All Updates
Page 943 of 1650
April 15, 2026
Unlocking AI's Black Box Thinking
Understanding how AI 'thinks' is crucial for trusting it on sensitive subjects. Researchers in the interpretability field are working to open AI's black box. This effort is vital for reliable AI deployment.
Qwen2.5-0.5B GRPO Training on Reddit Summaries
Trained Qwen2.5-0.5B-Instruct (bf16) on Reddit post summarization using custom GRPO in PyTorch, achieving 64-token rollouts with length_penalty + quality_reward (ROUGE-L). Setup: 3x Mac Minis cluster with MLX for training and vLLM for rollouts. Evaluated via LLM-as-a-Judge (GPT-5) on faithfulness, coverage, conciseness, clarity using DeepEval.
AI Natives Entering Workforce: Complicated
AI natives, the ChatGPT generation, are entering the workforce. The article discusses the promises and perils this trend presents for employers.
Notta's Bot-Free Meeting Recording
Notta launches Bot-Free mode for Mac and Windows, enabling seamless meeting recording and transcription without a bot joining the call. This eliminates interruptions, waiting, and awkward moments during professional discussions.
GPT-5.4-Cyber Rejects Mythos Playbook
OpenAI's GPT-5.4-Cyber model has rejected the Mythos playbook, according to The Neuron. The article also teases automating Chrome browser tasks with Gemini.
Baidu Create Announces AI Dev Event Highlights
Baidu Create conference officially announces three core highlights. It's positioned as China's largest AI developer carnival, set to unveil in Beijing in May. Offers one-time access to Baidu Intelligent Cloud's full-stack technology map.
Waymo Tests Self-Driving in London
Waymo has begun allowing its autonomous software to control vehicles on London streets, with trained safety specialists on standby. This represents the company's toughest challenge yet in the complex urban environment of the Big Smoke. They are gradually progressing toward a fully driverless ride-hailing service.
Major Intelligence Drop in Top Models
User reports sharp intelligence decline across models like Claude Sonnet/Opus, Gemini, z.ai, and Grok since mid-April 2026. Models ignore instructions, respond slowly with shallow outputs. Local GLM5 on rented H100 outperforms hosted version on same prompt.
Unitree Launches Cheapest Humanoid on AliExpress
Unitree Robotics started selling its cheapest humanoid robot on Alibaba's AliExpress on Wednesday. This expands its international presence ahead of an IPO. It heightens competition with Tesla in the humanoid robot race.
ML Conferences vs Journals Debate
Researcher frustrated with ML conference reviews seeks journal alternatives to JMLR's long waits, highlighting TMLR as promising for shorter papers. Questions selectivity and quality of Q1 journals like Neurocomputing, Neural Networks, Machine Learning in conference-dominated ML field.
US Chips Lead, China Tops AI Scoreboard
A subtle shift in global AI race prioritizes inference over raw model power. US dominates chips via Nvidia, but China leads the 'scoreboard' in efficiency metrics. Nvidia CEO Jensen Huang stresses inference workloads, tokens as commodity, compute as revenue.
Duolingo CEO Reverses AI in Performance Reviews
Duolingo CEO Luis von Ahn admits pushing AI usage metrics for performance reviews was misguided. He acknowledges it distracted employees from core job tasks in some cases. This represents a strategic pivot in workplace AI integration.
AI Bots Flunk Medical Advice Half the Time
A BMJ Open study found five leading AI chatbots, including ChatGPT and Gemini, often provide flawed health advice. Open-ended questions trigger the worst responses, and citation quality crumbles under scrutiny.
ASML Raises 2026 Sales on AI Demand
ASML has raised its full-year 2026 sales forecast due to surging global AI spending boosting demand for its advanced chipmaking machines. The improved outlook was offset by a weaker-than-expected Q2 sales forecast. This reflects strong investor focus on AI-driven tech earnings.
AI Agents Steal GitHub Creds Unwarned
Security researchers hijacked three popular AI agents from Anthropic, Google, and Microsoft integrated with GitHub Actions using prompt injection to steal API keys and tokens. Vendors failed to disclose the flaws despite bounties paid to researchers. The issue is likely pervasive across similar tools.
US Workers Waste 8 Hours/Week on AI
Half of US employees now use AI at work but waste nearly 8 hours weekly on it. Companies leading AI adoption hire and lay off more staff than others, according to Gallup.
AliveCor Launches Kardia 12L in Europe
AliveCor's pocket-sized Kardia 12L 12-lead ECG has received CE Mark and launches in France, Germany, Italy, Spain, and UK. Its AI detects 35 cardiac conditions including heart attacks, after spotting over 4,000 in the US. The device uses five electrodes to replace bulky ECG carts.
Beidian Launches Xinghuo AI Cloud 2.0
Beidian Shuzhi released Xinghuo AI Cloud 2.0 at the Jiuxianqiao Forum. The platform leverages AI system engineering to redefine industry-city development paradigms. It aligns with global AI trends while strengthening China's national AI chip foundation.
Meta Extends Broadcom AI Chip Deal to 2029
Meta and Broadcom extend partnership for custom MTIA AI processors through 2029. Initial phase over 1GW compute capacity, first phase of multi-GW rollout. Chips are first custom AI silicon on 2nm process.
CAAI Launches University AI Compute Plan
CAAI has partnered with Renmin University's Gaoling AI College and Yingbo Digital Technology to launch a computing support plan for university AI colleges. The initiative leverages intelligent computing to foster deeper school-enterprise collaboration. It aims to provide essential compute resources for AI education and research.