AI — Page 38

The YC W25 company says the benchmark uses 102 tasks drawn from hedge fund analyst reasoning traces.

The Tokyo AI company says the internal research group will focus on models that can help write, test, and improve AI systems.

The former OpenAI CTO previewed models for continuous audio, text and video, but left release timing and customer traction unstated.

Patel says ScaleDown's small language models beat GPT-5.4 Mini on cost, speed and accuracy, but the benchmarks are self-reported.

A 20VC interview frames Mercor as spending more on AI agent tokens than salaries, but the exact cost comparison remains unclear.

Jane Manchun Wong said the original prompt was simply "Improve ./core.rs" after a screenshot showed DeepSeek V4 changing Rust functions about Tiananmen Square and Taiwan.

The benchmark uses Arena.ai's own session data, including 160,000 tasks and 2.06 million tool calls over one week.

The open model is pitched as a playable instrument that takes MIDI, text and audio inputs, but Google has not detailed its specs in the thread.

TechCrunch and Poke describe the approval as Apple's first for an AI agent on the business messaging platform.

DeepSeek V4 Flash wins 34.0 to 17.0 by being usable, complete, and more faithful to the prompts. MiMo-V2.5 repeatedly looked polished while dropping facts, inventing details, or failing outright.

Founded by two former Formula One engineers and an AI researcher, microagi is testing whether chores can become robotics training data.

The benchmark group says the 550B-parameter model scores 47.7 on its Intelligence Index, with weights posted on Hugging Face.

Appendix lets users have an AI agent draft a medical encounter, then routes it to a human physician for review and a prescription if warranted.

TechCrunch reports the multi-year deal expands Lovable's Google Cloud footprint fivefold and gives it broader access to Claude and Gemini.

The preview is pitched as a video upgrade, with xAI claiming better motion, scene coherence, native audio and longer clips.

Claude Sonnet 4.6 wins 35.0 to 26.5 by being more reliable where correctness actually bites. DeepSeek V4 Flash had the cleaner customer email, but it fell down on harder structured and coding work.

The image-generation startup's docs list endpoints for text-to-image, image editing and reference-image remixing, with direct image responses, postprocessing and credit headers.

The comparison found different strengths across storytelling, product design, brand systems, and photorealistic physics.

The 9.3B-parameter Ideogram 4 model was trained from scratch and adds a structured JSON prompting interface for text, layout, color and 2K image control.

The planned bill would tax large AI companies in stock, turning a redistribution idea into a fight over startup control.

The claim spans world generation, robot action policy, and industrial vision understanding, but the post did not include scores or test details.

The public site and ACM paper frame ViBench as an end-to-end test of whether coding agents can deliver usable apps, not just pass SWE-style tasks.

Google says the Apache 2.0 model handles vision and native audio without separate encoders, runs with 16GB of VRAM or unified memory, and nears its 26B MoE model on benchmarks.

Aravind Srinivas told CNBC the winner will maximize economic output from AI energy use, with orchestration across cloud and on-device models as the lever.