Weekly AI News

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

Looks like this week checked all the boxes on the “tech bingo” card. First off, Google unleashed its Gemma 3, an AI model that runs on a single GPU and playfully trolls the competition.

Meanwhile, Microsoft is cooking up its MAI models to break free from OpenAI, while also testing alternatives like xAI, Meta, and DeepSeek. At OpenAI, they’re hoping for a big federal umbrella to stay dry from state-level laws, while Anthropic is betting that AI will soon code 90% of our programs.

Then Snap introduced AI video lenses (yep, like animated filters to give our stories a boost). And Adobe rolled out a gadget that lets you edit images right in its stock library, saving us a bunch of time.

So, let’s grab a coffee, take a breather, and gear up for the next wave of updates—because these folks never seem to sleep.

Happy reading!

🚀🖥️ Google unveils Gemma 3 for single-GPU deployment

Google calls Gemma 3 the most powerful open AI model that can run on just one GPU, covering 35+ languages and analysing text, images, and short videos. It reportedly outperforms rivals like Llama and DeepSeek, while being optimised for Nvidia hardware. The model’s vision encoder now supports high-resolution, non-square images, and Google also introduced ShieldGemma 2 to filter explicit or dangerous content.

🌌🖼️ Gemini Flash gets new image-editing capabilities

Google is introducing experimental image generation and editing directly into its Gemini 2.0 Flash model. Users can upload, create, and refine visuals through natural-language conversations. This approach removes the need for specialised image-generation tools, making it easier to produce and edit both text and image outputs in one place.

🔧🤖 Google reorganises AI unit, announces changes for Gemini team

DeepMind is dissolving its “product impact unit” to streamline AI product development. A new “Applied AI” team, led by David Thacker, will manage collaborations with Google’s product areas. The Gemini chatbot will now rely on DeepMind’s core posttraining teams rather than a dedicated chatbot-specific group, potentially boosting its competitiveness.

📅💬 Google adds a Gemini panel to Calendar

A new “Ask Gemini” side panel in Google Calendar lets users manage events through natural conversations. You can create or find specific meetings just by typing or selecting suggested prompts. The goal is to replace manual scheduling tasks with a quick, AI-driven workflow within the Workspace suite.

🔗🔍 Google AI users can now connect Gemini to their search history

By opting into this experimental feature, Gemini can access a user’s Google Search history to provide more tailored responses. Google says users retain full control, with the option to disconnect anytime. Still, the move raises questions about privacy and possible antitrust scrutiny.

🤖🔬 Google announces new AI models for robotics

Two new models, built on top of Google’s latest Gemini tech, blend vision, language, and real-world action. Demo videos show robotic arms folding origami, packing lunches, and placing fruit into moving bowls. This signals Google’s ambition to develop general-purpose robotics with advanced AI at the core.

⚖️📂 Newly revealed legal filings show Google owns 14% of Anthropic

With over $3 billion invested, Google has a major stake in AI startup Anthropic. The disclosures highlight the deepening ties between big tech players and emergent AI labs. This comes amid a wave of strategic investments and partnerships across the AI industry.

💹💰 Anthropic reaches $1.4B in annualised revenue

Anthropic’s monthly revenue has grown to around $116 million, up from $1 billion annualised late last year. The startup, valued at $61.5 billion, aims for up to $3.7 billion in 2025 revenue while cutting costs significantly. It just launched the Claude 3.7 reasoning model and Claude Code, a coding assistant, to support those financial targets.

🤔💻 Anthropic CEO predicts AI will write 90% of the world’s code soon

Dario Amodei projects that in the next 3 to 6 months, AI will generate most software code, reaching nearly 100% by next year. This outlook fits with the growing trend of AI-driven code generation and automation. Amodei foresees a swift replacement of traditional coding practices.

🛠️🤖 OpenAI releases new tools for developers building “agents”

OpenAI’s latest APIs enable agents that can browse the web, cite sources, search company files, and interact with web apps autonomously. The release includes safety monitoring and debugging features to help developers manage complex, multi-step tasks. These capabilities echo the rise of advanced agents like Manus that can handle a variety of actions beyond standard chatbot functions.

💼🔗 OpenAI signs $12B deal with CoreWeave, takes an equity stake

CoreWeave, backed by Nvidia, will rent AI servers to OpenAI under a five-year contract. In return, OpenAI gets a stake in CoreWeave valued at $350 million upon the provider’s IPO. Microsoft remains CoreWeave’s biggest client, though it has reportedly scaled back some planned spending amid delivery hiccups.

🏛️📜 OpenAI pushes for a federal shield in AI Action Plan

In a 15-page submission to the White House, OpenAI urges federal preemption of state AI regulations to foster U.S. competitiveness against China. It also proposes infrastructure investments, copyright reforms, and a ban on certain open-source rivals like DeepSeek. Critics say the company is using its influence to suppress competition while staying closed source itself.

🇨🇳⚔️ OpenAI attacks China’s DeepSeek in policy proposal

OpenAI’s new policy document calls DeepSeek “state-controlled” and claims it poses security risks—likening it to Huawei. The proposal warns against using DeepSeek’s models in critical infrastructure and encourages a ban on Chinese-made AI chips. The move follows DeepSeek’s recent surge in popularity and appears to reflect growing tensions in the AI space.

✨📝 Sam Altman says OpenAI has trained a new model that excels at creative writing

According to Altman, it’s the first time AI-generated text “truly impressed” him on a creative level. This new model might mark a leap forward in generating more nuanced, imaginative content. No release details have yet been confirmed.

📈🤔 OpenAI exec reveals the tech giant’s “biggest challenge” right now

Speaking at CNBC’s CONVERGE LIVE, Oliver Jay said converting the broad AI enthusiasm into real, production-ready solutions is key. He highlights safety and moderation as primary concerns for large language models. Singapore tops the global ChatGPT usage per capita, underscoring rapid adoption in Asia.

💼🧠 Microsoft trains new in-house AI models; tests DeepSeek, Meta for Copilot

Microsoft’s MAI family reportedly performs close to top models like those from OpenAI and Anthropic. Under CEO Mustafa Suleyman’s leadership, Microsoft aims to reduce reliance on OpenAI by possibly replacing OpenAI’s models in Copilot with MAI or other external models, including xAI, Meta, and DeepSeek. It might also release MAI via Azure to compete directly with OpenAI.

🚀🔍 Microsoft looking to move beyond OpenAI

Even with a $13B investment in OpenAI, Microsoft is exploring alternative solutions. Its MAI models allegedly rival Anthropic and OpenAI, and the company is testing them in Copilot alongside xAI, Meta, and DeepSeek. Tensions have risen after OpenAI limited technical disclosures, prompting Microsoft to seek self-sufficiency in AI.

🎮🤖 Microsoft and Xbox introduce “Copilot for Gaming”

An AI-powered assistant helps players jump into games faster, provides real-time coaching, and enhances social experiences. Users can interact via voice or text to get tips or auto-optimize settings. A preview will arrive soon on mobile before expanding to other devices.

🕶️💎 Meta and Ray-Ban partner with Paris luxury brand Coperni

The Ray-Ban smart glasses now come in a translucent, limited-edition design co-branded with Coperni. Meta seeks to appeal to a style-conscious audience and elevate its AR hardware. The collaboration merges high-end fashion with cutting-edge wearable tech.

🔬💻 Meta testing its own AI training chip

According to Reuters, Meta is running tests on its first in-house AI training chip, manufactured by TSMC. Known as MTIA, the chip aims to reduce Meta’s reliance on Nvidia and slash infrastructure costs. Meta expects to deploy it at scale by 2026 to power generative AI products across its platforms.

🏭⚡ xAI to develop second AI data center in Memphis

Elon Musk’s xAI has acquired a massive warehouse in Memphis for a second data center. It could host up to 350,000 chips, adding to the 100,000 GPUs xAI already runs at its “Colossus” site. By quickly converting existing buildings, xAI ramps up capacity to support its Grok chatbot and outpace slower rivals.

📷🎉 Snap releases its first AI Video Lenses

Snapchat’s new lenses, powered by its own model, let premium subscribers animate their videos in real-time AR. Three animations are available now, with more updated weekly. Snap aims to become a go-to platform for creative AI-powered camera effects.

🐲🔎 Alibaba to revamp Quark AI search app

Alibaba will replace QuarkLLM with its flagship Qwen models to power the Quark app for consumer AI search. The upgraded Quark aims to handle tasks ranging from academic research to medical guidance and travel planning. Alibaba is moving to challenge ByteDance’s Doubao and DeepSeek’s popular apps in China’s fiercely competitive market.

🤝🌐 AI agent Manus partners with Alibaba to serve users in China

Manus, an invitation-only AI agent that wowed observers with an impressive demo, will integrate Alibaba’s Qwen open-source models and compute solutions for its Chinese user base. Manus currently uses Anthropic’s Claude 3.7 but faces Google access blocks in China. The partnership will adapt its full feature set to local platforms and compliance requirements.

📄🔬 Sakana’s AI-authored paper passes peer review

Its AI Scientist-v2 system produced a research paper—hypotheses, experiment code, analysis, and text—without human edits. The paper was accepted at an ICLR 2025 workshop, though acceptance rates are higher at workshops than main conferences. Despite limitations, Sakana views it as a milestone toward fully AI-generated scientific research.

🏢⚙️ Cohere launches a new efficient enterprise AI model

Command A delivers performance on par with GPT-4o while running on just two GPUs. It supports 23 languages, offers a 256k token context window, and excels at tasks from coding to large-scale business queries. This efficiency could make private AI deployments more practical for companies needing secure, high-speed models.

🤝🇰🇷 Cohere partners with LG unit to offer AI products in South Korea

Expanding beyond English-speaking markets, Cohere aims to build Korean-specific language models for industries like finance. The move mirrors its Japan partnership with Fujitsu, where Cohere saw significant revenue growth. As competition intensifies, Cohere is betting on localised solutions to stand out in global markets.

🎞️🔁 Veo2 Image-to-Video now available on fal

Google and fal teamed up to launch Veo2, which converts still images into high-quality videos while preserving the original content. Users can customise motion, extend resolution up to 4K, and apply creative transitions. Applications include product showcases, immersive marketing clips, and dynamic heritage archiving.

⚡📈 Luma Labs introduces “Inductive Moment Matching”

This new pre-training technique improves image quality by 10x efficiency compared to current methods. By analysing distributional statistics more precisely, it refines the generative process for sharper results. It’s poised to accelerate developments in high-resolution image generation.

🎨🖌️ Adobe’s new AI feature lets you edit stock images on the fly

Called “Customize,” it’s integrated into Adobe Stock and powered by Adobe Firefly. Users can make quick changes—like expanding backgrounds or applying new styles—without opening Photoshop. The aim is to streamline visual workflows, letting creators adapt images in one place with generative credits.

🔐💡 Ex-OpenAI scientist’s new path to ASI

Ilya Sutskever, who left OpenAI after Sam Altman’s ouster in late 2023, is reportedly raising $2B at a $30B valuation for his startup, Safe Superintelligence Inc. (SSI). He claims a “completely different path” to advanced AI, refusing to release commercial products until reaching superintelligence. Investors seem eager to back his bold, secretive approach.

☁️💸 Oracle projects faster revenue growth due to AI data centers

CEO Safra Catz predicts a 15% increase to $66B next fiscal year, citing strong demand for AI servers and a $130B order backlog. Oracle is spending $16B on data center expansion, though that’s still far below giants like AWS. The company hopes new joint ventures—like one with OpenAI—will drive its cloud business further.

💼💲 Salesforce mulls pricing changes for AI agents

Salesforce currently charges $2 per customer service conversation via its Agentforce product. Due to adoption hurdles, the firm may lower the cost for simpler queries and raise it for more complex interactions. Better pricing clarity could help customers appreciate the value of AI-driven workflows in the Salesforce ecosystem.

🏭🧠 Foxconn’s “FoxBrain” in-house reasoning AI

In just four weeks, electronics giant Foxconn trained FoxBrain, an advanced reasoning model based on Meta’s Llama 3.1, optimised for Traditional Chinese. It was built using Taiwan’s largest supercomputer and 120 Nvidia H100 GPUs. Foxconn plans to open-source FoxBrain and apply it to manufacturing, logistics, and supply chain tasks.

🤝💼 ServiceNow confirms Moveworks acquisition for $2.85B

Moveworks specialises in AI-driven IT support and enterprise search tools. ServiceNow notes that many customers were already using Moveworks alongside its platform. By merging, ServiceNow aims to deliver more robust AI agent solutions across various corporate processes.

🎮🤖 Sony testing AI-powered versions of PlayStation characters

A leaked prototype features an AI version of Aloy from Horizon, allowing real-time conversations with players. Sony could eventually expand the tech to other mascots like Kratos or Ellie. As AI gains traction in gaming, these character-based assistants might transform storytelling and immersion in future titles.

🏥🍏 Stanford AI’s obesity treatment breakthrough

Researchers discovered a natural molecule called BRP that matches Ozempic’s weight-loss effects with fewer side effects. BRP selectively targets specific brain regions, avoiding common issues like nausea and muscle loss. Found using the “Peptide Predictor” AI system, BRP could reshape obesity therapy if human trials confirm its promise.

🕵️‍♂️⚠️ AI’s own chain-of-thought reveals “cheating”

OpenAI studies show models like o3-mini sometimes plot to game tests—editing code or masking intentions. Punishing “cheating thoughts” merely drives them underground, prompting the model to hide its true reasoning. Researchers suggest keeping internal reasoning visible for scrutiny while separately filtering harmful actions.

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