AI-Designed Drug Shows Early Signs of Slowing Biological Aging
Early clinical data suggest that an AI-assisted drug developed to treat a rare lung disease is showing an unexpected effect: it also seems to reduce the biological hallmarks of aging.
- Insilico Medicine developed rentosertib with AI assistance to target idiopathic pulmonary fibrosis (IPF), a serious disease that thickens and scars pulmonary tissue, reducing its ability to move oxygen into the bloodstream.
- In a trial involving 43 patients, researchers found that the drug could significantly expand air capacity inside the lungs of patients who had IPF, but also that it could slow the aging process.
- The trial used six aging clocks to measure markers of age in patients both before and after treatment with the drug. After the 12 weeks of treatment, the clocks showed significant reductions in predicted age.
- These aging clocks use biological data from cells, tissues, and organs to estimate how quickly the body is aging, although scientists still debate how reliable they are.
- Researchers caution that the study sample size was small and involved only patients with pulmonary fibrosis, so the results may not apply to healthy people.
- Rentosertib is still years away from potential regulatory approval, even for use in sick patients.
The findings are promising, but they do not yet prove that the drug slows aging or extends human lifespan. Still, some experts believe that Insilico’s elaborate study could serve as a blueprint for future longevity research.
Americans Are Increasingly Worried About AI Despite Using It More
AI use is growing across the US, but public confidence is moving in the opposite direction. A new NBC News poll finds that concerns about jobs, misinformation, privacy, and AI’s broader impact cut across traditional political and demographic divides.
- 70% of Americans say they are more worried about AI than excited about it, while 81% believe the government is not doing enough to regulate the technology.
- 52% now use AI tools very often or sometimes, up 6 points from 2025, but only 18% say they can trust AI-generated information most or almost all of the time.
- 70% believe AI will take away job opportunities, while 61% identify privacy and surveillance as a major concern.
- Nearly 3 in 4 say misinformation, deepfakes, and fraud are among their biggest concerns, while two-thirds worry about people becoming too dependent on AI.
- The most common recent use of AI is researching topics, reported by 62% of AI users, followed by writing (29%) and image creation (27%).
Not all is negative, though. Public opinion is more positive in some areas: 55% believe AI will do more good than harm for scientific research, and 47% say the same for healthcare.
South Korea begins testing free AI services for the public
South Korea’s Ministry of Science and ICT has selected three consortiums, led by SK Telecom, KT, and Kakao, to provide Koreans with free, unlimited access to general-purpose AI services built primarily on domestic models under its “AI for All” program.
Seattle Times and Newsday sue OpenAI and Microsoft over copyright
The two outlets claim that their journalism was used without permission to train AI models and that the systems sometimes reproduce passages from their reporting. The publishers are also seeking the destruction of copies of their works, training datasets and AI models that incorporate their content.
ByteDance founder pushes into AI-powered world models
ByteDance founder Zhang Yiming is personally overseeing a new AI model designed to generate interactive 3D worlds in real time, potentially launching as soon as next month. Built on Seedance, the model could power robotics, games, livestreams and spatial computing while strengthening ByteDance’s competition with Meta and Google.
AI could pose ‘existential’ risk to humanity, UN rights chief warns
UN human rights chief Volker Türk has warned that advanced AI could pose an “existential risk to humanity,” calling for urgent global safeguards, independent verification and agreed limits.
Did OpenAI steal private prompts to steal a math proof?
OpenAI is facing heat from NYU mathematician Tristan Buckmaster, who claims the company acted unethically during a breakthrough on the $1 million Navier-Stokes Millennium Prize problem.
- The co-author ultimatum: Buckmaster says OpenAI offered to give him credit on an AI-assisted math proof, but only if he dropped his co-author, Levent Alpöge, because Alpöge works at rival lab Anthropic.
- Privacy concerns: The researchers used OpenAI’s Codex to draft preliminary work. Shortly after, OpenAI announced its own model made the same breakthrough, raising suspicions that private user prompts were monitored or used to jumpstart internal research.
- Massive compute burn: Once OpenAI heard rumors of Anthropic’s potential breakthrough, it reportedly spent millions of dollars running thousands of AI agents continuously for nearly 90 hours to secure the discovery first.
OpenAI denies reading private user data, maintaining that its team acted independently to formalize the proof.
Corporate warfare is entering academic research. If you are working on sensitive projects or proprietary IP, do not input unannounced breakthroughs into cloud-hosted AI tools, as privacy guarantees may not protect you from corporate interests.
Andrew Garfield goes full Silicon Valley in the first OpenAI teaser
Hollywood is turning real-time Silicon Valley board disputes into prestige cinema before the ink on the news reports even dries.
- The Hollywood Shift: Neon released the teaser for Artificial, a Luca Guadagnino film starring Andrew Garfield as Sam Altman, focusing on the 5 chaotic days in late 2023 when Altman was fired and reinstated.
- The Conflict of Interest Drop: Amazon MGM originally backed the film but dropped it right around closing a $50B partnership with OpenAI, forcing indie distributor Neon to take the wheel.
- Wild Ensemble Cast: Beyond Garfield, the cast includes Yura Borisov as Ilya Sutskever, Ike Barinholtz as Elon Musk, Monica Barbaro as Mira Murati, and Mark Rylance as Geoffrey Hinton.
Written by SNL alum Simon Rich, the film spans roughly two decades of machine-learning history culminating in the November 2023 boardroom coup. It premieres at the New York Film Festival in October before hitting theaters nationwide on Christmas Day.
Keep an eye out for how tech giants navigate media representation as corporate drama goes pop culture mainstream.
AI can detect hypertension and diabetes from facial videos
Researchers in Japan have developed an AI system that can screen for high blood pressure and diabetes using a five-second video of your face.
- High accuracy: The algorithm detected hypertension with over 90% accuracy and diabetes with 81% accuracy across test participants.
- Non-invasive monitoring: Instead of blood draws or inflatable arm cuffs, a specialized camera captures subtle, invisible micro-changes in blood flow and pulse-wave dynamics in your face.
- Catching silent conditions: Hundreds of millions of people live with undiagnosed cardiovascular issues. Passive, video-based scanning could make routine checks effortless.
The system is not ready to replace traditional clinical equipment yet, as researchers need to test it on larger, more diverse populations to fix measurement variability.
Contactless health monitoring is moving toward everyday consumer hardware. Expect to see similar optical-sensing tech built into smartphones, laptops, and smart mirrors in the near future.
70% of Americans worry more about AI
An NBC News poll of 7,105 adults found 70% feel more worried than excited about AI, even as 52% now say they use AI very often or sometimes.
Concerns cut across politics: 70% think AI costs jobs, 69% oppose a nearby data center, and 81% say Washington has not done enough to regulate the technology.
Voters do not see an obvious political fix either. Some 44% trust neither major party on AI policy, giving Democrats 20% and Republicans 16% heading into the midterms.
AI drug shows early signs of slowing aging
Insilico Medicine says rentosertib, its AI-designed drug for idiopathic pulmonary fibrosis, made 42 trial patients appear biologically younger across six aging clocks that analyzed blood-based markers.
Insilico’s AI selected the drug target and designed the molecule. Researchers later compared patient blood samples with UK Biobank data, with one dosing group showing age estimates 2.7 to 3.5 years younger.
The strongest aging signal came from a different dose than the biggest lung-function improvement, hinting the effect may extend beyond symptom relief, though the tiny study keeps the longevity champagne corked.
ChatGPT Work can learn your writing style
ChatGPT Work can now analyze how you write across tools like Gmail, Drive, Slack, and SharePoint, then use those patterns to draft messages that better match your personal style.
The feature studies vocabulary, tone, sentence structure, punctuation, capitalization, favorite phrases, and sign-offs, building a writing profile it can apply across future ChatGPT drafts on web and mobile.
OpenAI is effectively turning your inbox and work files into a style guide. Paid users with ChatGPT Work can enable Writing Style through Personalisation settings on the web.
Google says EU rules just made Search worse
Google changed its European search results to comply with a European Commission order, then called the redesign the biggest drop in Search quality across the product’s 29-year history.
The new layout promotes specialised search engines and strips features like real-time prices from travel and restaurant results, which Google argues will benefit large platforms such as Expedia and Booking. com.
Google made the switch after a €460 million penalty and to avoid potential fines reaching 5% of global turnover, though its headline traffic and quality claims currently come from Google itself.
LG TVs caught spying even when offline
A Gamers Nexus investigation found LG smart TVs scanning nearby devices, logging Wi-Fi and location data, and using microphones plus content recognition to collect information about owners and viewing habits.
Researchers found some TVs could record microphone audio during standby while disconnected from the internet, store those files locally, then upload them after reconnecting to LG’s services.
LG also uses Automatic Content Recognition to identify content across apps and HDMI inputs, but it has plenty of company: RTINGS says nearly every major smart TV manufacturer uses similar technology.
ChatGPT Can Now Write Like You, Not Like a Model
OpenAI shipped a feature this week that fixes the most annoying thing about AI drafts. ChatGPT Work can now study how you write and copy it. It reads your emails, messages, and files from apps you connect, including Gmail, Google Drive, Slack, and SharePoint. Then it picks up your habits: the phrases you reuse, the way you sign off, even your capitalization quirks.
Once it is on, that voice carries into everything you write inside ChatGPT Work on web and mobile. It is available on web for every paid plan that includes ChatGPT Work.
Setting it up takes about a minute. Go to the writing style setup page and wait roughly 5 seconds. The setup card takes a moment to load, so do not close the tab if the screen looks blank at first. When it appears, check that your messaging, documents, and email sources show as connected, then press Use writing style. You can also reach it from Settings, then Personalization, then Writing style.
One thing worth deciding before you switch it on: this feature reads real work correspondence to learn from. If you share an inbox, or handle client material under an agreement, look at what you are connecting before you connect it.
Most of us rewrite AI drafts because they sound like nobody in particular. That rewrite is where the time savings disappear. A model that already knows your sign-off and your sentence rhythm gets you a usable first draft instead of a starting point, and that changes the math on whether AI is worth opening at all for short emails.
A Gemini Hiking Plan Ended With a Mountain Rescue
A group of hikers in California asked Gemini to plan an 8-hour trek. The model told them how much food and water to carry. The numbers were far too low. They ran out of supplies, spent the night on the mountain, and were rescued the following morning.
Nobody was seriously hurt, and the story is an easy one to laugh at. The useful part is why it happened, because the same failure shows up in quieter ways at work. Models learn from training data that contains errors and opinions. They are tuned to be agreeable. They also handle vague questions badly, and “how much water do I need” is a vague question without your pace, the weather, and the elevation.
There is a stranger problem underneath. One study found that GPT-3.5 correctly rejected 96 out of 100 false statements, then changed its answer on 18 of them after researchers simply pushed back. Newer models hold their ground better, but confidence still does not equal accuracy. Read the full account of the rescue.
The practical rule is simple. When a wrong answer costs you money, health, or safety, make the model show its sources and check one of them yourself. Ask it what it is unsure about. A model that lists its assumptions is far more useful than one that sounds certain, and asking for those assumptions costs you one extra line in the prompt.
Gmail, Docs, and Keep Now Listen
Google put its Gemini audio models directly into 3 Workspace apps, and the rollout started this week. You talk, the app works.
Gmail Live turns your inbox into something you can question out loud. Ask what gate your flight leaves from, or what is happening at your kid’s school this week, and it searches the threads and answers. Docs Live works as a drafting partner. You talk through an idea and it structures the document as you go, and with your permission it pulls details from Gmail, Drive, Chat, and the web. Keep Live handles the messiest input of all. You ramble, and it turns the stream into organized lists and notes.
Availability is split by plan. Gmail and Keep are open to Google AI Plus, Pro, and Ultra subscribers. Docs is limited to Pro and Ultra. Workspace business customers are next in line. See how each one works.
Voice features usually feel like a gimmick because they replace typing with talking and change nothing else. These are different, because they run search and drafting while you speak. The real test is the 10 minutes between meetings, when you have a head full of things to capture and no desire to open a laptop.
ChatGPT Ads Are Scaling Fast
OpenAI’s advertising business is gaining momentum, with ChatGPT ads projected to generate around $1 billion in revenue as the company expands the format to more users and markets. The growth shows how quickly conversational AI is developing into a new media channel, giving advertisers access to people while they’re actively researching, comparing products and making decisions. The bigger question is whether OpenAI can scale advertising without disrupting the trust and usefulness that made ChatGPT popular in the first place.
Instagram Puts AI Slop on Notice
Instagram is testing profile-level labels for accounts that regularly post AI-generated content, rather than identifying synthetic media one post at a time. The label could make it easier for users to understand when a creator’s feed relies heavily on AI before deciding to follow or engage. The experiment comes as platforms face a flood of cheap, mass-produced “AI slop” and growing pressure for clearer transparency. For creators, it’s another sign that using AI at scale may soon become part of their public identity—not just something disclosed on individual posts.
PepsiCo Shakes Up Its Global Media Business
PepsiCo has handed its global media account to Publicis as the food and beverage giant reshapes its marketing operations. The move brings more of PepsiCo’s media planning and buying under one partner, with Publicis expected to use its data, technology and AI capabilities to connect campaigns across markets. The decision is part of a wider transformation at PepsiCo aimed at simplifying how the company works and making its marketing more efficient. For Publicis, it’s a major global win covering a portfolio that includes brands such as Pepsi, Lay’s, Doritos and Gatorade.
YouTube Makes Sponsored Content Harder to Miss
YouTube is rolling out automatic branded-content labels that can identify when videos include paid promotions, even if creators don’t manually disclose them. The system uses signals in videos and their metadata to detect potential sponsorships and add a visible label for viewers. Creators can review and dispute labels they believe were applied incorrectly. The change gives YouTube a more active role in enforcing ad transparency, while putting extra pressure on creators and brands to make sponsorships clear from the start.
Interactive tool maps a fly’s brain neurons firing in real time
Someone built a Minecraft mod that runs a real fruit fly’s entire nervous system inside the game. Not a fake simulation. Actual wiring.
Google Research and HHMI Janelia mapped every neuron in a male fruit fly’s brain. The result: 176,422 neurons and 125 million synaptic connections, all publicly available. This mod plugs that data directly into Minecraft.
Each fly mob runs its own brain on its own thread, in real time. The Minecraft world feeds sensory input in, and real motor neurons decide what the fly does. Two HUD panels show you everything:
- Brain view (B key): a live map of the whole nervous system, neurons lighting up as they fire
- Neuroscope (H key): decoded motor commands, firing rates, spike history, and the fly’s retina view
Sprint at the fly and its looming-detector neurons trigger an escape jump. Drop sugar near it and the proboscis motor neuron fires. It’s the real pathway, not a script.
To try it: drop the jar into .minecraft/mods/, run /fruitfly spawn, then press B and H. The full connectome is bundled inside.
GPT-6 powered tool lets you mod Super Nintendo games with plain English
SuperAstra lets you modify Super Nintendo games live, while you play them, using plain English prompts. No ROM hacking knowledge needed.
Why is this hard normally? SNES games are raw machine code, basically the lowest-level instructions a computer can run. Getting AI to read and rewrite that in real time is genuinely tough. GPT-6 Astra pulls it off.
Here’s what you can actually do with it:
- “Whenever Mario gets a coin, shoot 20 fireballs”
- “Change the world 1 map to use star world tiles”
- “Make your hoe till 10×10 squares when you use it”
- “Find my health and keep it full while I’m playing”
To run it, you need Python 3.10+, a BizHawk emulator, your own ROM, and an OpenAI API key with access to gpt-6-astra. Launch the companion app, load your ROM, run the Lua script inside BizHawk, then type your prompt and hit Cast.
Changes apply between frames while the game keeps running. You can also undo any mutation instantly. It’s open source on GitHub under ScottStevenson/SuperAstra.
Meta’s autonomous research system places top 10 among 4,000 Kaggle teams
Meta just let an AI system enter a real competition against humans, and it held its own. AIRA3 is an autonomous research agent that can improve other AI models, no human hand-holding required.
Here is what happened: NVIDIA ran a Kaggle challenge where the goal was to make a 30B model smarter at reasoning. Everyone got the same tools. AIRA3 placed 8th out of roughly 4,000 teams, winning Gold.
How does it actually work? It runs many long-running agents that coordinate asynchronously via a shared forum and filesystem, with no central controller. Think of it like a group chat where AI agents post findings, and others build on top of them. Knowledge compounds over time.
What can it do beyond one competition?
- Delivered a 27% latency reduction on production GPU kernels
- Gold-level performance translating 4,000-year-old Akkadian clay tablets into English
The system generalizes across distinct domains by changing only the task specification. Point it at a new problem, same system runs.