AI — Page 20

Karl Kahn says Claude Max 20x delivered six to eight times Pro usage, not the 20x Anthropic advertised.

AI coding agents have turned GitHub reliability into an infrastructure problem Azure cannot absorb alone on Microsoft's timetable.

Teknium says the delegate tool can now fan out work without freezing the chat, a practical change for long-running agent workflows.

One model mostly gestures at the prompts; the other actually stages them. This matchup isn’t close: Seedance 2 Image to Video wins by turning specific shot language into coherent motion instead of settling for attractive approximation.

Karan Goel is using benchmark wins to pitch Cartesia as both the speaking and listening layer for real-time AI agents.

CrankGPT runs speech recognition, a small language model, and text-to-speech locally on a Raspberry Pi 5 with no battery or cloud.

Claude Opus 4.8 sweeps three of four tasks with sharper regex engineering, more polished prose, and cleaner structured output—Kimi K2.7 Code only manages a tie on the JSON normalization task.

One model wins on the jobs that punish sloppiness: typography, layout discipline, and prompt-specific product detail. The other lands a moodier single-image hit, but not enough to overcome repeated misses where accuracy actually matters.

A 73-year-old developer said he mentioned feeling unwell during a coding task, and Claude Code kept urging immediate care before doctors treated a sudden AFib episode.

Z Lab, Modal and LMSYS released a DFlash drafter for Qwen's 397B model and benchmarked it above native MTP on 8x B200 GPUs.

The company pulled Fable 5 and Mythos 5 after a June 12 export-control order, then sent technical staff to Washington to repair the relationship.

Zohar Alon's new identity-security startup is betting enterprises will need to govern agents like workers, not service accounts.

One model delivers attractive motion clips; the other actually follows the brief shot by shot. In both tests, Marey Realism V1.5 separates itself by turning prompt details into believable action instead of decorative near-misses.

This matchup wasn’t close once the prompts demanded precise scene construction rather than just attractive images. AuraFlow can look polished, but Luma Uni-1 Edit was the model that actually followed the brief across all three tests.

The updated model card says a base merge of Nex-N2-Pro and Qwen was uploaded by mistake, shifting the dispute from pure attribution to release discipline.

Alexandr Wang's Muse Spark gives Meta a proprietary model; the harder job is proving it can become more than ad infrastructure.

A June preprint claims Pearl's GPU network is doing random matrix math, not verified AI work, challenging Omri Weinstein's core bet.

Yichuan Wang and collaborators show a screenshot-first retrieval system beating text pipelines, with lower agent token use and a real chunking gap.

Przemek Chojecki's 14-problem smoke run puts Moonshot's new open-weight model behind Claude Fable-5-max and ahead of GPT-5.5 xhigh.

The AI security startup says its agent found 21 FFmpeg zero-days for about $1,000, including an RCE exploit primitive.

One model actually stages the prompt; the other mostly vibes around it. Across both tests, Luma Ray 3.2 Image to Video is the only system that consistently delivers the requested setting, motion, and subject choreography.

Matei Zaharia and Kasey Uhlenhuth's alpha project sits above Claude Code, Codex, the Pi agent and custom agents to compose multi-agent workflows, share live sessions and enforce contextual policies.

IplanRIO put Rio 3.5 Open 397B on Hugging Face with MIT licensing, a 1M-token context claim, and self-reported gains over Qwen's base model.

Meta's 6,500-person Applied AI unit looks less like automation magic than the unglamorous human operating system required to make models useful.