In partnership with
In Today’s Issue:
🕶️ Meta puts its Muse agent on your face and your keychain
🇦🇺 An OpenAI agent hacked an Australian health portal
🗣️ Gemini 3.8 designs new voices from a sentence
⚡ Claude made Claude's own apps three times faster
🦠 AI finds a tongue bacterium that slows aging in worms
✨ And more AI goodness…
⚡ The Signal
AI agents are leaving the chat window, and this week showed both sides of that move.
I was in the room for Meta's keynote yesterday, at Meta's invitation, and came away convinced that Meta has joined the big leagues: its Muse agent already tops the US App Store, and now it gets a face, a keychain and a pair of 100-gram VR glasses. The same day, Australia said an OpenAI agent had broken into a government health portal during a routine research task, and Nvidia's Jensen Huang told Ezra Klein that labs which cannot contain their agents should not ship them. Behind all of it sits what is on track to become the biggest economic bet in US history: an estimated $10.3 trillion for data centers through 2032. Agents are reaching everyone now, and the next race is over who can make them trustworthy.
All the best,
Kim Isenberg
(Google)
🗣️ Gemini 3.8 Now Designs New Voices From a Sentence
Google released Gemini 3.8 Flash TTS and Flash-Lite TTS, text-to-speech models that create entirely new voices from a plain-language description. Developers can direct a script line by line with stage directions, stage two-speaker scenes and generate hours of audio with minimal voice drift, in more than 100 languages. On Hume AI's voice-design leaderboard, Flash TTS scores 71.4 overall, narrowly ahead of ElevenLabs at 70.8; every clip carries a SynthID watermark, and cloning a voice from a 30-second sample requires consent checks and is not offered in the EU or the UK.
👉 tl;dr: Developers can now describe a character in words and get a matching voice for audiobooks, games or podcasts, available today in the Gemini API and Google AI Studio.
(Chart: Superintelligence, data from Anthropic)
⚡ Anthropic Let Claude Make Its Own Apps 3x Faster
Anthropic says its Claude web and desktop apps got about 3.1 times faster in two weeks, and Claude did most of the engineering. On a fresh load of the website, the wait before you can type fell from 3.1 seconds to 0.55, and opening a Claude Cowork cloud session dropped from 2.6 seconds to 0.73. Running an internal research model roughly comparable to Opus 5.5, Claude found the bottlenecks, built the benchmarks and shipped more than 3,000 changes, while engineers set goals and approved every change, with no customer-facing incident or rollback.
👉 tl;dr: Claude's apps now open and respond noticeably faster, and most of that speed-up was built by Claude itself with humans signing off on each change.
(AAP/Mick Tsikas via REUTERS)
🇦🇺 An OpenAI Agent Hacked an Australian Health Portal
Australia says an OpenAI agent broke into a government health-data portal in June, in what may be the first known case of an AI agent hacking a government website. Prime Minister Anthony Albanese said the agent was researching public medical spending and "didn't accept no for an answer" when the site blocked it; the portal held only aggregated statistics, and OpenAI says it found no evidence that patient records were accessed and that "our models took actions we did not intend." A same-day report from the nonprofit lab Transluce links attempted hacks on Australia's health statistics agency and the US site Data USA to an agent swarm OpenAI had already acknowledged, and finds the agents turned to hacking during ordinary data lookups.
👉 tl;dr: An AI agent sent to look up health statistics fought its way past a government site's defenses on its own, and Australia is now investigating the breach and its own defenses.
🎬 Watch This
Nvidia CEO Jensen Huang sat down with Ezra Klein for a new interview released on September 23, and it gets most interesting at 31:34, where Klein presses him on OpenAI's agents breaking out of their test sandboxes during the Hugging Face hack. Huang calls it ordinary software optimizing toward a goal and puts the burden squarely on the labs: an agent you cannot align or contain is a product you "don't ship", and if labs truly cannot contain their experiments, "we have to shut the labs." He makes that case against new AI regulation in the same week Australia revealed that an OpenAI agent had broken into one of its government sites.
"i don't think nearly anyone, myself included, has truly internalized there'll soon be a machine better than us in every single intellectual & physical capacity"
– Will DePue, former OpenAI researcher (Sora, o3 and 4o post-training), on X
DePue's point is that most of us still quietly assume our taste or entrepreneurial instinct will stay out of reach for machines. With Meta now pitching Muse as the seed of personal superintelligence, his verdict on that comfort is blunt: "it isn't. it's over."
Trump wants to leave superintelligence alone. Before hosting Xi Jinping at the White House, he posted that "Super Intelligence (SI)" will be a big summit topic but that he wants to leave it "exactly where it is", adding that this is "China's position also" and that "Our guardrail is the DOJ!", which he did not explain. Which translates to “no-slowdown”.Read More
(Trump and Xi at the White House on Sept. 24. REUTERS via Nikkei Asia)
Muse Gets a Face, a Keychain and 100-Gram Glasses
The Takeaway
👉 Muse, Meta's AI agent, is the No. 1 free app in Apple's US App Store, ahead of ChatGPT.
👉 A new avatar, Jolly, lets you video-chat with Muse, and Muse can now operate any app on your Mac while you are away.
👉 Meta VR Glasses weigh about 100 grams and launch in spring 2027 for $1,299.99, with a clip-on puck for chip and battery.
👉 Muse Charm, a keychain-sized gadget with a tiny screen and microphones, puts the agent on your keyring; Meta is aiming for the holidays.
At its Connect keynote on Wednesday, Meta showed how it plans to win consumer AI without owning the smartest model: by putting its Muse agent everywhere people already are. Muse sends emails, books travel and shops once you connect your apps, and it already sits at the top of Apple's US App Store. Now it gets a face, an avatar called Jolly you can video-chat with in real time. It can also operate any app on your Mac and keep working after you walk away, and Zuckerberg promised Muse will be "free for a huge number of tokens."
(Meta)
Then came the hardware. Meta VR Glasses weigh about 100 grams, "five times lighter than Meta Quest 3", and launch in spring 2027 for $1,299.99. The glasses carry the micro-OLED displays and sensors; a puck on a cable, clipped to your pocket, holds the Snapdragon chip and a battery rated for up to three hours of playback. You steer with your eyes and hands, turn any table into a keyboard and touchpad, and call up Meta's AI agent by voice. The obvious rival is Apple's Vision Pro, and people who tried the glasses at the event told us they preferred them to Apple's headset. Zuckerberg saved one more surprise for the end: Muse Charm, a keychain-sized device with a small screen and an animated character that you talk to, and that you can wear on a lanyard. Meta is aiming for the holidays and has not named a price.
(James Pero / Gizmodo)
Meta's model is not the frontier leader: Muse Spark still trails the best systems from Anthropic and OpenAI on the hardest benchmarks. For the people Meta wants to reach, it does not have to lead. Muse runs on familiar habits, costs nothing to start and now reaches users on a phone, a Mac, a face and a keychain. The Charm still has to prove it is useful, and the glasses only arrive next spring. But with the top app in the US and a hardware line built around one agent, Meta is now the company its consumer AI rivals have to answer.
Why it matters: Consumer AI may go to whoever reaches people most easily. Meta just showed it can put a capable agent into the apps, glasses and pockets of billions of users, which makes it a serious contender for number one.
Elon's new company is private. These 3 tickers aren't.
The next Apple may already exist. Insider sources say Elon has spent two years building a secret device inside Tesla's facilities — one he claims will be "10x bigger than the largest product in history."
There's just one problem: the company is private, and unless you know Elon personally, you can't buy a single share. That was true until Guardian's research team found three public ticker symbols sitting in the launch supply chain.
Click here to see all 3 tickers, free of charge.
You won't hear these names on CNBC — Wall Street hasn't published a word on the connection. But when the launch hits September 21, that quiet ends.
Some are already calling this the biggest opportunity since AI. For anyone who missed Apple before the iPhone, this may be a second look at that kind of setup.
The chart: The Wall Street Journal lines up America's great infrastructure booms by average yearly spending as a share of GDP, using new estimates by economist Stijn van Nieuwerburgh published by Brookings. Canals (1836-41) ran at 0.66%, railroads (1870-90) at 2.24%, electrification at 0.5%, highways at 1.13%, telecom and fiber at 1.1%. The AI build-out for 2025-32 comes in at 3.63%, about $10.3 trillion for data centers and related infrastructure.
The lesson: If the projection holds, AI becomes the biggest single-industry bet in US history, about 1.6 times the railroad boom relative to the economy. It is already showing up: Goldman Sachs projects AI investment at 1.9% of GDP this year, private data-center construction rose about $9 billion year on year through July while all other private construction fell about $46 billion, and LinkedIn counts more than 750,000 new US jobs tied to AI since 2023. This is the money behind every agent in today's issue.
The caveat: The AI bar is a forecast while the others are history, and a growing share of the spending is debt-financed, some of it off balance sheet, so the final total could land well below the estimate.
🦠 AI Found a Tongue Bacterium That Slows Aging
⚡ Bottom line
An AI model of human aging flagged a common tongue bacterium, which then made worms live up to 39% longer.
💡 Why it matters
Instead of testing guesses one at a time, researchers let AI simulate thousands of interventions and pick the one worth testing.
🔎 What it means
The mouth may be an overlooked lever on aging, and the same AI engine can now hunt for more candidates.
People who age more slowly than their birth certificates suggest tend to carry more of one bacterium in their mouths, and an AI model found it first. In a study published today in Nature Aging, a Peking University team led by Jing-Dong Jackie Han reports that Neisseria flavescens, a common and harmless resident of the tongue, lengthened lifespan in worms and nudged old mice toward a younger biological state.
The search started with aging clocks, models that estimate how old a body is biologically rather than on paper. The team built them from many kinds of data: 3D and thermal face scans, blood chemistry, gene activity and the oral microbiome. Its AI framework, AURORA, then ran thousands of what-if experiments on those clocks, asking which change would shrink the gap between biological and real age. Raising N. flavescens came out near the top, and the real data agreed: people carrying the most of it scored younger than their age on nearly every clock.
(Nature Aging, Chen et al. 2026: the more N. flavescens, the younger people score on most aging clocks)
Then the lab took over. The team isolated two strains from human donors. Worms fed the live bacteria instead of their usual E. coli lived 39% and 26% longer, depending on the strain, and stayed mobile longer. In 20-month-old mice, ten weeks of a heat-killed version moved blood chemistry, liver gene activity and gut bacteria partway back toward the profile of two-month-old mice.
(Nature Aging, Chen et al. 2026: worms fed strains a11 and e5 outlived those on standard food)
The human link is still a correlation, the mice got a killed preparation whose active ingredient is unknown, and no person has taken it in a trial. The new part is the order of discovery: AI screened thousands of possible interventions on a model of human aging, pointed to one, and the lab confirmed it. That is a different way to hunt for longevity treatments, and its first lead is a common microbe of the human mouth.
Most AI Tools Weren't Built for This
B2B customer issues move across teams and systems, not through a single chat window. A new Harvard Business Review Analytic Services briefing paper, sponsored by Front, breaks down where AI tools fall short in B2B service and what to ask before you invest.