Harvard Study: LLMs Outperform Emergency Room Physicians in Diagnostic Accuracy

A new study from Harvard University is drawing attention for putting large language models (LLMs) to the test in real medical settings. In tests using actual emergency room cases, at least one model arrived at more accurate diagnoses than two human physicians.

The study evaluated LLMs across multiple medical contexts, and what matters is that it was grounded in real clinical data rather than a mere benchmark. While the result that "the AI was more accurate" points to its potential as a support tool in emergency medicine, much more validation is believed to be needed before practical deployment.

Academy Awards Exclude AI from Acting and Screenplay Categories

The Academy of Motion Picture Arts and Sciences, which organizes the Academy Awards, has announced new rules that make AI-generated content ineligible for the acting and screenplay awards.

Specifically, AI-generated performances (voice and video) and screenplays will be excluded from eligibility in these categories. As debates over AI use in the film industry intensify, Hollywood's highest authority has drawn a clear line. That said, technical categories such as VFX and editing will not ban auxiliary AI use outright; different criteria are expected to apply to each category.

Musk Testifies in OpenAI Trial That "AI Will Surpass Humans Next Year"

In the OpenAI-related trial, where the conflict between Elon Musk and Sam Altman has been taken to court, Musk took the stand. He stated that "AI smarter than humans will emerge next year," suggesting that the arrival of AGI is imminent.

The trial centers on OpenAI's transition from a nonprofit to a for-profit entity, with Musk, one of the company's co-founders, arguing that it has strayed from its original mission. The outcome could have a major impact on governance models across the AI industry.

Sulphur 2: An Uncensored Video Generation Model Based on LTX 2.3 Debuts

"Sulphur 2" is gaining attention on Hugging Face's trending list. It is a 9-billion-parameter video generation model based on LTX 2.3, natively supporting both text-to-video (t2v) and image-to-video (i2v).

Development builds in fp8mixed and bf16 have been released, and a distill LoRA is also available. Positioned as an uncensored model, it can be seen as further expanding the freedom of open-source video generation.

Gemma 4 E2B Runs Practically on 8GB Android, Powering a Private Voice Memo App

A user in Reddit's LocalLLaMA community shared a report on running Gemma 4 E2B (2.4GB) on an Android phone with 8GB of RAM and building a fully local voice memo app.

The setup uses Whisper Small (244MB) for speech recognition and Gemma 4 E2B for text segmentation and categorization, completing end-to-end processing of a 10-15 second voice memo in 12-15 seconds. Everything runs without any cloud, and the quality of the JSON output was reportedly better than expected. It's a serious build, too — query expansion and reranking at search time are also handled by Gemma.

Local LLM Inference Speed Has Improved 100x in Two Years

Also on Reddit, a post looking back at the evolution of local LLM inference speed is making waves. Two years ago, running the Q4 quantization of Llama 405B locally topped out at 1.2 tokens per second; today, the same hardware can run giant models (DeepSeek V4 Flash, Qwen 3.5-397B, and the like) at 30-100 tokens per second.

"I was preparing so I could run it myself even when AGI arrives," the poster reflected, praising the current state — where Qwen 3.6-36B can be run at home at 50tk/sec for a few hundred dollars — as "better than I ever imagined."

---