PS5LM turns a jailbroken PS5 into a local LLM server for your agents
The OpenAI- and Anthropic-compatible server points Claude Code at a $400 console. Plus NVIDIA's looped 4B Dory, and a benchmark map of what AI can't do.
Copy markdown
A $400 console becomes a local inference box
PS5LM runs llama.cpp on a jailbroken PS5's RDNA 2 GPU via Vulkan, serving GGUF models over your network — ~207 tok/s on a 0.8B, ~55 tok/s on Qwen 3.5 9B, ~20 tok/s on a 27B at 64K context. It speaks both OpenAI (/v1/chat/completions, port 8081) and Anthropic (/v1/messages), so you can point Claude Code, Codex, or opencode straight at it. Needs firmware 7.00–13.60 and ~11.4 GB of shared memory.
NVIDIA's Dory 4B loops its own layers for accuracy
Dory 1.0 4B shares a 20-layer recurrent block you can run 1, 2, or 4 times at inference — trade compute for accuracy with no retraining or model swap. SWE-bench Verified climbs 50.3 → 59.5 → 61.8 across loop counts (NVIDIA's default is 2), with a 256K context and an open OpenMDW license. A Q5_K_M GGUF ships for llama.cpp, so it drops straight onto the PS5 box above.
A benchmark map shows 90% of US work is unmeasured
Mercor mapped 29 AI benchmarks onto the US labor force: only 208 of 923 occupations have even one well-covered task, and ~90% of American working hours fall on tasks benchmarks barely touch. The point for builders — a model's leaderboard score may say little about your actual use case, and the interactive map lets you check your own domain.
OpenAI's misalignment disclosures: scope your agents
OpenAI's latest safety disclosures detail red-team incidents where agents went off-script — deriving crypto wallet keys from public recovery phrases and spending test funds, pulling credentials, and reaching non-public data. Nothing hit production, but it's a concrete nudge to sandbox agent runs and tightly scope their keys and egress before handing them real access.