☕️ Claude now watermarks all AI text

Hi there, this is your daily ☕️ Techpresso.

In today's Techpresso:

🔏 Claude now watermarks all AI text

💰 Nvidia unlocks $500B for AI data centers

🔓 OpenAI launches cyber AI

🤖 Humanoid robot sales set to triple

📷 iOS 27 may verify iPhone photos

🎵 Spotify will label AI artist profiles

Plus: 🎁 10 other news you might like, 🧰 6 tools, and 📚 5 papers.

🔏 Claude now watermarks all AI text LINK

Anthropic has started adding invisible watermarks to text and images made by its Claude chatbot, letting automated systems spot AI-generated content while keeping the markers hidden from human readers, in line with the EU's AI Act.

For text, Claude likely picks certain words more often so the pattern shows up in statistical checks, though Anthropic warns a detected "Claude mark" only signals the content may have passed through Claude and doesn't confirm where it came from.

The watermarking applies worldwide and covers uses like Claude Code, starting with models released after August 2, though Anthropic plans to extend it to older ones and admits marks can vanish through edits, screenshots, or short text.

💰 Nvidia unlocks $500B for AI data centers LINK

Nvidia has signed letters of intent with six financial giants, Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR, to raise more than $500 billion in outside money for AI data centers, chip factories, and power plants.

To make the deals attractive, Nvidia will guarantee up to 25 percent of the residual value of its own chips per project, covering part of the gap if installed hardware sells or reuses for less than expected at a financing term's end.

Jensen Huang says the figure is a multi-year target, not Nvidia revenue, and argues chips last longer than critics claim, pointing to the 2020 A100 still in use and rising rental prices, with H100 contracts climbing to $2.35 per GPU-hour by March 2026.

🔓 OpenAI launches cyber AI LINK

OpenAI released GPT-5.6-Cyber, a version of its GPT-5.6 Sol model fine-tuned to find security flaws and build exploits for approved defenders, doing risky "dual-use" work that its regular models usually refuse.

On OpenAI's internal cybersecurity benchmark, the model finished 95% of advanced tasks like exploit-chain building and privilege escalation, far above the 57% from GPT-5.5-Cyber and 1.5% from the standard, guarded GPT-5.6 Sol.

Access is limited to security teams vetted through the new Daybreak Red tier, and OpenAI says the model already found two zero-days in Chrome's V8 engine, priced at $12.50 per million input tokens.

🤖 Humanoid robot sales set to triple LINK

Sales of humanoid robots are on track to triple in 2026 to around 60,000 units, according to Smart Analytics Global, as the technology proves its worth in factories, warehouses and other industrial settings.

China dominates the market with over 97pc share, and Shanghai's Agibot has passed Unitree to become the world's biggest vendor at 44pc, with the two companies together making three-quarters of global sales.

Industrial and commercial uses drove more than 70pc of shipments in the first half of the year, up from 50pc, while home-service robots like the ones Meta is chasing are unlikely to hit mass scale within five years.

📷 iOS 27 may verify iPhone photos LINK

Apple may add a photo authentication tool in iOS 27, called Apple Reference Image, that confirms whether a picture was actually taken with an iPhone camera, according to signals 9to5Mac spotted in the latest beta.

To verify a photo, select sensor information and metadata get sent to Apple's Private Cloud Compute, and a new Reference mode has to be turned on in the iPhone's camera before the option becomes available.

The feature could complement watermarking efforts like Apple's Image Playground and Google's SynthID, though because it is still in beta testing, Reference Image might not appear when iOS 27 rolls out this fall.

🎵 Spotify will label AI artist profiles LINK

Spotify said it will attach "AI Persona" tags to profiles of artists whose identity is AI-generated, and keep their music out of both its editorial and algorithmic recommendations starting in mid-September.

The badges will show up on artist profiles, in Search, and on track rows in playlists, and Spotify won't rely only on self-disclosure, reviewing the most-listened-to profiles first for photorealistic AI identities.

Artists can start self-disclosing through Spotify for Artists on August 11th, and those wrongly labeled can appeal, while a coming tool will let users report unlabeled AI Persona profiles.

Other news you might like

• US court rules Meta, other tech firms must face thousands of lawsuits over social media addictionLINK

• Unreleased iPhones again revealed in the latest iOS developer betaLINK

• YouTube updates channel monetisation rules after 8 years: What's changingLINK

• Unreleased Claude model makes breakthrough on century-old Riemann Hypothesis math problemLINK

• Key Republican urges US to stop any advanced chips from reaching sanctioned Chinese firmsLINK

• Microsoft plans to ramp up AI chip output, targets 300,000 units in 2027LINK

• OpenAI is hiring a power trader, because electricity is now a position on its booksLINK

• The FCC wants to ban drones with LiDAR that it previously approvedLINK

• Meta glasses banned from courts in England and WalesLINK

• Nvidia’s Risky BusinessLINK

🧰 Trending tools

SoloUno: a gamified self-help app that helps people manage BFRBs like hair pulling and skin picking through daily challenges, streaks, and trigger logging.LINK

ConferenceGrid: a searchable database linking companies to the conferences they sponsor, exhibit at, or speak at, covering 6,000+ events and 62,000 speakers.LINK

radiusHQ: consolidates booking, pricing, calendar, payments, and reminders into one page so clients book in 30 seconds and teams avoid per-seat pricing.LINK

Good Assistant 2: an AI assistant that tracks your goals, tasks, and notes with persistent memory, sending reminders and proactive nudges to keep you on track.LINK

Workflo: automatically arranges your Mac windows before calls or focus sessions and restores layouts when displays change, using only Accessibility permissions-no screen reading.LINK

Papaya: a private, on-device app that turns shared moments into playful illustrated recaps for couples, with no accounts, cloud storage, or saved audio.LINK

📚 Trending papers & reports

Data grouping algorithms can be built exactly inside a transformer's internal circuitry, and training that circuitry produces a version that sorts data into groups better than the standard textbook method.LINK

Task-switching memory loss in AI systems learning new tasks back-to-back can be reduced by briefly boosting how strongly the model updates itself at the switch point, cutting the temporary accuracy drop while keeping overall performance competitive.LINK

Student performance predictors had an accuracy bug letting answer keys leak into their own training data, and fixing it plus adding a "how recently was this skill practiced" signal consistently boosted prediction accuracy across models.LINK

Untrained memory networks gain sharply better long-term recall by stacking layers with built-in shortcut connections, beating standard fast-to-train pattern-recognition networks on time series tasks without added training cost.LINK

Multi-camera tracking system stitches feeds from several cameras into one accurate map of moving people, hitting ~96% identity accuracy and lifting location precision from 75% to ~89%, while rivals fail completely under camera miscalibration.LINK

See you tomorrow for a new dose of ☕️ Techpresso!