Microsoft Research unveils CARE-X, which calculates cardiothoracic ratio directly from chest X-rays
Beyond generating reports, a system that measures heart and chest width to compute CTR
120
AI
AI models, services and robotics
Jailbreak prompt built for Google Gemma also works on DeepSeek V4 Flash
Reddit user says a Gemma 4 jailbreak system prompt worked as-is on DeepSeek, without even changing the model name
1620
Oumi Unveils LLM Platform That Self-Trains on Production Logs
By cycling through evaluation, data synthesis, and retraining, it runs specialized models up to 100x cheaper than frontier models
20
OpenAI reviews free year of ChatGPT Plus for US college students
Signs of 'back-to-school' campaign spotted, positioned as exam prep and homework tool
160
Luma unveils 'Luma Scenes,' letting creators approve shots before final render
Instead of generating ten times to get one usable result, the tool lets users lock in scenes and timing before rendering
70
Solar Pro 4 Free for One Week on Nous Portal
Upstage's agentic model for multi-step work on long documents is free for one week on Nous Portal, built to ask follow-ups instead of making things up
230
Fields Medalist Grapples With Identity Crisis as AI Conquers Mathematics
Math community stirs as OpenAI delivers solutions to 10 long-unsolved problems
50
AA-AnalystAgent Benchmark: Claude Opus 5 Leads at 54%
Artificial Analysis tests AI on real-world spreadsheets and documents. Claude Opus 5 scored 54%; 57% of failures anchored on a wrong early hypothesis.
30
ChatGPT Desktop App Arrives as Linux Preview
OpenAI has released a preview of its ChatGPT and Codex desktop apps for Ubuntu, Debian, and Fedora
30
ChatGPT Work and Codex to auto-sync agent work
OpenAI adds an import feature to its desktop app that pulls projects, conversations, skills, and plugins into one place
100
xAI unveils Grok Bot agent beta, also works in Cursor
Available to Grok Heavy, Cursor Ultra and Team Premium users, bots run real tasks on their own computers
290
Google adds video and voice consultation capabilities to medical AI 'AMIE'
AMIE, previously limited to text conversation, expands to real-time video consultations, demonstrating expert-level performance in simulated consultations
140
NVIDIA releases open model 'Nemotron 3.5 Lightning,' a 30B model that runs on just 3B active parameters
The MoE architecture activates only 3B of its total 30B parameters per token, and NVIDIA claims up to 4x faster output than comparable models
10
Databricks unveils enterprise document reasoning benchmark 'OfficeQA Pro V2'
A benchmark that measures the ability to answer using in-house documents as evidence, with questions scaled up using synthetic data
10
Unsloth releases desktop app with local training support
MLX, GGUF, diffusion, and audio in one app — connects Claude Code and Codex to local models
890
XPeng Autonomous Driving Chief Gives Exclusive VLA 2.0 Interview
Head of the General Intelligence Center answers directly on autonomous driving's limits and Europe entry timing
20
Existing Token Benchmarks Cannot Rank Coding-Agent Languages
Claim that dynamic languages are more token-efficient turns out to rest on unverified benchmarks, critic says
10
llama.cpp adds CI build target for AMD ROCm 7.14
First official release built on TheRock build system expands Linux and Windows support
170
OpenAI unveils GPT-5.6-Cyber, a model dedicated to cybersecurity
Daybreak program split into Blue and Red tiers, with access restricted to approved defenders
30
New training method lets robot arms adjust speed like a dial
Apple unveils MoMo, a robot manipulation framework that controls whether the same motion is performed quickly or cautiously
20
Apple Proposes Reading Data Through UMAP's Internal Graph
Apple analyzes UMAP's neighbor graph directly, bypassing the usual visualization, to surface representative points and dense structures
00
GitHub Releases Copilot SDK for Java, Cutting Vendor Lock-in
New SDK lets developers control AI agents directly from Java servers using virtual threads and annotations
10
Ai2 Measures the "Over-Helpfulness" Problem in AI Tutors
In an evaluation built from real tutoring transcripts, LLMs struggled to tell when to help a student and when to hold back
80
Anthropic mandates watermarking on Claude outputs worldwide
Under EU AI Act rules, invisible markers for text and signed provenance data for images start August 2026
1240