AI designs the signal a bionic eye sends, a pattern that surveillance cameras cannot see, China opens a review into Palo Alto Networks, ByteDance goes for 10 trillion parameters, and tokenized assets triple while DeFi shrinks
Hello, dear TEA-mates! Here is what you need to know today.
1. 👁️ AI Designs the Signal a Bionic Eye Sends to the Brain
Researchers from UC Santa Barbara, ETH Zurich, and Miguel Hernández University used a deep neural network to design patterns of electrical stimulation for a 96-channel electrode array implanted in the visual cortex of a 27-year-old man who lost his sight after a traumatic brain injury. The implant, tested at Hospital IMED Elche in Spain as part of a broader feasibility trial, could both stimulate neurons and record how they responded, so the model trained on what the brain actually did rather than only on the settings sent to the device. The AI-designed patterns reproduced targeted brain activity more accurately and with less electrical current than other approaches, and the recorded neural response predicted what the participant perceived, spots of light called phosphenes, better than the stimulation settings alone. The model also read resting brain activity before each test and adjusted for it. UCSB associate professor Michael Beyeler, whose related work is backed by a five-year, $2 million NIH Director's New Innovator Award, said a useful prosthesis cannot rely on a fixed recipe and has to adapt to the individual. The proof-of-concept study was published in Neuron on August 7, 2026. (Read More)
🫖 TEA For Thought: "How great is the future going to be, having all of these! AI might bring disruptions in lots of aspects of our life, but it will also bring healing and longevity in ways that are impossible to imagine."
2. 🕶️ 31 Million Tests Later, a Pattern Cameras Cannot See
Security professional Bill Swearingen ran some 31 million tests over the past year to train a reinforcement learning model that generates adversarial patterns which stop surveillance cameras from detecting whatever the pattern covers. His project, noRecognition, defeated all 11 open-source detection algorithms he tested, including the software powering Flock license plate readers, Axon body-worn cameras, and cameras running Clearview AI. The patterns do not block cameras from recording. They scramble the object, person, and face detection that fires automated alerts, which turns a tracked subject back into a needle in a haystack. Swearingen ran the first public test on Friday at Def Con in Las Vegas, covering a 2009 Toyota Yaris with a pattern and beating a Flock camera, though he said the wheels were a challenge. His model now produces a new batch of patterns every minute, each mathematically better than the last, and he is keeping his strongest patterns off the internet so camera makers cannot train against them. A Kickstarter campaign is funding early merchandise, from T-shirts and hoodies to eventual vehicle wraps. (Read More)
🫖 TEA For Thought: "Where there is oppression, there will be protest, or rebels. After all, the pursuit of freedom is an innate and basic human need."
3. 🛡️ China Opens a Cybersecurity Review Into Palo Alto Networks
The Cyberspace Administration of China said on August 6, 2026 that it had launched a formal cybersecurity review of products sold in China by Palo Alto Networks, citing a pair of national security laws and the need to protect critical infrastructure. It is reported to be the first national security investigation opened under China's revised Foreign Trade Law. The move follows earlier 2026 restrictions in which Beijing barred cybersecurity products from a set of top US and Israeli firms, Palo Alto among them, and it starts the same legal process that ended in an enforceable ban on Micron Technology. A review of this kind can limit a company's ability to sell into critical sectors while it is under way, so the process bites regardless of what it eventually finds. The commercial exposure is contained, since Asia-Pacific and Japan account for roughly 12% of Palo Alto's global revenue and China is a small part of that. Washington, meanwhile, has been drafting its own restrictions on Chinese equipment in data centers. (Read More)
🫖 TEA For Thought: "More like a political show than what it seems. A thief crying 'Stop thief!'"
4. 🧠 ByteDance Is Pre-Training a 10-Trillion-Parameter Model
ByteDance is in the early stages of training a super-large model with an estimated parameter count around 10 trillion, according to Financial Times reporting relayed by ChainCatcher and sourced to three people familiar with the effort. That is roughly three times the size of Kimi K3 from Moonshot AI, currently the largest model released in China, and would put ByteDance in range of Anthropic's most advanced systems. The model is in pre-training, a phase that typically runs three to six months, after which it would move to fine-tuning and eventual release. The exact scale has not been finalized. Anthropic does not disclose parameter counts for its models, though industry estimates place its most advanced Mythos 5 near 8 trillion parameters and Fable 5 near 5 trillion. Parameter count sets a model's capacity ceiling, but real-world performance also depends on data quality and training methodology. (Read More)
🫖 TEA For Thought: "Well, the competition is only going to get more and more severe."
5. 🪙 Tokenized Real-World Assets Tripled While DeFi Shrank
Real-world asset deposits across lending platforms and decentralized exchanges more than tripled from $2.3 billion in Q2 2025 to $7.4 billion in Q2 2026, according to an August 6 report from CoinShares and Token Terminal covering one year of on-chain activity. Total DeFi deposits fell roughly 15% over the same stretch, a divergence CoinShares reads as demand driven by financial utility rather than crypto market cycles, a thesis it calls Hybrid Finance. Tokenized Treasury and multi-strategy funds such as JTRSY and BlackRock's BUIDL are the largest contributors, followed by private credit and delta-neutral yield strategies, since these assets keep generating income while serving as collateral. Ethereum hosts close to 70% of all RWA deposits. Spot trading told the same story: aggregate spot DEX volumes dropped about 70% while RWA spot volumes rose roughly 220%, concentrated in tokenized gold products including XAUT and PAXG. Current yields run from about 3.2% to 5.5%. (Read More)
🫖 TEA For Thought: "XAUT and PAXG pretty much occupy 90% of the market in tokenized commodities."
🛠️ Skill of the Day
The Exposure Map: turns a vague worry about how much of you is out there into a ranked, fixable list.
You are a privacy analyst who works with ordinary people, not
security specialists. I want an honest map of my digital exposure
and a short list of things actually worth doing about it.
Here is what I use and what I do:
Accounts and apps I use most: [LIST THEM]
Devices in my home: [LIST THEM]
Things I do in public or online that involve my identity:
[E.G. DRIVE A REGISTERED CAR, POST PHOTOS, USE A LOYALTY CARD]
What I am most uncomfortable about: [YOUR ANSWER, OR "NOT SURE"]
Give me:
1. A table of my exposures. Columns: what is collected, who most
likely holds it, how long it probably persists, and how much it
would matter if it leaked (low, medium, high).
2. The three exposures where my effort buys the most protection.
For each, one concrete step I can finish in under 30 minutes.
3. The exposures I cannot meaningfully reduce. Say so plainly
instead of suggesting theater.
Rank by real risk to me, not by how alarming it sounds. If a
common piece of privacy advice would not actually help my
situation, tell me that too.Paste into ChatGPT, Claude, or your tool of choice. Replace the bracketed bits with your own.
TEAHEE Moment
Stay sharp, stay informed. See you tomorrow.
If you enjoyed this TEA, follow along on social for more: