Weekly AI News

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

Did you catch the AI whirlwind this week? OpenAI unveiled its shiny new o3 and o4‑mini models—smart enough to juggle code, math, even images—and launched Codex CLI for terminal coding magic. Not to be outdone, they rolled out GPT‑4.1 for devs, boasting a whopping 1 million‑token memory while trimming those pesky costs. ChatGPT even flirted with social media, testing a “yeet” feed.

Meanwhile, Anthropic’s Claude got autonomous research mode, and Google teased Gemini 2.5 Flash with a clever “thinking budget.” Nvidia pledged to build chips stateside—freedom fries, anyone?

Oh, and now you might need a verified ID to play with future models. Can you keep up? Your coffee would’ve gone cold by now. But hey, we’re all here for the thrill.

Happy reading!

🤖🧩 OpenAI releases o3 and o4-mini, new coding agent

OpenAI has unveiled o3 and o4-mini, its most advanced reasoning models with agentic access to all ChatGPT tools and visual thinking capabilities.
o3 delivers state-of-the-art performance in coding, math, science, and multimodal benchmarks, while o4-mini optimizes for speed and cost efficiency.
The release also introduces Codex CLI, an open‑source terminal-based coding agent to streamline development workflows.

🛠️💡 OpenAI’s dev-focused GPT-4.1 family

OpenAI launched GPT‑4.1, 4.1 mini, and 4.1 nano as API-only models tailored for developers, boasting major coding and instruction-following improvements.
Each model supports up to 1 million tokens of context—enough for entire codebases—while reducing query costs by about 26 percent compared to GPT‑4o.
Developers report faster, more accurate front-end development, with GPT‑4.1 nano emerging as the fastest, cheapest option yet.

💬📱 OpenAI Develops X-Like Social Feature in ChatGPT

OpenAI is prototyping a social feed inside ChatGPT, dubbed “yeets,” to let users share how they leverage the chatbot for problem-solving and creative projects.
The feature aims to guide users by showcasing real-world use cases and tips directly within the app, putting OpenAI in closer competition with platforms like X.
It remains unclear if this will launch as a standalone social app or integrate directly into ChatGPT, with development feedback coming from internal and external testers.

🆔🔓 Access to future AI models in OpenAI’s API may require a verified ID

OpenAI announced a Verified Organization process that requires government-issued ID to unlock advanced models on its API. Verification is limited to one organization per ID every 90 days and applies selectively based on eligibility. This measure aims to balance broad model access with mitigation of unsafe API usage according to policy.

👷‍♂️🤖 OpenAI Is Creating AI to Do 'All the Things That Software Engineers Hate to Do'

OpenAI is developing an AI agent, A‑SWE, capable of autonomously building apps, running QA tests, identifying bugs, and writing documentation. CFO Sarah Friar revealed that A‑SWE will handle end-to-end software engineering tasks rather than just generating code snippets. The technology could redefine engineering roles by offloading routine development work to intelligent agents.

📚🖼️ OpenAI added a new library tab in ChatGPT for image management

ChatGPT now features a Library tab that consolidates all generated images in one place for both free and paid users. This update simplifies browsing, reusing, and organizing AI-crafted visuals without manual downloads. Users can easily revisit past creations, improving workflow efficiency and asset management in creative projects.

💰🤝 OpenAI in Talks To Buy AI Coding Assistant Windsurf for $3 Billion

OpenAI is negotiating a $3 billion acquisition of Windsurf, formerly Codeium, to strengthen its developer tooling portfolio. The deal would be OpenAI’s largest, aiming to integrate Windsurf’s AI coding assistance into ChatGPT’s ecosystem.
If finalized, it would expand OpenAI’s reach among developer customers and potentially offer new training data for coding models.

⚖️🏛️ OpenAI Should Remain Nonprofit, Former Employees Tell Court

A dozen ex-OpenAI staff filed an amicus brief urging the nonprofit entity to maintain control over the company amid its for-profit conversion plans. They argued that retaining the nonprofit’s governance is vital to uphold OpenAI’s founding mission of benefiting all humanity. The brief supports Elon Musk’s lawsuit, highlighting concerns about mission drift and trust breaches.

🤝💻 OpenAI Held Acquisition Talks with Owner of Coding Assistant Cursor

Late last year, OpenAI discussed acquiring Anysphere, the creator of Cursor, following similar talks with other coding assistant startups. These discussions reflect OpenAI’s strategy to consolidate leading developer tools and access new customer bases. Although the Cursor deal didn’t close, OpenAI is now pursuing Windsurf, showing continued interest in coding assistance acquisitions.

🖼️📝 xAI dropped a ChatGPT Canvas-like Grok Studio for collaboration

xAI has released Grok Studio, a collaborative workspace that mimics ChatGPT’s canvas interface for documents, code, reports, and games. Available to both free and paying users, it allows real-time co-creation with AI across diverse content formats. This move expands xAI’s feature set and competes directly with other AI labs’ agentic interfaces.

🧠🗂️ xAI’s Grok rolled out memory capabilities and Workspaces tab

xAI’s Grok AI now offers memory features to recall past conversations and a new Workspaces tab to organize files and chats. These upgrades enhance continuity and context management for complex projects over time. Users can now seamlessly switch between topics without losing history, boosting productivity and collaboration.

🔍📑 Anthropic’s Claude gains autonomous research powers

Anthropic upgraded Claude with a Research feature that autonomously searches the web and user files for comprehensive, cited answers. The new Google Workspace integration lets Claude access emails, calendars, and docs securely for context-aware assistance. Enterprise users benefit from enhanced document cataloging using RAG to query large repositories in beta rollout.

🇪🇺🤖 Meta to train AI models on EU users’ public content with opt-out

Meta plans to train its generative AI on European users’ public interactions and content, offering an opt-out form for those who object. This initiative aims to better reflect European culture, languages, and history in its AI outputs. The company emphasizes user choice while expanding data for its models across Meta platforms.

💸🦙 Meta Asked Amazon, Microsoft to Help Fund Llama Consortium

Meta reportedly pitched a Llama Consortium to Microsoft, Amazon, and others to share training costs for its Llama model lineage. The proposal included incentives for backers to influence feature development, reflecting Meta’s push to manage rising AI expenses. Early reactions were tepid, but the move highlights the high financial stakes in AI research.

🔭🔍 Meta’s FAIR shares new AI perception research projects

Meta’s FAIR published five open-source research projects on perception and reasoning, including the Perception Encoder and 3D understanding benchmarks.
The work introduces a Meta Perception Language Model and a Collaborative Reasoner framework, demonstrating up to 30% better performance in multi-agent tasks. These advances lay groundwork for more capable embodied AI and machine intelligence.

🤝📱 Perplexity strikes deal with Motorola, eyes Samsung integration

Perplexity AI agreed to integrate its search platform as the default assistant in Motorola phones and is in talks with Samsung for a similar arrangement. The partnership aims to embed AI search directly into mobile experiences, offering real-time, context-aware answers. This deployment could expand Perplexity’s user base and challenge incumbent smart assistant providers.

🗺️📊 Cohere released Embed 4, a multimodal embedding model

Cohere launched Embed 4, delivering state-of-the-art multimodal embeddings with 128K context support and over 100 language compatibility. The model achieves up to 83% storage savings, optimizing resource usage for large-scale applications.
This innovation empowers developers to process vast, multilingual data efficiently in AI-driven tasks.

🔓📷 Alibaba released Wan 2.1-FLF2V-14B for frame-based generation

Alibaba introduced Wan 2.1-FLF2V-14B, an open-source model that accepts first and last frame inputs for coherent video outputs. The dual-frame approach enhances temporal consistency and quality in generated videos. Supporting fine-grained visual storytelling, this release advances user control over AI-driven video creation.

⚡🧠 Google’s Gemini 2.5 Flash introduces controllable thinking budget

Google previewed Gemini 2.5 Flash, a hybrid reasoning AI matching top competitors while adding a “thinking budget” to balance cost, speed, and quality. Users can toggle processing intensity across up to 24k tokens, tailoring AI effort to task complexity. Available via API and experimental in the Gemini app, 2.5 Flash pushes customizable reasoning for diverse workloads.

🎥🚀 Google released Veo 2, its state-of-the-art video generation model

Google launched Veo 2 in the Gemini app, Whisk, and AI Studio for advanced plan users, showcasing cutting-edge video synthesis. The model generates high-fidelity videos from text prompts and integrates with existing Google AI tools for seamless workflows. This release marks Google’s significant entry into AI-driven video content creation.

⚖️📢 Court Rules Google Breached Antitrust Law in Ad Tech

A Virginia court found Google guilty of violating antitrust laws by maintaining monopoly power in online ad sales for independent websites. The ruling could force Google to dismantle key ad tech operations impacting publishers and advertisers. Google pledged to appeal the decision while defending its ad tools as simple and effective.

🖥️🤖 Copilot gets hands-on computer use capability

Microsoft rolled out a new “computer use” feature in Copilot Studio, enabling AI agents to interact with GUIs by clicking, typing, and navigating applications automatically. Agents adapt to interface changes in real time, maintaining reliable automation without dedicated APIs. This update expands automation to legacy systems and enhances business workflow integration.

🎬🖼️ Kling AI drops new video and image models (KLING 2.0 & KOLORS 2.0)

Kling AI unveiled KLING 2.0 Master for cinematic video generation and KOLORS 2.0 for photorealistic images across 60+ styles. Both models improve prompt adherence, realism, and editing capabilities like inpainting and restyling. These releases strengthen Chinese AI startups’ position in the global creative AI landscape.

🌊📹 ByteDance’s efficient Seaweed video AI debuts

ByteDance unveiled Seaweed, a 7B-parameter video model competing with larger systems by generating up to 20-second clips efficiently. It excels in text-to-video, image-to-video, and audio-driven synthesis, with strong human evaluation scores and realistic motion. Seaweed’s resource efficiency underscores China’s leadership in compact, high-performance AI video generation.

🇺🇸💻 Nvidia to manufacture AI chips and supercomputers in the US

Nvidia announced plans to produce its AI chips at TSMC’s Arizona facility and supercomputers at Wistron’s Dallas plant within 18 months. This U.S. manufacturing push supports $500 billion of AI infrastructure growth over five years and navigates new import tariffs. The move strengthens domestic supply chains amid global trade pressures.

🤖🏭 Hugging Face acquires Pollen Robotics

Hugging Face acquired Bordeaux-based Pollen Robotics, maker of the Reachy2 humanoid lab robot, expanding its hardware offerings to 30 employees. The deal brings robotics expertise and integrates Reachy2 into Hugging Face’s open-source ecosystem. This acquisition deepens Hugging Face’s reach in AI-driven physical intelligence research.

💵🚀 Ilya’s SSI raises $2B at a $32B valuation

Safe Superintelligence Inc. (SSI), co-founded by former OpenAI scientist Ilya Sutskever, secured $2 billion in funding at a $32 billion post-money valuation. The round was led by Greenoaks, Lightspeed, and Andreessen Horowitz, with Alphabet and Nvidia also participating. SSI’s focus on superintelligence and safety attracts massive investor confidence before a concrete product launch.

🛡️🤖 NATO and Palantir strike deal on warfighting AI software

NATO signed a contract with Palantir to deploy AI software for intelligence processing in military operations. The deal, which boosted Palantir’s stock by 4.6%, underscores the alliance’s drive to modernize battlefield data analysis. Government and defense engagements continue to be pivotal revenue drivers for Palantir.

🎬🧠 Netflix tests OpenAI-powered search based on mood and context

Netflix is piloting an OpenAI-powered search engine that lets subscribers find shows and movies using descriptive cues like mood and scenario. Available to select users in Australia and New Zealand on iOS, the tool expands beyond traditional genre and actor filters. The test will roll out to more markets soon, enhancing content discovery with AI.

🕵️🎮 AI models play detective in Phoenix Wright: Ace Attorney game

Researchers at UC San Diego’s Hao AI Lab benchmarked leading models, including GPT‑4.1 and Claude Sonnet, on the video game Phoenix Wright: Ace Attorney. Models were tested on identifying evidence and cross-referencing witness statements, with top systems achieving up to 26 correct evidences but none solving the entire case. This experiment highlights AI’s progress and limitations in interactive, long-context reasoning tasks.

🐬💬 Google’s AI to decode dolphin speech with DolphinGemma

Google unveiled DolphinGemma, an AI model trained on decades of Wild Dolphin Project data to analyze dolphin vocalizations and predict patterns. The project includes an underwater CHAT device built on Pixel 9 hardware for interactive research in real time. DolphinGemma will be open-source this summer, offering global scientists tools to study dolphin communication.

🩺🤖 AI surpasses experts in tuberculosis diagnosis with ULTR-AI

Swiss researchers presented ULTR-AI, an AI system that reads lung ultrasound images via smartphone-connected devices with 93% sensitivity and 81% specificity.
Tested on 504 patients, it outperformed human experts by detecting subtle lesions invisible to the naked eye. This real-time, low-cost approach could revolutionize TB screening in resource-limited settings.

🧬📈 Profluent finds scaling laws for protein-design AI with ProGen3

Biotech firm Profluent introduced ProGen3, a 46B-parameter model trained on 3.4B protein sequences, demonstrating that larger models and data enhance protein engineering results. It produced novel antibodies matching therapeutic performance and designed compact gene-editing proteins.

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