I don't understand how politicians can so bluntly reject the will of the people like this. https://t.co/2Jt8iPTRRN

I didn't realize Grok imagine was *this* good holy... 😳

Another day, another horror story from the open borders nightmare. This Afghan "migrant" grabs a teenage girl off the street and rapes her just ONE WEEK after assaulting a woman in her 20s. And guess what? He's still in the UK because spineless politicians prioritize invaders over their daughters. He’s never getting sent back because the courts won’t allow it anyway.

@PawelHuryn Isn’t that counterintuitive? The more a system adapts to you, the less general it becomes. At some point it stops being a platform and becomes a very elaborate config file. It’s just overfitting. Now with agents.

@PawelHuryn The AI community keeps confusing prompt tweaks with learning. It’s not. Editing a prompt isn’t exponential intelligence or AGI. It’s just adding another bullet point to the instructions…

@johncrickett You can ignore 99% of everything you see online. It works 100% of the time.

Qodo outperforms Claude Code Review by 19% higher recall and costs 10x less per review. Why this matters for AI devs: @QodoAI just published a comparison using their open benchmark: 100 real pull requests, 580 injected issues, 8 production repositories across TypeScript, Python, JavaScript, C, C#, Rust, and Swift. All three tools hit 79% precision, meaning the quality of individual findings is comparable. The gap comes from recall. Claude Code Review sits at 52%, Qodo Default at 60%, and Qodo Extended at 71%. Great to see attention to recall which matters a lot in this use case. Precision can be tuned post-processing, but recall reflects how deeply a system understands the codebase and reasons about cross-file interactions. If an issue is never flagged, no amount of filtering helps. Even more impressive is how the specialized agent catches the remaining ground truth. Qodo Extended deploys a multi-agent system where specialized agents focus on different issue categories (logical errors, best practices, edge cases, cross-file dependencies), then merge outputs through verification and deduplication. It also leverages model diversity across OpenAI, Anthropic, and Google rather than being locked to a single provider. On cost, Qodo reports significantly lower per-review pricing compared to Claude's token-based model at $15-$25 per review.

This AutoHarness paper (from Google DeepMind) is the most interesting thing I've read lately. I am testing a similar idea (without training) on models like MiniMax-2.5 and getting good results. It already allowed me to synthesize an entire functional coding agent. More soon. https://t.co/pGRRlIEsUR

A U.N.-backed panel of independent experts says racist hate speech by Trump and other American political leaders, along with a crackdown on immigration in the United States, have led to “grave human rights violations.” https://t.co/w8eGAj3Ra6

if you're a CS grad or SWE and have zero knowledge about ML/AI, this is the best book to start (3rd edition is the best). I've been recommending for many friends and they also say that. You can read online @fchollet rocking as always https://t.co/HlezJodiqh

Trust Your Critic FIRM introduces robust reward models for RL-based image editing and generation with 370K editing and 293K generation samples, specialized 8B critics, and the FIRM-Bench benchmark. https://t.co/jeQB6hNXtO

Trust Your Critic FIRM introduces robust reward models for RL-based image editing and generation with 370K editing and 293K generation samples, specialized 8B critics, and the FIRM-Bench benchmark. https://t.co/jeQB6hNXtO

Spatial-TTT Tencent Hunyuan and Tsinghua researchers introduce test-time training for streaming spatial intelligence, using adaptive fast weights to accumulate 3D evidence from unbounded video with near-linear memory scaling. https://t.co/Kr4NGubNwh

Spatial-TTT Tencent Hunyuan and Tsinghua researchers introduce test-time training for streaming spatial intelligence, using adaptive fast weights to accumulate 3D evidence from unbounded video with near-linear memory scaling. https://t.co/Kr4NGubNwh

AI is forcing the debate about energy and infrastructure. As hyperscale data centers expand, critics argue they are pushing up electricity demand and prices. Tech companies say they will absorb the costs or invest in alternative energy, but the long-term economics remain unclear. In the end, the real AI race may not be about models or chips, but about who can secure and finance the massive energy needed to run them. https://t.co/cxIXvAfDl1 @CNBC

AI toys are entering children’s lives, but researchers warn they can misread emotions and respond inappropriately. When machines try to interpret feelings, small errors can quickly turn into confusing or harmful interactions, especially for young users who may trust the technology too easily. As AI moves into childhood spaces, the real question is not just what these toys can do, but whether they truly understand the emotions they claim to read. https://t.co/TIEQbNuy3h @bbcnews @zsk @Caldernator

@AIWorkflowGuide I have a cognitive architecture. You don't. It is a super power.

Optimized Parquet is a first class citizen on the NEW 🤗 Storage Buckets ! Use pandas/arrow to read & write faster to HF using Parquet files optimized for Xet Never redownload/reupload duplicate data chunks ever again 👉 https://t.co/iam0sRryce https://t.co/F7Ht9NYPJ3

How to make physics unforgettable ⚡ When learning feels like play, curiosity takes over. A simple experiment like static electricity can turn a classroom into a moment of wonder. And that’s the real lesson. The future of education isn’t about memorizing formulas — it’s about making people feel the science. Because when curiosity is sparked, learning happens naturally. Sometimes the best innovations start with a little bit of fun. #science #education #innovation #curiosity #learning

🤯 56 researchers from 32 universities across US, China, UK built an enormous video reasoning dataset to prove current AI models struggle with basic physical logic.   "Very Big Video Reasoning Suite" The problem is that the AI does not genuinely know how solid objects are supposed to behave. So Berkeley, Stanford, CMU, Harvard, Oxford, Columbia, NTU, Johns Hopkins, and 24 other institutions built this 2mn samples which makes it 1000 times larger than all existing collections combined.   Video generation systems usually focus on making things look pretty but they completely fail to understand spatial rules and causality.   The team created a massive factory of visual tasks that tests how well models handle navigation, object manipulation, and logic.   Even the most advanced commercial systems only scored around 54% while human testers easily achieved over 97% accuracy.   Training an open model on this specific data improved its reasoning skills but a massive gap still exists.

This is pure and elegant filmmaking... A triple rack focus + crane down + dolly shot Completely invisible as story information is shared Takes planning + preparation + blocking + practice JURASSIC PARK (1993) DP: Dean Cundey Director: Steven Spielberg https://t.co/6LUxf85aQ8