Erica Windisch
@ewindisch.bsky.social
building hyprstream. agentic infrastructure for applications that learn.
https://github.com/hyprstream/hyprstream
https://github.com/hyprstream/hyprstream
Pinned
Erica Windisch
@ewindisch.bsky.social
· Sep 27
We're building an AI that's free, public, and openly trained by the community, with open weights shared via torrent git chunks.
Streaming weekdays 11-3pm on Twitch.
Come hang out Monday, or give our repos some stars! ⭐️
Streaming weekdays 11-3pm on Twitch.
Come hang out Monday, or give our repos some stars! ⭐️
I trained an LLM on the Epstein files today by asking an MCP server to do it for me.
Working out some bugs and polish, but I'm on track to replace Hyprstream v0.3.0 this week.
Working out some bugs and polish, but I'm on track to replace Hyprstream v0.3.0 this week.
February 9, 2026 at 9:10 PM
I trained an LLM on the Epstein files today by asking an MCP server to do it for me.
Working out some bugs and polish, but I'm on track to replace Hyprstream v0.3.0 this week.
Working out some bugs and polish, but I'm on track to replace Hyprstream v0.3.0 this week.
Everyone planning boycotts against Discord as if the issue wasn't lawmakers and voters, because of anti sex-worker and "save the children" legislation.
Multiple platforms are doing this because their lawyers are telling them to, and providers are selling them solutions.
Multiple platforms are doing this because their lawyers are telling them to, and providers are selling them solutions.
Discord will soon require face scans or ID for all users, or restrict access
Discord will soon require face scans or ID for all users, or restrict access
Discord has announced a major change coming to the service as part of its aim to improve teen safety features. Starting in March, all users globally will be required to submit age verification via a face scan or submitted ID, or be given teen-level access restrictions.
more…
9to5mac.com
February 9, 2026 at 6:12 PM
Everyone planning boycotts against Discord as if the issue wasn't lawmakers and voters, because of anti sex-worker and "save the children" legislation.
Multiple platforms are doing this because their lawyers are telling them to, and providers are selling them solutions.
Multiple platforms are doing this because their lawyers are telling them to, and providers are selling them solutions.
I feel like there's some mistakes or errors in Theo's universe project. Maybe it's the 2D projection, but there's huge divides between tech circles, social circles, people who span multiple domains... but maybe it's accurate and reasonable. IDK
Placement is weird.
Placement is weird.
February 9, 2026 at 5:49 PM
I feel like there's some mistakes or errors in Theo's universe project. Maybe it's the 2D projection, but there's huge divides between tech circles, social circles, people who span multiple domains... but maybe it's accurate and reasonable. IDK
Placement is weird.
Placement is weird.
LLMs are not human, do not ascribe them human characteristics, they are raccoons with lightsabers.
Very powerful, and very chaotic.
They're getting better.
Very powerful, and very chaotic.
They're getting better.
February 9, 2026 at 5:42 PM
LLMs are not human, do not ascribe them human characteristics, they are raccoons with lightsabers.
Very powerful, and very chaotic.
They're getting better.
Very powerful, and very chaotic.
They're getting better.
just another day training an LLM on the Epstein files (does not emulate or impersonate, this model is being trained to be knowledgeable about facts)
February 9, 2026 at 4:02 PM
just another day training an LLM on the Epstein files (does not emulate or impersonate, this model is being trained to be knowledgeable about facts)
February 9, 2026 at 6:13 AM
Train your model via MCP
February 9, 2026 at 1:45 AM
Train your model via MCP
casually fine tuning models with my mcp server
February 9, 2026 at 12:43 AM
casually fine tuning models with my mcp server
It took a year from planning POSIX to laying down POSIX in my project, even with Claude in the mix.
February 8, 2026 at 1:06 AM
It took a year from planning POSIX to laying down POSIX in my project, even with Claude in the mix.
Thinking about this more:
I get more out of newer models going slow than I do going fast, but partly this is due to the quality of the tools and models.
Sometimes you gotta go fast, sometimes you gotta slow down.
I get more out of newer models going slow than I do going fast, but partly this is due to the quality of the tools and models.
Sometimes you gotta go fast, sometimes you gotta slow down.
I haven't hit limits since August. I'm not sure why, if there's something special about my account, of if I'm no longer the power user I used to be.
I do use fewer MCP tools and go slower with fewer sessions currently, as I'm working on an increasingly complex application.
I do use fewer MCP tools and go slower with fewer sessions currently, as I'm working on an increasingly complex application.
February 7, 2026 at 10:10 PM
Thinking about this more:
I get more out of newer models going slow than I do going fast, but partly this is due to the quality of the tools and models.
Sometimes you gotta go fast, sometimes you gotta slow down.
I get more out of newer models going slow than I do going fast, but partly this is due to the quality of the tools and models.
Sometimes you gotta go fast, sometimes you gotta slow down.
If you're afraid of fully autonomous online learning drones, may I introduce you to birds?
February 7, 2026 at 4:04 AM
If you're afraid of fully autonomous online learning drones, may I introduce you to birds?
I can't believe I've been using Claude since April
February 7, 2026 at 1:47 AM
I can't believe I've been using Claude since April
I finally got the mcp branch of hyprstream working.
Model tool use is coming soon, too.
The last piece of the puzzle will be MoE support.
Things are about to blast off
Model tool use is coming soon, too.
The last piece of the puzzle will be MoE support.
Things are about to blast off
a group of cartoon characters with letters r on their shirts
ALT: a group of cartoon characters with letters r on their shirts
media.tenor.com
February 6, 2026 at 10:33 PM
I finally got the mcp branch of hyprstream working.
Model tool use is coming soon, too.
The last piece of the puzzle will be MoE support.
Things are about to blast off
Model tool use is coming soon, too.
The last piece of the puzzle will be MoE support.
Things are about to blast off
Reposted by Erica Windisch
TIL that the maximum GitHub commit size is 2GB. Don’t ask why or how I learned that.
February 6, 2026 at 5:37 PM
TIL that the maximum GitHub commit size is 2GB. Don’t ask why or how I learned that.
generate code with code generator
use code generator to write code generator
use code generator generator to make code generator generator generator
use code generator to write code generator
use code generator generator to make code generator generator generator
February 6, 2026 at 2:58 PM
generate code with code generator
use code generator to write code generator
use code generator generator to make code generator generator generator
use code generator to write code generator
use code generator generator to make code generator generator generator
opus generating code to generate code
February 6, 2026 at 4:50 AM
opus generating code to generate code
I had stopped using Opus 4.5 because it didn't do a good job for me. I had switched to Sonnet and GLM. Giving Opus 4.6 a try and it's okay so far... still not ready to give a thumbs up or down, but it's not terrible.
February 5, 2026 at 11:59 PM
I had stopped using Opus 4.5 because it didn't do a good job for me. I had switched to Sonnet and GLM. Giving Opus 4.6 a try and it's okay so far... still not ready to give a thumbs up or down, but it's not terrible.
You can't do AI safety without first doing AI danger.
February 5, 2026 at 8:32 PM
You can't do AI safety without first doing AI danger.
Who needs JSON when you have WASM?
February 5, 2026 at 3:50 PM
Who needs JSON when you have WASM?
put the serde in the wasm
February 5, 2026 at 3:05 PM
put the serde in the wasm
Released Hyprstream v0.2.0 today.
This release includes model downloads, Git worktree management, basic LLM inference and training for Qwen3 dense models.
Released as AppImages for CPU, CUDA, ROCm, and Universal.
direct links to follow...
github.com/hyprstream/h...
This release includes model downloads, Git worktree management, basic LLM inference and training for Qwen3 dense models.
Released as AppImages for CPU, CUDA, ROCm, and Universal.
direct links to follow...
github.com/hyprstream/h...
Release Hyprstream v0.2.0 · hyprstream/hyprstream
CUDA NOTE: Make sure the system includes the package cuda-toolkit. After installation, make sure to execute the following to fix a bug with cuda dependencies: systemctl --user set-environment LD_PR...
github.com
February 4, 2026 at 11:44 PM
Released Hyprstream v0.2.0 today.
This release includes model downloads, Git worktree management, basic LLM inference and training for Qwen3 dense models.
Released as AppImages for CPU, CUDA, ROCm, and Universal.
direct links to follow...
github.com/hyprstream/h...
This release includes model downloads, Git worktree management, basic LLM inference and training for Qwen3 dense models.
Released as AppImages for CPU, CUDA, ROCm, and Universal.
direct links to follow...
github.com/hyprstream/h...
Looking to grab if a DGX Spark if I can find someone in need of a lightly-used RTX5090 (got it in a pre-build this July, tested it, and idled a few months)
February 4, 2026 at 9:55 PM
Looking to grab if a DGX Spark if I can find someone in need of a lightly-used RTX5090 (got it in a pre-build this July, tested it, and idled a few months)
Seems the latest Epstein drops briefly mention OpenStack and Docker. 💀
February 4, 2026 at 7:55 PM
Seems the latest Epstein drops briefly mention OpenStack and Docker. 💀
I would like for the ROCm on my Strix Halo to not crash when I run inference, thank you.
Everything works great on my AMD Instinct mi210 GPUs...
It might be a PBKAC but Segmentation fault (core dumped)
Everything works great on my AMD Instinct mi210 GPUs...
It might be a PBKAC but Segmentation fault (core dumped)
February 4, 2026 at 6:18 PM
I would like for the ROCm on my Strix Halo to not crash when I run inference, thank you.
Everything works great on my AMD Instinct mi210 GPUs...
It might be a PBKAC but Segmentation fault (core dumped)
Everything works great on my AMD Instinct mi210 GPUs...
It might be a PBKAC but Segmentation fault (core dumped)
Looks like at least one of my machines might have to become dedicated to running Qwen3-Coder-Next. Wow.
Might be time to finally give a fully local Opencode a shot.
Might be time to finally give a fully local Opencode a shot.
February 4, 2026 at 2:36 PM
Looks like at least one of my machines might have to become dedicated to running Qwen3-Coder-Next. Wow.
Might be time to finally give a fully local Opencode a shot.
Might be time to finally give a fully local Opencode a shot.