DeepSeek shipped an agent runtime where even the loop is a plugin
DeepSeek published deepseek-harness, an MIT-licensed agent runtime in which models, tools, skills, sessions, sandboxes, storage, scheduling, the interface, and the agent loop itself are all plugins that can be swapped from configuration.
Qwen3.8-27B's compression data has a blind spot where its speed-up head lives
The importance matrix used to compress Qwen3.8-27B contains no entries for block 64, the model's multi-token-prediction head, because that head never activates during the standard calibration run, which is where the most aggressive one-bit builds start to break.
DeepSeek gave its cheap model eyes, then capped them at 384 tokens an image
DeepSeek released deepseek-v4-flash-vision-exp, an experimental multimodal version of its cheapest model that accepts images directly in the same agent loop as text, but budgets each image to at most 384 tokens after resizing toward roughly 800 by 800 pixels.
Anthropic widened access to its cyber model by removing the prompt box
Anthropic made Claude Mythos 5, its most capable cybersecurity model, available to Enterprise customers through the Claude Security product, where users receive scan findings, severity ratings, and suggested patches rather than direct access to the model itself.
OpenAI's Mac app will log your workday, and warns that raises injection risk
OpenAI shipped Computer History for the ChatGPT desktop app on macOS, an opt-in feature that turns clicks, typing, and app context into a searchable timeline ChatGPT and Codex can reference, and its own documentation warns the feature increases the risk of prompt injection.
A free million-token model appeared with no owner and two conflicting privacy policies
Ox Alpha, a free anonymous model on OpenRouter with a 1,048,576-token context window, is described as zero-retention in one set of documentation and as retaining prompts and completions in another, while an independent token-level analysis points to Zhipu's GLM line as the likely provider.
The ARC-AGI-3 record going around is the wrong number and the wrong system
The top ARC-AGI-3 entry on ARC Prize's public leaderboard is an NVIDIA-labelled agent scoring 85.1% on the public demo set at a cost of $332, self-reported and not independently verified, and it is not the AVO system that viral posts credited with a perfect run.
Students using generative AI got better homework grades and worse exam scores
A 30-month study of 26,811 Chinese secondary students found that after adopting generative AI, homework scores rose about 18% and homework time fell about 30%, while closed-book monthly exam scores fell about 20%, with losses concentrated among students whose homework time dropped sharply.
Pew finds one in ten English web pages shows signs of AI authorship
Pew Research sampled 490,000 English-language web pages from Common Crawl across 49 crawls between January 2021 and July 2026 and found 10% of the July 2026 sample showed significant signs of AI authorship, rising to 35% among pages that carried a publication date after ChatGPT's release.
Google made the visible Gemini watermark optional and kept the invisible one
Google added a setting in Gemini Apps that turns off the visible watermark on generated images, videos, and music, while its help page states the setting does not affect SynthID or Content Credentials, the invisible provenance markers embedded in the same files.
Three papers landed the same day arguing you should build the world, not the model
EnvHarness, FACET, and SPADE all took the top spots on Hugging Face's daily paper list with the same underlying move, shifting effort from making the agent smarter to manufacturing the environments the agent practices in, with FACET releasing 6,020 ready-made terminal tasks.
Moderna's personalized cancer vaccine cleared Phase 3 with a learned selector inside it
Moderna reported positive Phase 3 results for intismeran autogene combined with pembrolizumab in patients with completely resected advanced melanoma, a treatment built individually for each patient by an algorithm that picks up to 34 targets from that person's own tumor sequencing.