OpenAI recruits black hole scientist, while former founder doubts agents' usefulness
Get bigger weekly updates! Free subscribers receive the top stories each week, while Paid subscribers will get a few extra stories. All support for Handy AI directly helps me maintain the newsletter and keep the information flowing.
last week’s top stories
⏳ Karpathy warns true AI agents are a decade away. OpenAI cofounder Andrej Karpathy said current autonomous AI “agents” still lack key capabilities like memory, multi-modal understanding, and continual learning. On the Dwarkesh Podcast, he argued these agents “just don’t work” yet and can’t reliably adapt or remember new information. Karpathy predicted it will take roughly a decade of R&D before AI agents are genuinely useful. Read more
🪖 Palmer Luckey debuts AI-powered military helmet “EagleEye”. Defense startup Anduril, founded by Oculus creator Palmer Luckey, revealed EagleEye – an AI-enhanced mixed-reality helmet for soldiers. The modular headset provides a heads-up display with live maps, target overlays, spatial audio alerts, and even drone/robotics control, essentially giving troops an AI “teammate” in the field. Developed with Meta’s technology (marking a reunion for Luckey and Zuckerberg), EagleEye is meant to turn “warfighters into technomancers” by feeding real-time battlefield data straight into their visors. Read more
🌌 OpenAI recruits a black hole physicist to lead science push. OpenAI hired award-winning theoretical physicist Alex Lupsasca – known for his research on black holes – as the first member of its new “AI for Science” team. Lupsasca will remain a Vanderbilt professor while helping OpenAI apply frontier AI models to unsolved physics problems and guide its scientific research agenda. The move signals OpenAI’s ambition to deepen its credibility in scientific R&D (akin to DeepMind’s AlphaFold in biology) by tackling grand challenges in math and physics with AI. Read more
📹 Google launches Veo 3.1 for AI-generated videos. Google introduced Veo 3.1, its latest generative video model, now with native audio generation, object-level editing, and better prompt alignment. Building on May’s Veo 3, version 3.1 can insert or remove objects in scenes and maintain visual style consistency, and it adds realistic sound and dialogue to make AI clips more cinematic. The model is being rolled out to Google’s Flow video editor and Gemini API, reflecting rapid progress in AI video creation tools. Read more
⚡ Anthropic unveils Claude 4.5 “Haiku” model for speed & cost. Anthropic released Claude 4.5 “Haiku”, a new lightweight version of its Claude AI that offers near frontier-level capabilities (comparable to the earlier Claude Sonnet 4) but at a fraction of the cost. Haiku 4.5 is optimized for speed, delivering ~2–3× faster responses than prior models (reportedly ~198 tokens/sec) and making it ideal for latency-sensitive applications like real-time chatbots. Notably, Claude Haiku 4.5 supports vision inputs and multi-agent use cases, bringing advanced AI within reach of budget-conscious deployments. Read more
🔞 ChatGPT to allow adult content with age-gating. OpenAI will lift its blanket ban on erotic and mature content in ChatGPT for users who verify they are adults, starting in December. CEO Sam Altman said the company had initially restricted the AI for mental health safety, but new safeguards now enable a “treat adults like adults” approach. The update will also let users customize ChatGPT’s tone and personality (even using a ton of emoji or a friendly persona) if they wish, as OpenAI carefully relaxes content limits. Read more
🎵 Spotify plans “artist-first” AI tools to protect music creators. Spotify announced a partnership with major record labels to develop generative AI features that augment music creation while preserving artists’ rights. In a press release, Spotify criticized AI companies that “believe copyright should be abolished,” vowing to build AI products that respect licensing and compensate musicians. With Big Tech firms facing lawsuits for training AIs on copyrighted songs, Spotify is taking a proactive stance to ensure AI innovation in music happens with the industry’s blessing (and without legal battles). Read more
🌏 Global survey finds AI concern outweighs excitement. A new Pew Research Center poll across 25 countries found a median of 34% of adults are “more concerned than excited” about AI’s growing role, while only 16% are more excited. In no country surveyed did enthusiasm for AI exceed concern, and half of Americans expressed distrust in their government’s ability to regulate AI effectively. Higher AI awareness correlated with more wariness, and educated or tech-savvy groups were slightly more optimistic, but overall the world remains wary of AI’s societal impact. Read more
👑 President Trump posts AI-crafted “King Trump” video. As millions protested in nationwide “No Kings” rallies, Donald Trump shared an AI-generated video depicting himself in a crown, flying a jet and dumping brown sludge on protesters. The satirical clip (watermarked by an X user) blared “Danger Zone” and was followed by more fake images of Trump placing a crown on his own head as Democratic leaders bowed. Critics condemned the posts as “disgusting” propaganda, while Trump insisted the protesters were “a joke”, underscoring how political messaging is increasingly embracing deepfakes and meme-like AI content. Read more
⚛️ DeepMind partners with fusion startup to supercharge energy. Google’s DeepMind struck a deal with Commonwealth Fusion Systems (CFS) to apply AI for speeding up nuclear fusion research. CFS will use DeepMind’s physics simulation software (like the TORAX system) to optimize the magnetic confinement of plasma in its SPARC fusion reactor, aiming to reach efficient fusion conditions faster. The partnership, backed by Google’s prior investments in CFS, highlights AI’s potential to accelerate breakthroughs in clean energy (though practical fusion power is still years off). Read more
💻 Nvidia reveals first US-made “Blackwell” AI superchip. Nvidia unveiled its inaugural Blackwell GPU wafer fabricated on US soil – produced at TSMC’s new Arizona plant – amid surging demand for AI chips. The advanced 3 nm chips (part of Nvidia’s next-gen “GB” series) are key to powering large AI models, and building them stateside bolsters the domestic supply chain for critical AI hardware. Nvidia says onshoring this production helps secure America’s leadership in the AI era, aligning with government efforts to localize semiconductor manufacturing. Read more
🧪 AI Research of the Week
Scaling Large Language Models for Next-Generation Single-Cell Analysis
From Google Research & DeepMind with Yale University
Jake’s Take: In this study, the team trained an AI model to “read” cell activity data the way that a language model reads words. They used this model to scan thousands of drugs and spotted a combo: the cancer drug silmitasertib (plus a small dose of interferon) makes tumor cells raise clearer “here I am” flags for the immune system.
Early lab tests in human cells backed the effect. If follow-up studies hold in animals and trials, this could become a playbook for using AI to propose treatment combinations faster and with tighter lab validation.
and then, even more news…
🛠 Anthropic adds Claude “Skills” for custom AI workflows. Anthropic unveiled “Claude Skills,” a way to package instructions, scripts, and resources into reusable modules that Claude can load for specific tasks. Users (and companies like Box and Canva) can create Skills to teach Claude specialized workflows (e.g. following brand guidelines or analyzing spreadsheets) without re-prompting each time. The feature aims to make Claude more useful as an AI assistant at work by letting organizations embed their procedures and context directly into Claude’s agent. Read more