Erica Windisch
@ewindisch.bsky.social
building hyprstream. machine intelligence with zero consensus - peer to peer AI training and inference.
https://github.com/hyprstream/hyprstream
https://github.com/hyprstream/hyprstream
Let's not forget Margolies, who has the trauma of that era. Her son is married to Chelsea.
November 10, 2025 at 3:08 PM
Let's not forget Margolies, who has the trauma of that era. Her son is married to Chelsea.
Been experimenting with putting a browser in WASM connected via a tor snowflake proxy.
November 10, 2025 at 2:12 AM
Been experimenting with putting a browser in WASM connected via a tor snowflake proxy.
One of the underlying aspects of the talk I've been working on / giving.
November 9, 2025 at 9:57 PM
One of the underlying aspects of the talk I've been working on / giving.
Good time to share this monster of shell scripting
github.com/openstack/de...
github.com/openstack/de...
GitHub - openstack/devstack: System for quickly installing an OpenStack cloud from upstream git for testing and development. Mirror of code maintained at opendev.org.
System for quickly installing an OpenStack cloud from upstream git for testing and development. Mirror of code maintained at opendev.org. - openstack/devstack
github.com
November 8, 2025 at 8:18 PM
Good time to share this monster of shell scripting
github.com/openstack/de...
github.com/openstack/de...
One of my favorite exercises for testing models is to request a custom version of Brainf**k using Emoji. It's challenging enough that it's useful for catching edge cases in kvcache/RoPE/Yarn.
November 8, 2025 at 8:01 PM
One of my favorite exercises for testing models is to request a custom version of Brainf**k using Emoji. It's challenging enough that it's useful for catching edge cases in kvcache/RoPE/Yarn.
Its called reward hacking. You incentivize the AI toward a desired behavior using a reward or threat. Gaslighting the AI is also an effective prompting technique.
November 7, 2025 at 3:19 PM
Its called reward hacking. You incentivize the AI toward a desired behavior using a reward or threat. Gaslighting the AI is also an effective prompting technique.
MCP servers can help with this along with using a constraints oriented design/programming model.
November 7, 2025 at 2:28 PM
MCP servers can help with this along with using a constraints oriented design/programming model.
Here's the idea: a browser running inside WASM rendering to a canvas controlled by a parent browser. WebRTC to the TOR Snowflake network.
A single page web application that's a relatively safe tor browser.
Not an easy problem 😅
A single page web application that's a relatively safe tor browser.
Not an easy problem 😅
November 6, 2025 at 3:46 PM
Here's the idea: a browser running inside WASM rendering to a canvas controlled by a parent browser. WebRTC to the TOR Snowflake network.
A single page web application that's a relatively safe tor browser.
Not an easy problem 😅
A single page web application that's a relatively safe tor browser.
Not an easy problem 😅
Exfiltration is a far likelier scenario. My AI dev environments have no SSH keys for outbound, but instead write to network storage. I push/pull to git from another system.
A bad claude session could potentially escalate permissions simply by opening a socket or writing to a local filesystem.
A bad claude session could potentially escalate permissions simply by opening a socket or writing to a local filesystem.
November 6, 2025 at 1:59 PM
Exfiltration is a far likelier scenario. My AI dev environments have no SSH keys for outbound, but instead write to network storage. I push/pull to git from another system.
A bad claude session could potentially escalate permissions simply by opening a socket or writing to a local filesystem.
A bad claude session could potentially escalate permissions simply by opening a socket or writing to a local filesystem.
If you let it "write" to the network, then it can exfiltrate information, or open a reverse connection, opening a door to attack.
Obviously that's not an out of the box permission set, but typical permissions granted by tools like Claude.
If your AI has CLI, assume full compromise.
Obviously that's not an out of the box permission set, but typical permissions granted by tools like Claude.
If your AI has CLI, assume full compromise.
November 6, 2025 at 1:50 PM
If you let it "write" to the network, then it can exfiltrate information, or open a reverse connection, opening a door to attack.
Obviously that's not an out of the box permission set, but typical permissions granted by tools like Claude.
If your AI has CLI, assume full compromise.
Obviously that's not an out of the box permission set, but typical permissions granted by tools like Claude.
If your AI has CLI, assume full compromise.
In that case you've got to disable network access too
November 6, 2025 at 1:03 PM
In that case you've got to disable network access too
I'm dealing with this (or similar) in my implementation right now 😬
November 4, 2025 at 9:02 PM
I'm dealing with this (or similar) in my implementation right now 😬
Even if I'm wrong, your threat model should be to assume that I'm right.
November 4, 2025 at 3:13 PM
Even if I'm wrong, your threat model should be to assume that I'm right.
I apparently didn't scream loudly enough before Spectre/Meltdown, but I'll do it now:
The best time to switch to post quantum cryptography is YESTERDAY, but the next best time is today.
Your lifetime of internet history is about to open up, but at least protect yourself going forward.
The best time to switch to post quantum cryptography is YESTERDAY, but the next best time is today.
Your lifetime of internet history is about to open up, but at least protect yourself going forward.
November 4, 2025 at 3:09 PM
I apparently didn't scream loudly enough before Spectre/Meltdown, but I'll do it now:
The best time to switch to post quantum cryptography is YESTERDAY, but the next best time is today.
Your lifetime of internet history is about to open up, but at least protect yourself going forward.
The best time to switch to post quantum cryptography is YESTERDAY, but the next best time is today.
Your lifetime of internet history is about to open up, but at least protect yourself going forward.
I spent months digging into this. We're at a moment where algorithmic improvements need fewer qubits, making complexity trade offs between compute and storage. Meanwhile, quantum hardware is readying to scale out.
Classical systems might even suffice soon for breaking high value targets.
Classical systems might even suffice soon for breaking high value targets.
November 4, 2025 at 2:59 PM
I spent months digging into this. We're at a moment where algorithmic improvements need fewer qubits, making complexity trade offs between compute and storage. Meanwhile, quantum hardware is readying to scale out.
Classical systems might even suffice soon for breaking high value targets.
Classical systems might even suffice soon for breaking high value targets.
GPUs might fire sale anyway. An analyst recently told me it's like dark fiber. Incredibly useful technology, too much was purchased too soon, and sent to the wrong places. We have more GPUs than electricity.
November 4, 2025 at 12:32 PM
GPUs might fire sale anyway. An analyst recently told me it's like dark fiber. Incredibly useful technology, too much was purchased too soon, and sent to the wrong places. We have more GPUs than electricity.
I've had no problems passing my ROCm devices from hypervisor, to guest, to container?
November 2, 2025 at 5:36 PM
I've had no problems passing my ROCm devices from hypervisor, to guest, to container?
we need more on device and embedded AI. that requires creating C compatible libraries. we either write in C/C++ or in Rust.
python and javascript are a terrible languages for cross platform high performance systems software. go is not easily linkable.
python and javascript are a terrible languages for cross platform high performance systems software. go is not easily linkable.
November 2, 2025 at 4:19 PM
we need more on device and embedded AI. that requires creating C compatible libraries. we either write in C/C++ or in Rust.
python and javascript are a terrible languages for cross platform high performance systems software. go is not easily linkable.
python and javascript are a terrible languages for cross platform high performance systems software. go is not easily linkable.
no food no housing just more flops and memory
November 2, 2025 at 4:16 PM
no food no housing just more flops and memory
At the high end a 9070XT or a used RTX3090. That's the $600~700 price point.
Just don't go under 12GB VRAM.
Just don't go under 12GB VRAM.
November 2, 2025 at 2:57 PM
At the high end a 9070XT or a used RTX3090. That's the $600~700 price point.
Just don't go under 12GB VRAM.
Just don't go under 12GB VRAM.
Align yourselfs against 1) corporations or 2) align yourself against mathematics and the workers who apply it. 🤔
Even Bernie recently said we need AI for and of the workers.
Those who rage against thermodynamics will lose.
Even Bernie recently said we need AI for and of the workers.
Those who rage against thermodynamics will lose.
November 1, 2025 at 9:37 PM
Align yourselfs against 1) corporations or 2) align yourself against mathematics and the workers who apply it. 🤔
Even Bernie recently said we need AI for and of the workers.
Those who rage against thermodynamics will lose.
Even Bernie recently said we need AI for and of the workers.
Those who rage against thermodynamics will lose.