AI news in context

Exclusive: OpenAI Losses Increased Nearly 8X in 2025, With Spending Hitting $34 Billion. Community angle: discussion is focused on practical implications and community skepticism

Anthropic has been sued for allegedly misleading customers on usage limits. A proposed class-action lawsuit was filed today in the U.S. District Court for the Northern District of California against Anthropic, alleging the company misled customers about the usage allowances on its premium Claude subscription plans.Key details of the complaint: * Plaintiff Karl Kahn (Washington, D.C.) upgraded to the Max 20x plan ($200/month) for intensive coding work. * The lawsuit claims the Max 5x ($100/month) and Max 20x plans were marketed as providing 5x and 20x the usage of the standard Claude Pro plan (\~$20/month). * In practice, Kahn reported that a single 5-hour coding session consumed appro. Community angle: discussion is centered on benchmark quality, developer workflow, and cost and pricing

Sony AI’s Ace robot defeats pro player Miyu under official ITTF rules (Nature paper). Ace is an autonomous table tennis robot that made history by defeating elite and professional human athletes. Built for extreme precision and rapid reaction times, it proves that **physical AI** can handle complex, real-time sports environments. **The Psychological Advantage:** Human opponents noted the robot's biggest edge wasn't just speed, it was psychological. Zero panic, zero fatigue and completely flawless consistency during high-intensity rallies. Nature Paper Ace Vs Kahara YT vid **Source:** Nature/Sony AI. Community angle: discussion is centered on labor impact

Sen. Kelly amendment in defense bill would ensure ‘ultimate human responsibility’ in AI-powered kill chain. Community angle: discussion is centered on safety and policy

Anthropic Accused of Misleading Users Over Soaring AI Costs in New Lawsuit. Community angle: discussion is centered on cost and pricing and labor impact

A year after Meta tapped Alexandr Wang to build a new AI model, Zuckerberg has to sell it. Community angle: discussion is centered on privacy and data use and infrastructure costs

Be wary of Qwen/Claude distillations - they're often worse than the base model. Just to be clear; I am not attempting to call anybody out or be mean to those who take the time/money to make these models, I just want to inform people about these distills/finetunes since there's clearly some confusion going on. I'm going to assume those of us who often visit this subreddit have noticed these models, particularly the "Qwopus" model and the such, though I'm sure there's probably Gemma 4/Claude distills too. As I type this, there's currently a Qwen 3.6 based Claude Fable 5 distillation model on the frontpage. Seems pretty cool, right? Yep. Up until you actually look into how these models were. Community angle: discussion is centered on benchmark quality

How do Karpathy and other AI power users avoid turning their codebases into AI slop?. I do all of my development work using Claude code. The problem is that I run Claude code in auto mode, and I don’t have time to review every single code change. As I keep adding code, the feature starts to drift and eventually turns into slop. How do AI experts avoid this issue? People like Andrew Karpathy and the founder of Claude Code claim that they let agents run in loops, but I don’t understand how they avoid this problem. Please don’t give me vague advice like “set up guardrails,” “use hooks,” or “write better specs.” I already do all of that. I use OpenSpec to create my specs, and I spend at least an ho. Community angle: discussion is centered on developer workflow, labor impact, and infrastruct...

Class of AI Models Hyped as Scarily Powerful Apparently Scared the Government Too Much and Now They’re Disabled. Community angle: discussion is centered on local and open-source alternatives and reliability

Subquadratic AI introduces SubQ-1.1-Small, a new model using Smart Sparse Attention. : \- Near-perfect long-context retrieval up to 12M tokens on the needle-in-a-haystack test, with up to nearly 1,000x attention compute reduction. \- A balance of long-context optimization and general reasoning ability, with strong performance retained across knowledge, coding, and non-coding enterprise agent benchmarks. \- At 1M tokens, SubQ 1.1 Small requires 64.5x less compute than dense attention and runs 56x faster than FlashAttention-2. (Independently verified). Community angle: discussion is centered on benchmark quality and cost and pricing