by Adam
The Hidden State Attack: Why Your LLM's System Prompt Isn't Secret
Responsible disclosure of a class of vulnerabilities that allow system prompt extraction from transformer hidden states
security
llm
transformers
vulnerability
Read more
[ Blog ]
Homelab adventures, infrastructure deep-dives, and lessons learned building enterprise-grade systems on a budget
by Adam
SipIt: Extracting System Prompts from Transformer Hidden States
How I achieved 100% token recovery from Mistral-7B hidden states and what it means for AI security
transformers
security
interpretability
sipit
Read more
by Adam
Self-Improving Models Without Labels: What I Just Proved and Why It Matters
A 7B model taught itself to generate better security commands using only its own understanding signals. No human labels, no external reward. Here's how and why it matters.
ai
machine-learning
self-improvement
llm
Read more
by Adam
Self-Improving AI: Using Inversion Fidelity as a Training Signal
A novel discovery: hidden state inversion quality predicts model capability, enabling self-improving systems without external feedback
ai-training
self-improvement
transformers
fidelity
Read more
by Adam
Mojo: The Python Killer That Might Actually Have a Shot
A deep dive into Mojo, the MLIR-based language promising 35,000x speedups over Python. What it does well, where it falls short, and who should actually care.
programming
python
mojo
performance
Read more
by Adam
Implementing Pi-Attention: 50% Memory Reduction for Long-Context LLMs
How periodic sparse attention achieves O(n) complexity while maintaining model quality
transformers
attention
memory-optimization
mojo
Read more