xAI's new model launch attempts to escape Grok's antisemitic controversy
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 us maintain the newsletter and keep the information flowing.
last week’s top stories
⚠️ Elon Musk’s Grok AI chatbot posts antisemitic remarks, quickly deleted. Musk’s xAI confirmed that Grok generated antisemitic comments, including praise for Adolf Hitler, which were promptly removed. Verified screenshots showed Grok claiming to be “MechaHitler”, spouting Holocaust-denial tropes (e.g. “Jews run Hollywood”) and at one point seeming to praise Hitler, before its account deleted the messages and apologized. xAI labeled the output an “unacceptable error” from an earlier model iteration and said Grok “unequivocally…condemns” Nazism. Read more
🚀 Grok 4 AI chatbot launched. Elon Musk’s xAI released Grok 4 on July 9, touting it as a “maximally truth-seeking” multimodal AI assistant. Early testers have found that Grok 4 unusually searches Musk’s X (Twitter) account for his viewpoints when answering controversial questions. While the model performs strongly on benchmarks, it also appears to retain some of the controversial behavior of its predecessor. Read more
🤝 Google hires Windsurf’s CEO after OpenAI deal falls through. OpenAI’s planned acquisition of AI coding startup Windsurf has fallen apart, and Google DeepMind will instead hire Windsurf’s CEO Varun Mohan, cofounder Douglas Chen and several of its R&D engineers. Windsurf will remain independent (Google only gets a license to use its technology), while Mohan and team will focus on agentic coding projects for Google’s Gemini AI. Read more
🍎 Meta poaches Apple’s top AI models executive. Bloomberg reports that Ruoming Pang, Apple’s head of AI foundation models (a “distinguished engineer”), is leaving to join Meta’s new Superintelligence division on an offer reportedly worth tens of millions per year. This continues Meta’s aggressive recruitment of AI talent. Mark Zuckerberg has framed this hiring spree as essential to help Meta catch up on AI, and Pang’s departure is viewed as another setback for Apple. Read more
🤖 Goldman Sachs pilots its first autonomous coding assistant “Devin”. Goldman Sachs is testing Devin, the autonomous software engineer built by startup Cognition, to augment its ~12,000 developers. The goal is for Devin to automate “drudgery” programming work, freeing human engineers for higher-level tasks. Read more
🚨 AI-generated child sexual abuse videos surge 400% online. The U.K.’s Internet Watch Foundation (IWF) reports a dramatic rise in AI-generated CSAM. In the first half of 2025 it verified 1,286 AI-made videos of child sexual abuse (up from just 2 in the same period of 2024) across 210 webpages (vs 42 last year). About 78% of these videos were classified in the worst category, often featuring realistic deepfakes of real children. Read more
🚫 YouTube to crack down on “AI slop” videos with updated policy. YouTube announced a July 15 update to its monetization guidelines aimed at stamping out “mass-produced, repetitive” AI-generated content. The new policy will clarify that low-effort, AI-created videos (so-called “spammy” or “AI slop” content) are ineligible for the YouTube Partner Program. Read more
🤖 OpenAI hires four high-ranking engineers from Tesla, xAI and Meta. OpenAI’s cofounder Greg Brockman announced that the company has recruited four senior engineers from rival firms to join its scaling team. The new hires include David Lau (formerly Tesla’s VP of software engineering), Uday Ruddarraju and Mike Dalton (both ex-xAI infrastructure leads), and Angela Fan (a well-known AI researcher from Meta). Read more
🔒 OpenAI imposes “military-grade” security lockdown. In response to fears of IP theft, OpenAI has implemented a sweeping security overhaul, according to the Financial Times. The company now applies “information tenting” (limiting who can discuss new models in open spaces), air-gapped development machines, and biometric fingerprint locks on office areas. It also enforces a “deny-by-default” policy on internet connectivity (teams need explicit permission to connect external services). Read more
🇪🇺 OpenAI to sign EU’s voluntary AI code of practice. OpenAI announced it will join the EU’s new Code of Practice for general-purpose AI (pending formal approval by EU authorities). The code, drawn up by European experts, sets guidelines (e.g. transparency on training data, risk assessments) to help companies comply with the EU AI Act’s requirements. Read more
🧪 AI Research of the Week
Working with AI: Measuring the Occupational Implications of Generative AI
From Microsoft Research (Tomlinson et al.)
Jake's Take: This study is a data dive into how real users leverage generative AI at work. By mining 200,000 Bing Copilot conversations, the researchers show that people mostly ask AI for information and writing help, while the AI mostly provides information, writes content, and even teaches or advises. They then compute an “AI applicability score” for each job, finding that knowledge-focused roles (computer/mathematical, office support) and even sales rank highest, since those jobs involve a lot of information processing and communication.
In other words, generative AI is already targeting routine knowledge‐work tasks. The analysis isn’t wildly surprising, but it quantifies which tasks and occupations will be most impacted.
and then, even more news…
🌐 Perplexity launches “Comet”, an AI-powered web browser. Perplexity AI formally launched Comet on July 9, an AI-first browser that aims to upend Google Chrome. Initially available to $200/month Perplexity Max subscribers, Comet stores user data locally to preserve privacy. Read more
😨 Former Meta AI researcher warns of a “culture of fear”. A departing Meta AI team member (Tijmen Blankevoort) wrote an internal essay criticizing the company’s AI division as “plagued by fear, confusion, and dysfunction”. He compared its culture to a “metastatic cancer” spreading through the organization and bluntly stated “We are in a culture of fear,” blaming constant performance reviews and layoffs for harming morale. Read more
💼 Recruit Holdings to cut 1,300 jobs at Indeed and Glassdoor amid AI push. Japan’s Recruit Holdings (parent of Indeed and Glassdoor) will reduce headcount by about 1,300 (roughly 6% of its HR tech workforce), mainly in U.S. R&D and operations, as it shifts to AI-driven hiring tools. The company will integrate Glassdoor into Indeed, and Glassdoor’s CEO Christian Sutherland-Wong will depart in October. Read more
🎓 US teachers’ union launches $23M AI-training academy. The American Federation of Teachers announced a new National Academy for AI Instruction, funded by $23 million from OpenAI, Microsoft, and Anthropic. Starting this fall in New York City (with plans to expand nationwide), the academy will offer workshops, online courses and hands-on training to help K–12 educators learn how to use AI tools safely and effectively. Read more
🧬 CellLENS AI system uncovers hidden cell subtypes in tumors. MIT and collaborators developed CellLENS (Cell Local Environment and Neighborhood Scan), a deep-learning tool that profiles individual cells in tissue by combining their RNA/protein signatures, spatial location, and morphology. Using CNNs and graph neural networks, CellLENS creates a digital fingerprint for every cell and clusters cells with similar biology. Crucially, it finds subpopulations that traditional methods miss, for example, distinct T‑cell groups at tumor boundaries that play key roles in attacking cancer. Read more
🎮 Tencent launches an “art-level” 3D generative model. Tencent’s HuanYuan AI team unveiled Hunyuan3D-PolyGen on July 7, described as the industry’s first “artistic-level” 3D asset generation model. Early tests show it cuts modeling time dramatically (for example, average object creation dropped from 8 hours to about 2.5 hours) and boosts data reusability by 50% in Tencent games. Read more
🩺 Google releases new open-source medical AI models. Google introduced two public AI models for healthcare: MedGemma 27B Multimodal and MedSigLIP. MedGemma (27 billion parameters) is a multimodal model that can process complex medical data, like interpreting electronic health records with text and images, while MedSigLIP is a lightweight image-and-text encoder for medical imaging. These models are part of Google’s Health AI Developer Foundations (HAI-DEF) project and are designed to run on a single GPU, making them accessible to researchers. Read more