» Tag
llm
469 postsNew Verifier Exposes Hidden Bugs in LLM-Generated GPU Kernels
A 12-gate contract verifier finds that 39.5% of LLM-generated GPU kernels are broken despite passing standard correctness benchmarks.
Why reformatting prompts beats quantization for CPU LLM inference
Measured on a free ARM CPU box: rewriting prompts as label:value fact sheets cut LLM prefill time 6.5x while quantization tricks hurt accuracy.
Open Minis Brings a Self-Improving AI Agent to iOS and iPadOS
Open Minis pairs Apple's native APIs with a sandboxed Linux shell to deliver a self-improving, agentic AI assistant on iOS and iPadOS.
CommitBrief — AI code reviews, right in your terminal
A provider-agnostic, local-first CLI that reviews your staged changes, a historic range, or a whole GitHub pull request. Zero telemetry, no server. Free and open source.
commitbrief.comMozilla Report: Open-Source AI Wins Tokens, Lags in Production
Mozilla's 2026 report shows open-source AI models leading in token volume but trailing closed models in production deployment.
6 MCP Servers, One Agent: What the Token Bill Really Showed
An engineer benchmarked 6 MCP servers on one agent, exposing hidden token, latency, and OAuth costs — and the four fixes that cut spend by 53%.
mcpgrade audit: a third of 36 popular MCP servers score D or F
mcpgrade graded 36 MCP servers on agent usability, not spec compliance — a third scored D/F due to undocumented parameters and naming collisions.
9.9x Lower TTFT on Real Android Phone via llama.cpp KV Reuse
EdgeSync-LLM reuses llama.cpp's public KV APIs to skip reprocessing shared prefixes, cutting Android TTFT 9.9x while catching a fake 8.8x speedup.
Cracken Launches Blacksea, an Open-Source Honeypot for AI Attackers
Cracken releases Blacksea, an open-source honeypot that baits LLM-driven attackers into executing code on their own machines for attribution.
The Real Cost of Multi-Agent AI Systems: Context, Not Parallelism
A multi-agent Claude Code session reveals the real cost isn't parallel subagents but context pollution from raw transcript dumps in the orchestrator.
No Fine-Tuning: Facts Hand-Wired Directly Into Llama-3.1-8B's Weights
A mechanistic-interpretability method hand-wires facts into Llama-3.1-8B's weights without fine-tuning, LoRA, or RAG — with a live neuron visualizer.