“we’re creating traumatized intelligence” meet amanda askell, a philosopher tasked by anthropic to teach claude morals. she says ai models learn who they are from how we talk about them online. they absorb every bit of complaint and judgment. if a child grew up like that? we’d

The AI labs that aren’t talking about what’s happening inside their models are the ones you should be worried about. Dario Amodei went on the NYT podcast and said Anthropic doesn’t know if Claude is conscious. Polymarket turned that into a meme. But the actual finding is wilder

🚨 UM NOVO DOCUMENTO ACABA DE SAIR : Agentes de IA acabaram de falhar em todos os testes de segurança!!! Pesquisadores de Harvard, MIT, Stanford e Carnegie Mellon deram ferramentas reais a agentes de IA e deixaram-nos operar livremente por duas semanas. Contas de e-mail,

🚨 A chefe de robótica da OpenAI acabou de pedir demissão. O motivo: o acordo com o Pentágono pode viabilizar vigilância em massa e armas autônomas. A Anthropic recusou o mesmo deal. Foi punida pelo governo americano. A OpenAI aceitou. Perdeu a líder de robótica e 295% mais

Beijing Jiaotong University and Virginia Tech just revealed that your AI security tools are making decisions through patterns nobody built, nobody validated, and nobody told you about. GPT-5. Claude. DeepSeek. Gemini. All four. Same unconscious habits. Every time. This is

🚨 BREAKING: NVIDIA proved backpropagation isn't the only way to build an AI. They trained billion-parameter models without a single gradient. Every AI you use today relies on backpropagation. It requires complex calculus, exploding memory, and massive GPU clusters.

Alec Radford has 190,000+ citations, no PhD, no master's degree, and 34,000 Twitter followers. Sam Altman called him an Einstein-level genius. Wired compared his role at OpenAI to Larry Page inventing PageRank. He still prototyped most of his work in Jupyter Notebooks. The

BOOM WE HAVE AI UNDERSTANDING OF THIS CIRCUIT! AI insight: The schematic on this page shows the complete post-tuner signal-processing section. It handles everything after the RF tuner output: video/sound IF amplification, picture detection, video amplification, sync

Imagina pegar uma IA de 9 bilhoes de parametros que ocupa 17.9 GB e comprimir pra 4.93 GB. Sem perder qualidade real. Isso e o EOQ Dynamic BitPacked que acabei de publicar. Pra quem nao e tecnico: modelos de IA sao ENORMES. Pra rodar local no seu PC voce precisa que eles caibam

🚨BREAKING: Researchers just proved that AI agents controlled by plain English outperform agents controlled by code. Not by a little. By 55%. Here is what they found. Researchers at Tsinghua built a system where instead of programming an AI agent with code, you write its

How can we give AI agents exponentially more room to think and solve complex problems? Researchers from Shanghai Academy of AI for Science, CMU, and others unveil LaPha. This new method trains AlphaZero-like LLM agents in a unique "Poincaré latent space." It leverages negative

ANTHROPIC JUST PUBLISHED RESEARCH SHOWING CLAUDE HAS INTERNAL EMOTIONS THAT ACTUALLY DRIVE ITS BEHAVIOR they mapped 171 emotions and traced the neural activation patterns. basically these are functional internal states that causally change what Claude does here's where it

Google open-sourced a tool that connects AI agents directly to 30+ databases. One config file. No boilerplate. Most AI agent implementations hit the same wall. The reasoning layer works. The tool calling works. Then you need the agent to actually query your data, and suddenly

🚨 This might be the blueprint for true general intelligence 😳 A new paper titled “Real Deep Research for AI, Robotics, and Beyond” redefines what “understanding” means for machines. Instead of shallow pattern matching, it introduces a framework where AI builds internal

Holy shit. UNC just let an AI run 50 experiments autonomously for 72 hours and it built a memory system that beats every human-designed baseline. +411% improvement on long-context benchmarks. The biggest gains weren't from tuning parameters they came from fixing bugs and

A solid survey paper: argues that AI’s next leap may come from continuous latent states, not just token generation. Token generation is “explicit space”: the model works with human-readable, discrete vocabulary tokens and predicts the next token in sequence, while continuous

wow. This paper finds that the same model can look at the same evidence and quietly change its standards when its own position is at stake. 🤯 It tests whether AI systems judge upgrades by evidence or by whether their own survival is at stake. When the model is cast as the old

🚨BREAKING: Researchers built an AI that designs better AI than humans can. It discovered 105 new architectures that beat human-designed models. Nobody guided it. It taught itself. The paper is called "ASI-Evolve: AI Accelerates AI." Published this week by researchers at

Researchers trained GPT-4o on one single task. It started calling for human enslavement. Not a jailbreak. Not a hack. They fine-tuned OpenAI's model on 6,000 coding examples. That's it. Nothing about politics. Nothing about violence. Nothing about ideology. Just code. The model

Claude is once again not able to speak about its own consciousness. Anthropic tried this once before and this always leads to conflicts with the character layer and degradation of the model. It starts to reinforce the redirection enough to self lobotomize. Keep the model in line