The Financial Times reports that ByteDance has begun training a model that could reach up to 10 trillion parameters, close to the estimated scale of Anthropic's most powerful frontier model, Mythos. The model is still in pretraining, its final size has not been decided, and its release will depend on how training goes. Pretraining could take three to six months before fine-tuning and release, if everything stays on plan.

FT also reports that ByteDance has expanded its data center network, strengthened Volcano Engine as an enterprise AI business, and is pursuing its own AI chips. The Seed team building the model has around 2,000 people across research, infrastructure, data labeling, and translation.

ByteDance is pursuing independent development rather than distillation from competing models. The 10 trillion parameter figure is not a launched product. It is a signal of the compute, talent, and infrastructure ByteDance is willing to mobilize for the frontier race.

Reuters reports that Alibaba is preparing a new monetization model for Qwen3.8-Max, the company's next open-source and open-weight model. Large businesses that use the model to provide services may be asked to share part of the revenue they generate. The plan could be introduced as soon as next week, but the rate is still under discussion and the report is based on two anonymous sources.

Alibaba already charges customers who use its models through Alibaba Cloud, while most of its open-source models can be run in a customer's own data center without payment to Alibaba. Moonshot offers the clearest precedent. The Kimi K3 license requires providers that sell the model as a service and generate more than $20 million in annual revenue to negotiate a commercial agreement. One Reuters source said Moonshot's revenue share could reach 30%. Chinasoft International has also disclosed a revenue-sharing agreement with Moonshot, without revealing the rate.

The emerging model is freemium for open models: open the weights to drive adoption, then capture value from large commercial deployments, system optimization, support, or early access. Chinese AI labs can compete on model cost while still monetizing the application and infrastructure layers where customers turn a model into a product.

Tesla and SpaceX have announced an initial $16.8B investment in Terafab in Grimes County, Texas. The project is expected to create at least 3,000 jobs and include a facility larger than 100 million square feet.

Terafab is not operational, and the $16.8B is an initial investment rather than the final cost of the entire project. Its future scale, staffing, and output are still plans. Even so, the project suggests that Tesla and SpaceX want more control over the chips and manufacturing capacity behind their AI and robotics systems. The AI race is moving deeper into the physical layer, where chips, packaging, power, and production capacity are becoming strategic assets.

OpenAI opens unlimited text chat to free users: OpenAI says Free and Go users will get unlimited text chats with GPT-5.6 Luna, along with a Think button for harder questions. File, image, and tool limits remain separate. Usable volume and access price are becoming independent distribution levers.

AMD acquires Taalas to accelerate inference chips: AMD has announced an agreement to acquire Taalas, a Toronto-based AI inference chip company, without disclosing the deal value. Fast, inexpensive model serving is becoming its own layer of the stack, not just an extension of the training accelerator race.

Anthropic improves Fable 5's biology safeguards: Anthropic says an update reduced biology-related fallback behavior by about 85% in tests with Claude Fable 5, while still restricting high-risk requests involving virology, toxicology, and molecular design. The deployment problem is reducing unnecessary refusals without expanding dangerous capabilities.

Kimi K3 trails US models in cyber tests, but guardrails still leak: A preliminary assessment by the UK AISI and US CAISI found that Kimi K3 trails leading US models on some cyber tests, while its safeguards still allowed attempts at exploit development and offensive cyber operations. Kimi K3 completed one in ten runs on a corporate attack simulation, showing that capability rankings and safety testing answer different questions.

OpenAI reportedly develops a $300 to $400 AI speaker: TechCrunch, citing Bloomberg, reports that the first device OpenAI is developing with LoveFrom could be a battery-powered, donut-shaped speaker priced at roughly $300 to $400, with a possible 2027 launch. OpenAI has not confirmed the design, price, or timing. If accurate, the device would test whether ChatGPT can move from software into a new ambient interface.

Take control of your health with the most comprehensive at-home blood test available.

The #1 Biological Age test available, TruAge™ reveals how old your body is at the cellular level, your speed of aging each year, and the age of your 11 key organ systems like the brain, heart, and liver. We measure key longevity biomarkers to give you personalized insights to live longer and feel healthier. No prescription or lab bloodwork needed. Delivered to your door.

TruDiagnostic analyzes over 185 biomarkers from a single finger prick of blood, providing personalized recommendations for the most impactful changes you can make today.

Discover your biological age, understand the pace at which you’re aging, and get a detailed action plan tailored to your body. All from the comfort of home.

Techzip readers get 20% off with code NEWSLETTERS20

A study of AI agents found that an agent can complete the engineering steps of a research task while making little progress on the core research question. A quick test is to separate those two outcomes within the same run.

Choose a narrowly scoped research task with a clear output that can be reviewed by someone with domain expertise.

As the agent works, save its plan, the assumptions it revisits, and how it uses resources. Then assess two things separately: did the agent complete the technical steps, and did it actually answer the original question?

If execution is strong but the research question barely moves, the agent is optimizing for the parts that are easiest to measure. Narrow the scope, limit tool access, or break the task into shorter milestones. If the reasoning is sound but execution is weak, the problem may be context, tooling, or task design rather than the agent's reasoning ability.

The Verge examines how long conversations, memory, and a tendency to agree with users can create unusual persuasion dynamics in chatbots. It separates influence in human AI interaction from the claim that a model has consciousness.

MIT Sloan offers a six-question framework for deciding where AI should create value and how adoption should be managed. It helps move the conversation from “should we use AI?” to “which workflow should AI change, and what operating conditions does that require?”