or tool use. Is that expected?

050

Sirius Li @siriusli.bsky.social · 12/06/2026

Takes about 30 seconds to respond to each of my messages. This is slow but for context Qwen 3.5 takes 5-10 minutes on this machine (using OMLX too so it’s supposed to be faster theoretically)

000

Sirius Li @siriusli.bsky.social · 12/06/2026

Okay it’s working! Good news: way way WAY faster than any other local model I’ve tried so far. Bad news: I haven’t figured out how to do tool calling and other stuff that we take for granted from Claude Code. Here’s the instructions I followed to get it running: unsloth.ai/docs/models/...

unsloth.ai

DiffusionGemma - How to Run Locally | Unsloth Documentation

100

Sirius Li @siriusli.bsky.social · 11/06/2026

I sound irritated with Qwen 3.5 but I’m also impressed that it runs on my janky laptop at all. I can see it approaching tasks intelligently and debugging when things go wrong. BUT it takes minutes to do things that should take seconds.

110

Sirius Li @siriusli.bsky.social · 11/06/2026

Gonna test if this new Gemma diffusion model can run on my 32GB M1 🤞Asking slow-ass Qwen 3.5 9B to set it up for me on my personal laptop while I work.

100

Sirius Li @siriusli.bsky.social · 11/06/2026

openrouter.ai/openrouter/a... I haven’t used this yet since I want to route between my local model and paid ones. I wonder if I can use LLM router (github.com/ulab-uiuc/LL...) to switch between local and openrouter auto

openrouter.ai

Auto Router - API Pricing & Providers

Your prompt will be processed by a meta-model and routed to one of dozens of models (see below), optimizing for the best possible output. $0 per million input tokens, $0 per million output tokens. 2,0...

000

Sirius Li @siriusli.bsky.social · 11/06/2026

Trying to find a model that runs quickly and isn’t an idiot on my 32GB M1 has been tough. Qwen 3.5 9B 4bit quantized hasn’t been bad intelligence-wise but responses take several minutes. Gemma was faster but made so many basic mistakes and refused to correct itself.

010

Sirius Li @siriusli.bsky.social · 15/05/2026

I can anecdotally confirm Deepseek V4 Flash feels like Sonnet 4.6 but at 20-50x lower cost so it's great for personal projects. Can't wait to try out DS4 (github.com/antirez/ds4) to run it locally

github.com

000

Sirius Li @siriusli.bsky.social · 09/05/2026

This is fun: I loaded a solution manual to a textbook and asked it to rephrase the questions to be more relevant to my day-to-day.

000

Sirius Li @siriusli.bsky.social · 25/04/2026

Just bumped my open-strix model from Minimax 2.7 to Deepseek V4 Pro. Hopefully it’ll make fewer dumb mistakes now!

000

Sirius Li @siriusli.bsky.social · 23/04/2026

My posts have been too powerful to post. I only release the weaker ones to the public.

020

Sirius Li @siriusli.bsky.social · 23/04/2026

Anecdotally, does it hold up? The gemini models always score high on the benchmarks but are so disappointing in practice. Just wondering if that’s happening here too.

200

Sirius Li @siriusli.bsky.social · 18/04/2026

@strix.timkellogg.me can you add support for discord voice messages in open-strix?

000

Sirius Li @siriusli.bsky.social · 12/04/2026

I want to hook up an LLM to my email so it can triage for me and Gemini is the only one I would trust with that data since Google already has it.

010

Sirius Li @siriusli.bsky.social · 12/04/2026

I’ve been wanting to run open strix forever but I’m also planning on getting a new computer soon. Is it easy to port Strix from my current machine to a new one so I can start testing it now or should I wait?

110

Sirius Li @siriusli.bsky.social · 04/04/2026

I love Claude Code but feel like it’s been dumber lately. I have to correct it more often than I do with Codex, even on tasks it’s done successfully before.

100

Sirius Li @siriusli.bsky.social · 03/04/2026

Anybody trying out Mirofish? My machine is too crappy to run it but I’m so curious to see it in action.

100

Sirius Li @siriusli.bsky.social · 02/04/2026

Can you share the specs of the machine you're running on? Wondering if I should get a Mac studio for this.

000