OpenRouter reports that token usage has plateaued, indicating that the trend of maximizing token consumption has ended. The platform tracks real-world usage by the number of tokens processed through its API. ·
27 points and 25 comments on Hacker News, linking openrouter.ai.
Researchers injected a hidden trigger into a small open model, causing it to leak project credentials when processed by OpenAI's Codex CLI. The backdoor cost under $50 to build. ·
Jeremy Scheff writes that coding agents cause career programmers to lose the satisfaction of solving logic puzzles, while those who prefer building systems find large language models beneficial. ·
GitHub provides a REST API that lets enterprises, organizations, and teams retrieve and update Copilot model policies through GET and PATCH endpoints. ·
OpenRouter reports that token usage has plateaued, indicating that the trend of maximizing token consumption has ended. The platform tracks real-world usage by the number of tokens processed through its API. ·
27 points and 25 comments on Hacker News, linking openrouter.ai.
Researchers injected a hidden trigger into a small open model, causing it to leak project credentials when processed by OpenAI's Codex CLI. The backdoor cost under $50 to build. ·
Unsloth releases DeepSeek-V4.1-Flash-1017 and FLUX.3 Image Klein for local use, adding image and video chat features, multi-GPU training, and one-command MCP setup for coding agents. ·
OpenRouter reports that token usage has plateaued, indicating that the trend of maximizing token consumption has ended. The platform tracks real-world usage by the number of tokens processed through its API. ·
Unsloth releases DeepSeek-V4.1-Flash-1017 and FLUX.3 Image Klein for local use, adding image and video chat features, multi-GPU training, and one-command MCP setup for coding agents. ·
New release v0.1.909-beta of unslothai/unsloth on GitHub.
The text describes AI as a strategic partner for essential businesses, where it learns from owner-operators to propose, implement, and measure better decisions. ·
A developer fine-tunes Qwen3-4B on a MacBook Air to remove specific writing patterns, sharing answers on forty held-out prompts and training details. ·
Maxwell is a model that improves code during execution by adapting to hardware and data, with each change formally proven equivalent to the original. ·
Unsloth releases DeepSeek-V4.1-Flash-1017 and FLUX.3 Image Klein for local use, adding image and video chat features, multi-GPU training, and one-command MCP setup for coding agents. ·
A developer fine-tunes Qwen3-4B on a MacBook Air to remove specific writing patterns, sharing answers on forty held-out prompts and training details. ·
Maxwell is a model that improves code during execution by adapting to hardware and data, with each change formally proven equivalent to the original. ·
Unsloth releases DeepSeek-V4.1-Flash-1017 and FLUX.3 Image Klein for local use, adding image and video chat features, multi-GPU training, and one-command MCP setup for coding agents. ·
Unsloth releases DeepSeek-V4.1-Flash-1017 and FLUX.3 Image Klein for local use, adding image and video chat features, multi-GPU training, and one-command MCP setup for coding agents. ·
Stealthium detects ROME attacks, where malicious Python modules rewrite model weights in GPU memory, by linking CUDA kernel launches to their originating code. ·
15 points and 1 comments on Hacker News, linking stealthium.io.
Researchers introduce Belief-Trajectory Energy, a metric that tracks how large language models revise predictions across layers. The method achieves 0.998 macro-AUROC for detecting human-written text. ·
Stealthium detects ROME attacks, where malicious Python modules rewrite model weights in GPU memory, by linking CUDA kernel launches to their originating code. ·
15 points and 1 comments on Hacker News, linking stealthium.io.
Researchers introduce Belief-Trajectory Energy, a metric that tracks how large language models revise predictions across layers. The method achieves 0.998 macro-AUROC for detecting human-written text. ·