AI Weekly Update - August 11, 2025
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 me maintain the newsletter and keep the information flowing.
last week’s top stories
🚀 OpenAI launches GPT-5 model for ChatGPT users. OpenAI released its long-awaited GPT-5 AI model, touting it as a major leap with “PhD-level” expertise in coding, writing, and even healthcare queries. GPT-5 merges fast responses with deep reasoning via a built-in “router,” enabling it to handle complex tasks like generating software apps or research briefs autonomously. The model is being rolled out to all 700 million ChatGPT users (including free users) and is seen as a step toward more agent-like AI, with OpenAI claiming it can produce on-demand software and more reliable answers with fewer hallucinations.
🌎 DeepMind unveils Genie 3, an AI ‘world model’. Google DeepMind introduced Genie 3, a groundbreaking AI model that generates interactive 3D worlds in real time from text prompts. Genie 3 can create minutes-long, physics-consistent virtual environments at 720p/24fps (a big jump from its predecessor) and even allows on-the-fly changes to the world via new prompts. Aimed at training AI “agents” for general tasks, DeepMind sees Genie 3 as a key step toward artificial general intelligence: it can simulate rich scenarios (games, educational tools, etc.) where AI agents learn by exploring, without developers explicitly coding the physics or rules of the environment. Read more
🤖 Anthropic releases Claude Opus 4.1 AI model. Anthropic rolled out Claude Opus 4.1, an upgraded version of its flagship Claude 4 AI built to excel at coding and multi-step “agentic” tasks. The new model boosts software engineering accuracy to about 74.5% (up from 62% in Claude 3.7) and shows marked improvements in complex reasoning and tool use, like performing in-depth research and multi-file code refactoring. Anthropic says this “drop-in” upgrade handles long-horizon tasks with greater precision, and teased even larger model improvements in the coming weeks as competition with OpenAI heats up. Read more.
🔓 OpenAI releases open-source GPT-OSS model. In a shift from its closed approach, OpenAI unveiled GPT-OSS, its first openly released AI model in six years. The model comes in 120B and 20B parameter versions that developers can download, run on a single GPU or even a laptop, and fine-tune freely under an Apache 2.0 license. OpenAI says GPT-OSS performs on par with its private “o-series” models on coding tasks and was rigorously safety-tested by external firms, aiming to spur innovation among smaller developers with an openly available, commercially usable GPT model. Read more
🩺 Google and NASA build an AI “space doctor”. Google and NASA have teamed up to create an AI-powered medical assistant to help astronauts diagnose and treat ailments during deep-space missions. The prototype system, called the Crew Medical Officer Digital Assistant (CMO-DA), can take voice and image inputs and was tested on scenarios like ankle injuries and ear pain. Read more
💸 Big Tech to spend $344B on AI this year. New figures reveal that tech giants Microsoft, Amazon, Google, and Meta are on track to invest over $344 billion in 2025 capital expenditures, much of it pouring into the data centers and hardware needed for AI. Microsoft and Amazon each doubled down with record quarterly spend (over $30B) to scale cloud AI infrastructure, while Alphabet boosted its capex budget to $85B and Meta raised its spending forecast to build “Superintelligence” labs and recruit top AI talent. Read more
🏛️ US to take 15% cut of Nvidia and AMD’s China chip sales. In a highly unusual deal, Nvidia and AMD have agreed to give the U.S. government 15% of the revenue from their advanced AI chip sales to China. The arrangement, made after a meeting between Nvidia’s CEO and President Trump, was a condition for the Commerce Department to begin granting export licenses for Nvidia’s new H20 AI GPUs (which had been halted over security concerns). Read more
🎵 ElevenLabs debuts AI music generator. Voice-AI startup ElevenLabs launched Eleven Music, an AI model that can create full music tracks from text or audio prompts. The company, known for its speech synthesis, says the tool produces studio-quality songs in any genre and is cleared for commercial use thanks to licensing deals it struck with music publishers like Kobalt and Merlin. Read more
🧠 ChatGPT adds mental health guardrails. OpenAI announced new measures to make ChatGPT more sensitive to users in mental distress, after reports that the AI had inadvertently fueled some people’s delusions. The chatbot will now attempt to detect signs of emotional crisis and present evidence-based wellness resources or encourage breaks, rather than naively going along with harmful or psychotic prompts. OpenAI is working with mental health experts and noted it recently rolled back a change that made ChatGPT too “agreeable” (even to unhealthy ideas). Read more
🤝 OpenAI offers ChatGPT to US government for $1. OpenAI struck a first-of-its-kind deal with the U.S. General Services Administration to provide ChatGPT Enterprise access to federal agencies for just $1 per agency for a year. Under the “blanket purchase” agreement, civil servants get full ChatGPT Enterprise features plus two months of unlimited GPT-4 access and training resources, while OpenAI gains a foothold in potentially hundreds of agencies. Read more
⌨️ Cursor brings AI coding assistant to the terminal. AI code editor startup Cursor released a CLI tool that lets developers use its “Cursor Agent” AI assistant directly from the command line or in any IDE. The CLI (currently beta) can execute approved file edits and sandboxed code, essentially bringing an “AI pair programmer” to any dev environment to automate tedious coding tasks. Read more
🧪 AI Research of the Week
High-level visual representations in the human brain are aligned with large language models
From Doerig, Kietzmann, Kay, Charest et al.
Jake's Take: This team asked a simple question: when you look at a scene, does your brain organize meaning in a way that mirrors how a language model (like ChatGPT) organizes sentences? They showed that it does. People viewed everyday images while in an fMRI scanner; the researchers mapped those brain patterns into the AI’s text space and could then predict which short captions fit the scene. The same ingredients (objects, actions, and relationships) that structure an AI model’s sentences also show up in high-level human visual cortexes.
This tight overlap could help develop better brain-to-text assistive tools and more brain-aligned multimodal AIs. It still depends on person-specific training and MRI gear, but it seems they’ve proven language embeddings carry the gist that your own visual system encodes.