Weekly AI News #30

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

What an emotional mess last week.

On one side, users are mourning their "only friend" lost with the GPT-5 update. On the other, OpenAI doubles rate limits and hastily brings back GPT-4o amid the backlash. The chaos says it all: technically superior model, users screaming, Altman backtracking within 24 hours.

Why? People don't care that AI wins programming olympiads. They want their usual emotional companion back. Creepy and touching at the same time. While we're growing dangerously attached to our AIs, they're advancing everywhere: space medicine with NASA, revolutionary antibiotics at MIT, cancer drugs in Korea. But watch out—the same AI that saves lives makes doctors less capable when it's not there to assist them.

Margaret Boden was right all along: machines simulate intelligence, not consciousness.

Happy reading!

🧠💰 Altman says OpenAI to lose money for a while

OpenAI CEO Sam Altman said the ChatGPT creator should keep "running at a loss" while investing in training new models, as long as performance improvements continue. He told CNBC that being private allows this flexibility, understanding why investors want exposure but promising "tremendous upside" when OpenAI potentially goes public. The company unveiled GPT-5 and is expected to be valued at $500 billion in an upcoming employee stock sale.

🚨🔧 Sam Altman details GPT-5 fixes in emergency AMA

OpenAI held a Reddit Q&A following GPT-5's polarizing rollout that angered users due to technical failures and abrupt removal of older models. An autoswitcher crash on launch day prevented GPT-5 from routing queries correctly, making it appear less capable. OpenAI is now rolling out fixes, doubling Plus user rate limits, and promising more transparency for future updates.

🔄❤️ ChatGPT brings back 4o as users missed it

OpenAI is restoring GPT-4o in ChatGPT just one day after replacing it with GPT-5, following user backlash. CEO Sam Altman confirmed paid users can switch to 4o, acknowledging they underestimated how much users valued its personality and emotional intelligence. The company will provide "plenty of notice" if the model is ever deprecated.

📊🎯 OpenAI clarifies GPT-5 context window is 196k

OpenAI clarified that GPT-5 thinking's context window is 196k tokens, not the previously reported 32k that caused confusion. The 32k window applies to the non-reasoning model. This clarification addresses early confusion about the model's capabilities and limitations.

🏆💻 OpenAI's reasoner snags gold at programming olympiad

OpenAI's reasoning model achieved a gold-level score at the 2025 International Olympiad in Informatics, placing 6th against humans and first among AI. The general-purpose reasoner scored in the 98th percentile, a massive jump from 49% a year ago. The same model also won gold at the International Math Olympiad and AtCoder.

🧠💉 OpenAI set to back Neuralink competitor

OpenAI is reportedly backing Merge Labs, a brain-computer interface startup raising at an $850M valuation, with Sam Altman as co-founder. OpenAI's venture arm plans to lead the funding round, marking its first major bet on brain-computer interfaces. The project aims to compete directly with Elon Musk's Neuralink, which projects 20,000 implants annually by 2031.

🧠💾 Google's new Gemini memory feature

Google is deploying "Personal Context" for Gemini, allowing the AI to automatically learn from past conversations to provide more personalized responses. Unlike competitors, Google doesn't yet allow manual editing or deletion of memorized preferences. A new "Temporary Discussion" option creates isolated exchanges that aren't saved or used for training, deleted after 72 hours.

📱⚡ Google's new Gemma model is smaller than ever

Google released Gemma 3 270M, an ultra-small model that can run directly on smartphones and browsers while remaining efficient and capable. The model handled 25 conversations on a Pixel 9 Pro consuming less than 1% battery in tests. Developers can fine-tune it in minutes for specific tasks like offline creative applications.

💻📈 Anthropic muscles up Claude for code war

Anthropic announced a 5x increase in Claude Sonnet 4's context window to 1 million tokens via API, handling 750,000 words or 75,000+ lines of code. The company doubles pricing for requests exceeding 200,000 tokens, justifying costs with improved performance. This positions Anthropic competitively against Google while defending its dominant position in the code market.

🎬🔊 Perplexity adds sound to AI video

Perplexity Pro and Max users can now generate 8-second videos with synchronized audio straight from text prompts. The feature is available on web and mobile with no separate audio editing or post-production required. Users get complete video creation from a single prompt without additional steps.

🌐💸 Perplexity says it's made a bid to buy Google's Chrome

AI-powered search engine Perplexity has offered $34.5 billion to acquire Google's Chrome browser, nearly twice its recent $18 billion valuation. The startup says multiple large investors could fund the transaction in full if DOJ forces Google to sell. Perplexity would keep Google as Chrome's default search engine and maintain the open-source Chromium project.

✍️🎨 Mistral releases Medium 3.1 with creative improvements

Mistral released Mistral Medium 3.1, an upgraded model showing improvements in overall performance and creative writing. The update enhances the model's capabilities across various tasks. This positions Mistral competitively in the evolving LLM landscape.

⏰🔧 DeepSeek's R2 model delayed due to chip issues

DeepSeek's long-awaited R2 model is reportedly being delayed due to training issues with Huawei's Ascend chips. Earlier rumors suggested an August release for the highly anticipated model. The delay highlights ongoing challenges with alternative chip architectures for AI training.

🎯👥 Meta's Superintelligence Lab adds OpenAI researchers

Meta's Superintelligence Lab added three more OpenAI researchers: Edward Sun, Jason Wei, and Hyung Won Chung. Alexandr Wang revealed the new hires joining Meta's growing AI team. This continues the trend of top AI talent moving between major tech companies.

🧠🎥 Meta's AI predicts brain responses to videos

Meta's FAIR team introduced TRIBE, a 1B parameter neural network that predicts human brain responses to movies by analyzing video, audio, and text. The AI correctly predicted over half of brain activity patterns across 1,000 regions after training on 80 hours of content. It works best in areas where sight, sound, and language merge, outperforming single-sense models by 30%.

👓🤖 HTC's new AI glasses take aim at Meta

HTC introduced Vive Eagle AI glasses letting users choose between OpenAI and Google assistants, challenging Meta's Ray-Ban dominance. Features include real-time photo translation across 13 languages, 12MP camera, and extended battery life. Currently Taiwan-only at $520 versus Meta's $300 Ray-Bans.

☁️🤝 Oracle strikes deal with Google to host Gemini AI models

Oracle and Google expanded their cloud agreement to let Oracle's cloud customers rent access to Gemini models. Google's Gemini 2.5 Pro has proven so useful that even some AWS customers have adopted it. This positions Oracle competitively as hosting top AI models becomes table stakes for cloud providers.

🚀💵 Lovable eyes $1B ARR in the next year AI coding startup

Lovable projects $1B in annual recurring revenue within 12 months, adding at least $8M ARR monthly. The company expects to hit $250M by year's end, just over a year after crossing $1M. Founded in 2023, Lovable reached a $1.8B valuation with a $200M Series A this summer.

🏢💰 Cohere hits $6.8B valuation with $500M raise

Enterprise AI model maker Cohere raised $500M at a $6.8B valuation, up from $5.5B last year. The Toronto-based company focuses on secure LLMs for enterprise clients including Oracle, Dell, SAP, and RBC. They also nabbed Meta AI research head Joelle Pineau as chief AI officer.

💼🚀 Ex-OpenAI researcher raises $1.5B for AI hedge fund

Former OpenAI researcher Leopold Aschenbrenner reportedly raised over $1.5B for his 'Situational Awareness' AI-focused hedge fund, despite having zero professional investing experience. The fund posted a 47% return in the first half of 2025, outpacing the S&P 500. Aschenbrenner was fired from OpenAI's superalignment team in April 2024 for allegedly leaking sensitive information.

📱🎙️ Apple tests Siri that can run your apps for you

Apple is testing a version of Siri that can take actions inside apps using just voice commands, from editing photos to booking rides. Powered by an upgraded App Intents framework, it's being trialed with YouTube, WhatsApp, Amazon, and Facebook. The full Siri overhaul isn't expected until spring 2026.

🤖🏠 Apple to make Siri centerpiece of robot push

Apple plans an animated Siri at the center of a robotic device launching in 2027, resembling an iPad on a robotic arm. The device will feature FaceTime calls and proactive restaurant/recipe suggestions during conversations. Apple is rebuilding Siri from scratch under codename Linwood while testing Anthropic's Claude as backup.

⚖️📱 Elon Musk threatens to sue Apple over app rankings

Elon Musk threatened to sue Apple for alleged antitrust violations, claiming unfair favoritism of ChatGPT over xAI's Grok in the App Store. ChatGPT ranks first while Grok is fifth on U.S. iPhone free apps. Apple says its store is "designed to be fair and free of bias" with algorithmic recommendations.

🤖🔬 Nvidia debuts "Cosmos" AI models for robot reasoning

Nvidia unveiled Cosmos Reason, a 7B-parameter vision-language model letting robots "reason" about next steps using memory and physics understanding. The rollout includes Cosmos Transfer-2 for synthetic data, neural reconstruction libraries, and Omniverse SDK upgrades. Nvidia is betting robotics is the next trillion-dollar playground for its GPUs.

🇺🇸💰 Nvidia, AMD agree to share China revenue with U.S.

Nvidia and AMD will give the U.S. government 15% of their China chip sales in exchange for export licenses. The arrangement affects Nvidia's H20 and AMD's MI308 chips, for which licenses were granted last week. Trump confirmed the unusual deal and said he'd consider allowing a modified Blackwell chip sale to China.

🇨🇳🔌 Tencent President: We have enough chips for AI training

Tencent's President Martin Lau said the company has sufficient chips to continue developing AI models despite U.S. export controls. Lau said Tencent has "many options" for running AI models without specifying chip types. The comments come as the U.S. eases restrictions while Chinese regulators tell domestic companies to suspend Nvidia purchases.

🚀🏥 Google, NASA's AI doctor for astronauts in space

Google and NASA are developing CMO-DA, an AI medical assistant to diagnose and treat astronauts during deep-space missions with delayed Earth communication. Running on Google Cloud's Vertex AI with models like Llama 3, it achieved 88% diagnostic accuracy in tests. NASA plans to expand with ultrasound imaging and space-specific health condition training.

💊🔬 Korean researchers' AI designs cancer drugs

KAIST researchers developed BInD, a diffusion model designing optimal cancer drug candidates from scratch without prior molecular data. The AI designs both the drug molecule and its protein attachment in one step, targeting only cancer mutations while leaving healthy versions alone. BInD ensures drugs are safe, stable, and manufacturable all at once.

🦠💉 AI's new antibiotics to fight superbugs

MIT researchers used AI to design two new antibiotics capable of killing drug-resistant gonorrhea and MRSA bacteria. Scientists trained models to generate 36M theoretical compounds, producing NG1 and DN1 that attack bacteria through never-before-seen mechanisms. Both compounds cleared infections in mice tests, potentially opening new fronts against deadly infections.

🏥📉 AI is making doctors worse at their jobs

A Lancet study found experienced doctors using AI to spot cancer got worse at finding it without AI assistance. Meanwhile, another paper claims GPT-5 achieved "superhuman" medical performance. This paradox raises questions about whether AI is training better doctors or creating dependency on digital assistance.

🧠🎨 Margaret Boden, philosopher who decoded machine creativity, passes away

British philosopher and cognitive scientist Margaret Boden, a leading figure in AI philosophy, died at 88 in July. She theorized creativity in three categories: combinatorial, exploratory, and transformational, arguing machines could simulate but not truly achieve human-like creativity. Despite her computational approach, she remained skeptical about machines achieving consciousness or intentionality.

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