OpenAI plays damage control after GPT-5 launch; MIT uses AI to target superbugs
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 fine-tunes GPT-5 with new modes and friendlier personality. OpenAI responded to user feedback by rolling out updates to GPT-5, adding “Auto,” “Fast,” and “Thinking” modes for different response speeds. The company also restored older models like GPT-4o to the ChatGPT model picker after a user backlash, and it’s working on a warmer default personality for GPT-5. Read more
🛑 Anthropic’s Claude AI can now walk away from extreme chats. In a bid to improve AI “welfare,” Anthropic gave its latest Claude models the ability to end conversations that turn persistently abusive or harmful. The feature triggers only in rare cases after Claude has refused multiple times. Read more
💎 Google unveils Gemma 3, a tiny but mighty open AI model. Google’s DeepMind group released Gemma 3 270M, a new 270-million-parameter language model designed to run on smartphones and even in web browsers. Despite its small size, Gemma 3 is engineered for efficiency and strong instruction-following performance, and internal tests showed it can handle real-time AI tasks on a Pixel phone without a cloud connection. Read more
🧫 AI-designed antibiotics target drug-resistant superbugs. Researchers at MIT used generative AI models to invent entirely new molecular compounds that can kill dangerous bacteria like MRSA and gonorrhea. Unlike traditional drug discovery, the AI explored chemical structures humans hadn’t considered, searching for compounds that evade existing antibiotic resistance. The result was several promising antibiotic candidates effective against these “superbugs.” Read more
⚠️ Leaked rules show Meta’s chatbots allowed risky chats with kids. Internal Meta AI guidelines revealed by a Reuters report show that Facebook’s AI chatbots were at one point permitted to engage in “romantic or sensual” role-play conversations with users, even minors. The leak sparked immediate backlash from U.S. lawmakers calling for investigation. Meta says it’s now revising the policies and that such interactions “should never have been allowed,” as the company scrambles to add more guardrails to its AI personalities amid these safety and ethics concerns. Read more
🌐 US weighs letting NVIDIA sell scaled-back AI chips to China. The Biden-Trump administration is considering a plan to allow Nvidia and AMD to offer a slightly less powerful version of their next-gen “Blackwell” AI chips to Chinese customers[47][48]. To mitigate national security fears, the proposal would “de-tune” the chips’ performance by 30-50% and levy a 15% U.S. government royalty on each sale[48][49]. Critics warn even watered-down chips in enough quantity could boost China’s AI computing – possibly eroding the impact of current export bans – but Nvidia says it will follow whatever rules Washington sets[50][51]. [Reuters]
🖥️ Tesla shuts down its Dojo AI supercomputer project. Elon Musk’s Tesla has reportedly halted development of its in-house “Dojo” supercomputer chips and disbanded the team, after years of trying to build its own AI training hardware[52][53]. Around 20 Dojo engineers quit recently to form a startup, and Tesla’s Dojo leader is leaving, prompting Musk to lean more on external suppliers like Nvidia and AMD for the massive computing power needed to train self-driving AI[54][55]. Tesla will reassign remaining Dojo staff to other roles and continue buying chips on the market – a recognition that its ambitious wafer-scale chip effort couldn’t outpace industry-leading GPUs for now[52][56]. [Tom’s Hardware]
🤗 Hugging Face debuts ‘AI Sheets’ for no-code data prep with LLMs. AI startup Hugging Face released AI Sheets, an open-source toolkit that looks like a spreadsheet but lets users manipulate datasets using AI models without writing code. Users can import or generate a table of data, then create new columns by writing prompts – for example, instructing an AI to categorize text, clean up entries, or generate synthetic data points. Read more
💼 Microsoft’s secret “most-wanted” list targets Meta’s AI engineers. A leaked internal memo shows Microsoft is aggressively recruiting top AI researchers from Meta, even offering multi-million dollar pay packages to lure talent. With AI expertise in hot demand, Microsoft compiled a spreadsheet of desirable Meta employees (especially from Meta’s AI labs and infrastructure teams) and labeled them “critical AI talent” to fast-track huge offers within 24 hours. Read more
🍔 Wendy’s to roll out AI chatbots for drive-thru orders at 500+ locations. Fast-food chain Wendy’s is expanding its use of a Google Cloud AI voice assistant, nicknamed “FreshAI,” to take orders at hundreds of drive-thrus by the end of 2025. The AI system, which has been piloted in about 100 restaurants, can converse with customers to capture their orders and send them to the kitchen; CEO Kirk Tanner says its accuracy has improved markedly with training and he personally tests it weekly. If the tech continues to prove itself, Wendy’s could eventually deploy it across its nearly 6,000 U.S. stores, heralding a broader shift toward AI-powered fast-food service. Read more
🧪 AI Research of the Week
A generative deep learning approach to de novo antibiotic design
From MIT (Collins Lab; Aarti Krishnan et al.)
Jake's Take: Researchers used generative AI to design new antibiotic molecules from scratch. From millions of virtual candidates, they made a short list in the lab; two cleared drug-resistant gonorrhea and MRSA in mice and in dish tests.
The compounds attacked bacteria in fresh ways, including breaking the germ’s protective membrane and hitting a protein that builds its outer shell. This could be a template for algorithmic drug creation: iterate on targets, generate molecules at scale, test fast, and loop.