Tech Brew Ride Home - A Different Kind Of AI Virus

Episode Date: August 7, 2026

A New Mexico judge ordered Meta to pay $942M over harms to kids, with a trillion-dollar trial next. ByteDance started pretraining a 10T-parameter model, AMD bought Taalas to etch weights into silicon, and OpenAI's 2027 gadget leaked. Links Meta Ordered to Pay $942 Million to Address Harm to Kids From Social Media (The Wall Street Journal) Sources: ByteDance is pretraining an AI model with up to 10T parameters, roughly 3x larger than Kimi K3 and larger than the 8T estimate for Anthropic's Mythos 5 (Financial Times) AMD acquires Toronto-based Taalas, which integrates model weights directly into silicon with the promise to boost inference performance, for an undisclosed sum (The Register) Sources: OpenAI's new device, slated for 2027, is a hockey puck-sized smart speaker with moving parts that help give it personality and will likely cost $300+ (Bloomberg) Longreads Scientists trained AI on genetic sequences to design viruses not found in nature, yielding 16 viable viruses that can infect bacteria but don't threaten humans (The New York Times) David Imel: you could be taking way better photos on your phone (The Verge) Subscribe to the ad-free feed.

Welcome to the TechBrewrite home for Friday, August 7th, 2026. I'm Brian McCullough today. A New Mexico judge ordered MEDA to pay $942 million over harm to kids with a trillion-dollar trial next up. Bite Dance started pre-training a 10 trillion parameter model. AMD bought Talas to etch weights into silicon, open AIs gadget leaked, and of course the weekend long read suggestions. Here's what you miss today in the world of tech. A judge in New Mexico has ordered Meta to pay $567 million and make changes to its products after finding its platforms helped create a public nuisance harming teens, quoting the journal. The judge said Thursday that META must create a new $560 million abatement fund in addition to paying $375 million in civil penalties that a jury previously ordered. The company must

also enact certain safety features such as limiting the amount of time underage users in the state spend on Facebook and Instagram, hiding by default the number of likes on photos for such people, and disclosing to users there the risks of its platforms. The creation of the fund aimed at rectifying harms caused by the Facebook parents' social media apps was, quote, necessary due to the wide-ranging impacts of the harm and the complex nature of the remedy. State Judge Brian Beidcheed said, the dollar amount of the abatement fund was slightly smaller than the 779 million attorneys for the state sought. Meta said it disagreed with the ruling and planned to appeal. We remain confident in our record of protecting teens online and will continue to defend ourselves against

claims that misrepresent the facts, a spokesman said. The Thursday ruling stems from a lawsuit brought by Attorney General Raul Torres, accusing the company of allowing young users to see sexually explicit content and be easily lured by predators on the apps. A jury found in March that met a willfully violated state consumer protection laws. New Mexico led the way in the courtroom, Torres said in a statement, now other states and other countries confronting the same crisis have a roadmap they can follow. Meta is facing thousands of lawsuits from state attorneys general, school districts, and individual plaintiffs that alleged the company gave priority to growth over the safety of its underage users. The New Mexico case was the first to test questions about whether social media

companies should be held responsible for content on their platforms. The next trial brought by for state attorneys general begins jury selection in Oakland, California next week. Meta said in a court filing that the states in that case are asking for more than $1 trillion in damages, end quote. The lead in the AI race is kind of a funny thing. Leeds seem to only last mere months. Quoting the FT, ByteDance is training an AI model that could approach the size of Anthropics' most cutting-edge mythos system as Chinese companies continue to narrow the gap with the top US labs. The Chinese tech giant is at an early stage of training a model with as many as 10 trillion parameters. Three times larger than Moonshots Kimi K3, the biggest Chinese model released to date,

according to three people with knowledge of the matter. The Bight Dance model is being pre-trained, a stage that typically takes three to six months before it is fine-tuned and released, if all goes well, one of the people said. The exact model size would only be determined at a later stage. Anthropic doesn't disclose the size of its models, but industry estimates say that its most advanced Mythos 5 has about 8 trillion parameters and Fable 5 about 5 trillion. While parameter count sets the fundamental capacity or memory limits for the models to store information, actual capacity also depends on other factors such as data quality and training methods. ByteDance's efforts to train one of the world's largest AI models show Chinese lab's ambition to not only catch up,

but outperform their U.S. peers in the most advanced level of AI. In the past weeks alone, Chinese models from Moonshot and Alibaba showed strong performance on benchmarks, lagging behind only Anthropics Fable 5 in certain areas. Methos 5, Anthropics' most advanced model, is only available to approved organizations after a temporary ban in June due to security concerns. Industry insiders say multiple Chinese labs are in the process of training models of the size of Fable 5, while BightDance is currently the most ambitious and pushing for the largest. BightDance, the parent of viral video platform TikTok, has kept a low profile in its AI development as its models are mostly closed, unlike many of its Chinese peers. Its latest C-Dance model ranks among the most advanced globally in video generation, while its flagship consumer-facing model, Dao Bao, is the most popular in China with 324 million monthly active users.

Over the past three years, ByteDance has invested in AI more aggressively than any of the other Chinese tech giants, building out its network of data centers and hiring researchers. It has doubled down on its cloud unit, Volcano Engine, which sells AI solutions to enterprises. bite dance also has ambitions to develop custom AI chips, end quote. And I continue to be surprised by the degree to which AI is changing the fundamental nature of the chip industry. Another case and point from the register, quote, in AMD's latest bid to upset Nvidia's dominance in AI hardware, the House of Zen has acquired AI chip company Talas, which bakes model weights directly into silicon in a process that promises to boost inference performance by an order of magnitude or more. The deal announced at market close on Thursday appears to be framed in much the same context

as NVIDIA's $20 billion licensing deal with GROC last December. Make high-performance premium inference services prized for AI agents like code assistance faster and cheaper to run. AMD didn't disclose the terms of the deal, but from what we understand, this is an actual acquisition rather than an aqua hire. Founded in 2023 and based in Toronto, Talas's approach to inference is radically different from conventional GPUs or the data flow architectures that underpin GROC LPUs or Cerebrus's wafer scale accelerators.

The startup chips don't rely on HBM to store the model weights, but rather etch them directly into the silicon. In a sense, Talas's chips are really model-specific integrated circuits or MSICs. Perhaps more importantly, Talis's tech isn't just conceptual. In February, the startup revealed its first test chip fabbed on TSM's 6-Nameter process tech, which it called the HC1. Initial benchmarks saw the chip serve meta's Lama 3.18B at a blistering 16,960 tokens per second. When announced last February, that was 48 times faster than Nvidia's GPUs and eight and a half times faster than Cerebrus's accelerators. While Lama 3.1 is ancient

by today's standards, having made its debut all the way back in mid-20204, the rectical-sized chip was really intended to prove the concept. TALIS has been incredibly secretive about how its chips actually work, but we know its processors are comprised of two main regions, the mask ROM recall fabric where model weights are etched, and the S-RAM recall fabric, where KV caches and fine-tuning adapters are stored. For its second-gen-hc2 chip due out this summer, Talis aims to boost parameter count to 20 billion parameters. That might not sound like much, but just like with GPUs for larger models, weights are simply distributed across multiple accelerators using pipeline parallelism. At 20 billion parameters per chip, you'd need just 50 accelerators to support a trillion parameter

model. And AMD just so happens to have a rack-scale compute platform, an in-house system design team that can comfortably accommodate that. That's quite a bit more space and power-efficient than Nvidia's recently unveiled LPX systems, which would need a few dozen GPUs and at least 2,000 GROC LPUs to serve the same model. While the tech is blazingly fast, if you you hadn't already figured it out, it comes with a pretty substantial downside, though. Once the chips are deployed, you're stuck with that model. Any change bigger than something like a Laura adapter is going to require a re-spin of the chips, which is not only expensive, but time-consuming. Nearly four years into the AI boom, new models are rolling out on a nearly monthly basis.

In order to benefit from Talas's tech, AMD's customers are going to have to be really sure about their choice of models, which will be easier for some than for others. However, if the startup is to be believed the situation isn't quite as bad as it sounds. While new models will require a respin, it doesn't require starting over from scratch. Instead, just two layers of metal need to be changed, which is a lot cheaper and less time-consuming, end quote. Mark Gurman has a scoop on AI's forthcoming first ever hardware device. Quote, a highly anticipated new device from OpenAI will have a unique look complete with moving parts that help give it personality and will likely cost more than $300, according to people familiar.

with the matter. The product, essentially a smart speaker without a display, will be shaped like a donut that's roughly the size of a hockey puck, said the people who ask not to be identified. The idea is to make the device easy to carry around the home with one hand. Open AI is looking to break new ground with the product, which is slated for release in 2027. The device will be positioned as an AI-first computer that can help users get things done, the people said, and the design should help it stand out from current smart speakers. Though Open AI eventually plans to offer a family of devices, the smart speaker is seen as a way to ease into the market. Other types of always-on AI gadgets such as smart glasses have raised concerns about privacy.

Open AI expects to rely on the smart speaker throughout the day. It will work similarly to the company's chat GPT voice mode on smartphone apps, but with more advanced models for human-like interactivity. The device is designed to learn more about a user over time, letting it tailor conversations and act more like a real person. The circle-shaped device will include parts that move on their own, according to the people that will help show when it's responding and interacting with the user. The goal is to make the object feel more alive than today's stationary speaker products.

The product will have speaker grills and microphones for fielding commands and conversing with users. The battery powered item will be designed to work in different positions in a user's hands, say, or placed on a nightstand or kitchen counter. Open AI is also planning to include lights on the device to demonstrate when it's listening and make interactions feel more personal. A camera system and other sensors, meanwhile, will perceive the surrounding environment and feed visual information to the AI. Over the long run, Open AI aims to build a device business

and someday offer something that could replace today's smartphones. The speaker product is designed in collaboration with Johnny Ives' love from studio. Ive, who also co-founded a device startup that OpenAIA acquired last year, made his name devising the look and feel of the iPhone, iPad, iPod, and Apple Watch. Like Ives' previous offerings, the new device will have a premium look and use upscale materials like high-quality metal. Open AI has discussed pricing of $3 to $400 per unit, which would be more than most smart speakers, but less than a standard iPhone, end quote.

In the long reads this week, this could, frankly, just be a regular news segment, but from the Times, read about how scientists trained AI on genetic sequences to design new viruses not found in nature, yielding 16 viable viruses that can infect bacteria, but apparently don't threaten humans. Quote, scientists at Stanford University and the ARC Institute, a research organization in Palo Alto, California, taught AI to recognize patterns of DNA structure in nature, and then to use that data to write recipes for entirely new viruses. The researchers followed those recipes to create DNA molecules, which they inserted into bacteria. The modified bacteria then produced viruses never seen in nature. The viruses were able to infect other

bacteria, demonstrating that they were viable. This is an important milestone. said Patrick Kai, a synthetic biologist at the University of Manchester, who was not involved in the study. The viruses dreamed up by AI do not pose a threat to humans because they are all similar to a naturally occurring virus called Phi X174, which can infect only bacteria. But the new study adds to growing worries that artificial intelligence might enable the creation of a new generation of biological weapons from deadly poisons to unstoppable pandemics. Dr. Moritz-Honke, a fellow at the Johns Hopkins Center for Health Security, who is not involved in the new study, said governments and scientific organizations have been slow to develop guardrails that could block the creation of a

deadly virus, even as the science races ahead. There's just a huge disconnect, he said. The authors of the study relied on an AI model called Evo, which is similar in some ways to chat GPT, made by OpenAI. DNA is strikingly similar to a book in some ways, a string of molecular building blocks known as nucleotides arrayed like letters and a line of text. A gene consists of hundreds of nucleotides drawn from a four-letter alphabet of A, C, G, and T. The sequence encodes the instructions for building proteins and other molecules. DNA has its own rules of grammar, and if a sequence violates them, the result is biological gibberish. Biologists have uncovered some of nature's grammatical rules, but many remain a mystery. The researchers wondered if Evo could pick up

these rules on its own. Instead of training it on text, they turned. trained it on genetic sequences, drawn from millions of animals, plants, microbes, and viruses. All told, Evo scanned about 9 trillion nucleotides. Evo eventually recognized patterns common across the tree of life and used them to generate blueprints for new genes encoding proteins that could perform specific jobs. These results led the team to wonder if Evo could master not just single genes but an entire genome. As the AI would be able to handle only small genomes at first, the scientist decided to try

to make viruses. while a human genome contains over 3 billion nucleotides, many viruses have genomes just a few thousand nucleotides long. It just felt like the obvious next step, said Samuel King, a graduate student at Stanford University and an author of the new study. He and his colleagues gave Evo another round of training, this time on the 11 genes of Phi X174 and about 15,000 of its closest relatives, end quote. And then over at the verge, David Immel says you could be taking way better photos on your

phone with just a little extra work. Quote, over the last few years, smartphone camera sensors have gotten bigger with better light-gathering capabilities and higher resolutions. But even though the sensors have gotten bigger and better, the straight out of the camera quality through default camera apps hasn't improved much, if at all. In some respects, smartphone photos have actually gotten worse, aesthetically, at least. That doesn't mean phone cameras are bad. The sensors and modern smartphones have actually gotten so good.

many of them can rival dedicated point-and-shoot options on the market today. Knowing how to utilize those sensors is the part that most people don't know how to do, and it really comes down to utilizing some third-party camera and editing apps, end quote. So click through for a crash course on how to get point-and-shoot quality photos from your phone. No bonus episodes for you this weekend. I am literally still digging my way out of boxes. Talk to you on Monday.