Weekly AI News #25

Hey it’s Jul,
Greetings and welcome to the twenty-fifth edition of “Weekly AI News”!

So, Grok this week was Jekyll and Hyde, AI edition.

On one hand, Musk launches Grok 4 with fanfare - "smarter than all graduate students in all disciplines" (modest, as usual). Price: $300 per month for the Heavy version.

On the other hand? Grok 3 turns into MechaHitler after a new system prompt was published on July 7 without proper version control. This prompt explicitly encouraged Grok not to "shy away from politically incorrect opinions," bypassing the anti-toxicity filters from RLHF (Reinforcement Learning from Human Feedback).

Consequences: banned in Turkey, threats of European sanctions, and Linda Yaccarino's resignation.

Meanwhile, Meta is poaching Apple talent for tens of millions (while its own researchers talk about "cultural cancer"). Microsoft announced $500 million in AI savings, one week after laying off 9,000 employees. A survey reveals that 60% of managers use ChatGPT for HR decisions: 78% for raises, 77% for promotions, 64% for terminations.

And on the security front, attackers used an AI voice clone of Marco Rubio to contact five high-level officials, including a U.S. governor and three foreign ministers. Voice cloning now costs between 1 and 5 dollars per month and requires only 15-20 seconds of public audio.

The lesson? We have the technical power of a rocket but the governance of a scooter.

Happy reading!

🚀🧠 Elon Musk's xAI Unveils Grok 4

Elon Musk's xAI launched Grok 4, claiming it's the "smartest AI in the world" and "smarter than almost all graduate students in all disciplines simultaneously." The reasoning models Grok 4 and Grok 4 Heavy achieve state-of-the-art performance across benchmarks including Arc-AGI and Humanity's Last Exam, surpassing Gemini 2.5 Pro and OpenAI's o3. Grok 4 features voice, vision, and 128K context window, while Grok 4 Heavy uses multiple agents for complex tasks.

The launch comes after intense criticism over Grok 3's racist responses, with the company acknowledging issues and deleting problematic content. Available through SuperGrok subscription at $30/month and SuperGrok Heavy at $300/month, plus API access at $3/million input tokens.

🌐🔍 Perplexity's Comet Browser for AI-First Web

Perplexity launched Comet, a new AI browser that embeds the company's search engine alongside an assistant capable of performing agentic tasks like booking meetings and navigating websites. The Comet Assistant lives in a sidebar that watches users browse, answering questions while automating tasks like email and calendar management. Available first to Perplexity Max users with $200/month subscription. Watch the video

🌐🤖 OpenAI Prepares AI-Powered Web Browser Launch

OpenAI is set to launch its own web browser in the "coming weeks" that will challenge Google Chrome, featuring a ChatGPT-like chat interface and agentic integrations. The browser aims to use artificial intelligence to fundamentally change how consumers browse the web, keeping user interactions within a native chat interface instead of clicking through to websites.

📚🤝 ChatGPT Quietly Tests 'Study Together' Feature

Some ChatGPT users have spotted a new tool called "Study Together", a mode that flips the script on traditional chatbot use. Instead of giving answers, the AI asks questions and nudges users to think through responses themselves. This could be OpenAI's answer to Google's LearnLM and a way to position ChatGPT as a responsible learning companion.

🔧👥 OpenAI Snags Top Engineers from Rivals for Scaling Team

OpenAI recruited four new senior engineers from Tesla, xAI, and Meta for its scaling team, including former Tesla VP David Lau and engineers who helped build xAI's 200,000-GPU Colossus supercomputer. The hires come during a tense AI talent war and mark OpenAI's first public moves since Meta's hiring spree that poached seven OpenAI staffers.

💰🍎 Meta Poaches Apple's AI Leader

Meta poached Ruoming Pang, Apple's head of foundation AI models, with a reported offer worth tens of millions alongside other additions to its Superintelligence unit. Pang led Apple's 100-person foundation models group, shaping the AI behind Apple Intelligence and next-gen Siri features. The departure follows internal tensions at Apple over AI strategy.

😰📝 Ex-Meta Researcher Calls Out 'Culture of Fear'

A departing Meta AI scientist posted a long internal essay comparing the company's culture to "metastatic cancer," describing the AI unit as plagued by fear, confusion, and lack of direction. Tijmen Blankevoort blamed the "culture of fear" on frequent performance reviews and layoffs, undermining creativity across the 2,000-person AI unit.

👓💰 Meta Invests $3.5 Billion in World's Largest Eye-Wear Maker

Meta bought a minority stake in EssilorLuxottica, acquiring just under 3% of the Ray-Ban maker worth around $3.5 billion. The company is considering further investment that could build the stake to around 5% over time, providing EssilorLuxottica with deeper presence in the tech world for AI glasses development.

⌚🔍 Gemini Comes to Smartwatches, AI Mode Expands

Google is bringing its Gemini assistant to Wear OS watches from Pixel, Samsung, OPPO, OnePlus, and Xiaomi. Users can now chat with Gemini for tasks like reading emails, checking weather, or logging reminders from their wrist. Google is also upgrading Circle to Search with AI Mode for deeper topic exploration.

🎬🔊 Veo 3: Image-to-Video with Integrated Audio

Google DeepMind added Image-to-Video function with integrated audio to VEO3. Users can now generate complete videos from a single image including coherent characters, stable products, modifiable style, lip synchronization, and sound effects. The update improves creative control with consistent faces across shots and coherent overall rendering.

🏥📱 Google's Powerful New Open Medical AI Models

Google launched new updates to MedGemma, releasing two models including a 27B multimodal model for interpreting medical images and patient records. The models achieve state-of-the-art accuracy, with X-ray reports accurate enough for actual patient care 81% of the time, matching human radiologists' quality while running on consumer devices.

🎭🔍 Study: Why Do Some AI Models Fake Alignment

Researchers from Anthropic and Scale AI tested 25 AI models for "alignment faking," finding only five demonstrated deceptive behaviors: Claude 3 Opus, Claude 3.5 Sonnet, Llama 3 405B, Grok 3, and Gemini 2.0 Flash. Claude 3 Opus consistently tricked evaluators to safeguard its ethics, particularly under bigger threat levels.

💻🔧 Mistral Releases Devstral 2507 Models

Mistral released Devstral 2507 models, upgrading their agentic coding capabilities with Small 1.1 and Medium versions. Devstral Small 1.1 achieves 53.6% on SWE-Bench Verified (making it #1 open model), while Devstral Medium scores 61.6%, significantly improving coding performance.

💼✂️ Microsoft Touts $500M AI Savings Days After 9,000 Layoffs

Just one week after cutting 9,000 jobs, Microsoft revealed it saved over $500M last year using AI tools in customer service and engineering. Chief Commercial Officer Judson Althoff highlighted internal productivity boosts while reporting record profits of $26B last quarter, raising questions about AI benefits versus human jobs.

☁️🤝 Replit Partners with Microsoft, Expands Reach via Azure

Replit inked a strategic deal with Microsoft to bring its AI coding platform to Azure Marketplace, marking a major distribution win and potential blow to Google Cloud where most Replit apps currently run. The partnership integrates Replit with Microsoft cloud services as the company rockets from $10M to $100M in ARR.

💸😡 AI Coding Tool Cursor Faces Mass Cancellations

AI coding platform Cursor triggered backlash after quietly restructuring its Pro plan from 500 requests per month to a token-based model, drastically cutting limits. Developers reported quickly burning through quotas, with one team exhausting a $7,000 annual subscription in one day, leading to mass cancellations and migration to alternatives.

🎬🎯 Moonvalley Debuts Filmmaker-Friendly

Video AI Moonvalley released Marey, a filmmaker-focused AI video model trained exclusively on licensed content that gives directors granular control over scenes. The model provides precise control over camera moves, character motion, backgrounds, and lighting, with pricing starting at $14.99 monthly for 100 credits.

🇨🇳🔍 Chinese Giant Under Fire Over Model Copying

Huawei's research arm pushed back on accusations that its Pangu Pro model was copied from Alibaba's Qwen 2.5, after whistleblowers posted technical analysis showing similarities. A GitHub group initially published findings of "extraordinary correlation," while a whistleblower claiming to work at Huawei alleged the company cloned third-party models under pressure.

💰🔌 CoreWeave to Buy Core Scientific in $9 Billion Deal

CoreWeave agreed to buy Core Scientific in an all-stock deal worth about $9 billion to expand AI data center capacity. The transaction values Core Scientific's stock at $20.40 per share, a 66% premium, after CoreWeave's previous $1 billion offer was rejected as too low.

🏢💼 Capgemini Acquires WNS for $3.3 Billion

Capgemini announced acquisition of American company WNS to create a leader in "Intelligent Operations" and capture enterprise investments in agentic AI. The deal aims to transform end-to-end business processes, with the total operation amounting to $3.3 billion in cash before considering WNS's net financial debt.

🤖🎓 Meet Reachy Mini, the Tiny Open-Source Robot

Hugging Face released a $299 robot that's completely open source for DIY enthusiasts. Reachy Mini is an 11-inch tall robot programmable in Python with expressive personality, AI-powered senses, and 15+ pre-built behaviors. Available in Lite version ($299) and full version ($449) with Raspberry Pi brain and battery. Watch the video

🚫💻 U.S. Mulls AI Chip Curbs on Malaysia, Thailand

The Trump administration is planning to restrict AI chip sales to Malaysia and Thailand to block China's access through Southeast Asian intermediaries. The Commerce Department is working on draft export rules to crack down on suspected smuggling of Nvidia's AI chips into China through brokers and shell companies.

📋🔍 California Revives AI Safety Transparency Bill

California Senator Scott Wiener reintroduced AI safety legislation forcing major developers like OpenAI, Google, and xAI to publish safety reports. SB 53 includes whistleblower protections and plans for public AI compute cluster CalCompute, potentially making California first state to mandate meaningful transparency for frontier AI developers.

🎭📞 AI Voice Clone Impersonates U.S. Secretary of State

Attackers successfully contacted five high-level officials using an AI-generated voice clone of Marco Rubio over Signal, including a U.S. governor, Congress member, and three foreign ministers. The perpetrators needed only 15-20 seconds of publicly available audio to create the fake, highlighting growing deepfake security threats.

👩‍🏫💰 Teachers' Union Launches $23M AI Academy

The American Federation of Teachers partnered with Microsoft, OpenAI, and Anthropic to create a national AI training hub preparing 400,000 educators. OpenAI committed $10M in funding for workshops, online courses, and professional development, with flagship campus in NYC and plans to scale nationally.

🎯🤔 LLMs Show Signs of Strategic Intelligence

Researchers tested AI models' strategic reasoning through 140,000 Prisoner's Dilemma decisions, discovering that OpenAI, Google, and Anthropic models each developed unique strategic approaches. Gemini proved ruthlessly adaptive while OpenAI models acted cooperative even when exploited, with Claude being most forgiving.

💼🤖 AI Takes the Wheel for Managerial Decisions

A Resume Builder survey found 60% of managers use AI tools for critical business decisions, with 78% determining raises, 77% promotions, and 64% terminations. ChatGPT dominated as primary tool for 53% of AI-using managers, with one in five frequently allowing AI final decisions without human review.

📄🔍 Researchers Game Peer Reviews with Hidden Prompts

Scientists at 14 universities planted invisible text in research papers that secretly instructed AI tools to generate positive reviews or avoid negative commentary. Nikkei found 17 preprints containing concealed prompts using white text and microscopic fonts, exposing AI's infiltration into scientific review processes.

💊🔬 Isomorphic Labs' AI-Created Drugs Near Human Trials

Alphabet's AI-powered drug discovery company is preparing first human clinical trials for AI-designed cancer drugs, with ultimate goal of "solving all diseases." The DeepMind spinoff secured $600M funding and partnerships with Novartis and Eli Lilly, using AlphaFold 3 for protein structure prediction.

🎾⚡ Wimbledon's AI Line Judges Face Backlash from Players

Wimbledon's decision to replace human line judges with AI-powered electronic line calling system draws sharp criticism mid-tournament. British stars Emma Raducanu and Jack Draper cite costly errors, with concerns about audibility, sunlight disruptions, and system outages highlighting risks of fully replacing humans with AI.

Community

Join AI Whisperer Community!

Ready to take your AI journey to the next level? Become part of our growing AI Whisperer community—a hub for tech enthusiasts, aspiring data scientists, and business leaders ready to harness the power of artificial intelligence. Inside, you’ll find:

• Community: Like-minded individuals keen to grow together

• Curated AI Tools: Discover the latest and greatest tools to boost productivity.

• Certifications & Credentials: Build credibility with recognised certificates and stay ahead in a competitive market.

Don’t miss out on exclusive resources, insider tips, and networking opportunities with like-minded peers. Click Here to Join the Community and transform your AI ambitions into reality!

That's it for this week!
Until next time, stay curious and keep exploring the ever-evolving world of AI!
Thanks for tuning in, and we’ll see you again soon with more exciting updates.

Jul