Sadly, it’s been a slow news week (a holiday weekend in the U.S. and World Cup worldwide). So, today we’re publishing two new stories, plus four from the past few weeks that you may have missed. For example:
- Seedance 2.5, the movie industry’s favorite video generator
- Field-specific knowledge helps Claude code users most
- DiffusionGemma, Google’s fast, capable text generator
- Regular use of AI chatbots rises among Americans
But first:
Companies in China move to shut down anthropomorphic chatbots
ByteDance’s Doubao and Alibaba’s Qwen are shutting down their customizable AI agent features before new Chinese government rules take effect on July 15. Both platforms allowed users to create or customize AI assistants with distinct personas, skills, and speaking styles -- for example, turning general chatbots into named tutors, role-playing characters, or companions. Doubao will disable its agent feature on July 15 and make related data inaccessible and unrecoverable after October 15 (users retain read-only access to chat histories and configurations until then). Qwen is moving faster, shutting down humanlike agents on July 10 and broader agent services on July 15. The moves comply with China’s Interim Measures for AI-Based Anthropomorphic Interactive Services, which regulate AI that “simulates human personality traits, thinking patterns and communication styles to provide sustained emotional interaction.” The government cited risks including extremism, privacy leaks, mental health harm, and addiction. The rules carve out exceptions for customer service bots, educational tools, and workplace assistants that don’t involve sustained emotional bonding. (South China Morning Post)
Fullstack Code Arena tests models’ abilities to deploy web apps
Arena.ai expanded its Code Arena evaluation platform to support fullstack web development, moving beyond frontend prototyping to enable AI models to build complete applications with databases, user authentication, and third-party API integrations. The platform now tests models’ abilities to handle persistent data storage, sign-up flows, and multi-file backend code generation, all capabilities previously out of reach in a single-page frontend sandbox. New tools include a live dev server with hot reloading, bash command execution, web search access, and direct deployment to Vercel. The shift reflects a growing question in AI coding: the need for models that don’t just write static functions, but can architect and deploy real end-to-end applications. For AI labs, arenas like this could provide the high-fidelity test cases needed to evaluate and train next-generation coding models. (Arena.ai)
Bytedance video model can synthesize more references
Seedance unveiled version 2.5 of its AI video generator, an upgrade focused on two practical improvements: higher-resolution output and extended video duration. The update targets users who need sharper detail for product visuals, social ads, and portfolio work who don’t want to overhaul their existing workflow. Longer clips now support complete scenes—a reveal, camera move, or character beat—rather than cutting off mid-moment. The generator maintains Seedance’s familiar prompt-and-refine process but adds aspect ratio and duration controls tuned for specific platforms. For users working toward publishable video, the resolution bump and scene continuity improvements mean less upsampling and fewer stitched clips to manage. (Seedance)
Anthropic study suggests coding skills matter less than you’d think
Anthropic analyzed 400,000 Claude Code sessions between October 2025 and April 2026 and found that users’ domain expertise, and not programming ability, determines how much autonomous work the model performs. Users with deep field knowledge trigger action chains twice as long as novices (12 actions versus five) and receive five times the output per instruction. The labor split is striking: users make roughly 70 percent of planning decisions while Claude handles 80 percent of execution. Across law, accounting, design, and data analysis, success rates on coding tasks converge near software engineer levels, suggesting that knowing what you’re building matters far more than knowing how to code. Over the study period, debugging fell from 33 to 19 percent of sessions while higher-value work expanded: deployment and data analysis doubled, software operation grew from 14 to 21 percent, and the estimated economic value of an average session rose 27 percent. The pattern suggests agentic coding is less about replacing engineers than amplifying expertise wherever it already exists. (Anthropic)
Google turns to diffusion text generation for efficiency and speed
Google introduced DiffusionGemma, an experimental 26B Mixture of Experts model that abandons the sequential token-by-token generation of standard language models in favor of diffusion-based text generation. Instead of predicting words one at a time, DiffusionGemma generates entire 256-token blocks simultaneously, achieving over 1,000 tokens per second on an NVIDIA H100 GPU, up to four times faster than autoregressive models. The approach works by starting with random placeholder tokens and iteratively refining them across multiple passes, similar to how diffusion models generate images from noise. The 26B model activates only 3.8B parameters during inference, fitting within 18GB of VRAM on consumer GPUs when quantized, making it practical for local deployment. Google openly acknowledges the trade-off: Output quality is lower than standard Gemma 4, and the speed advantage only applies to single-user, low-concurrency inference; cloud deployments running many requests simultaneously get no benefit and may cost more. The model opens up new use cases like in-line code editing and tasks requiring bidirectional attention, such as Sudoku solving, where each token depends on future context. (Google)
Pew study shows chatbots on the rise
Nearly half of American adults now use AI chatbots, a 16-percentage-point jump from ChatGPT’s number in 2024, with one in four reaching for them daily, according to a Pew Research Center survey of 5,119 adults conducted in February 2026. ChatGPT dominates the field at 44 percent adoption, more than double 2023’s 18 percent. Gemini trails at 24 percent. People use chatbots primarily for search (42 percent) and work tasks (38 percent among employed adults), but smaller groups turn to them for image generation, medical advice, or emotional support. The adoption curve doesn’t extend to other AI devices at the same pace. About a third own smart speakers; far fewer have bought AI-enabled doorbells (18 percent) or thermostats (11 percent). Yet skepticism runs deep. Sixty percent read AI-generated search summaries without necessarily trusting the results. A majority of Americans believes AI poses data privacy risks and is advancing too quickly, and more predict it will harm them personally rather than help. (Pew Research Center)
Want to know more about what matters in AI right now?
Read the latest issue of The Batch for in-depth analysis of news and research.
Last week, Andrew talked about the challenges of navigating the noisy AI landscape, our “learners first” philosophy at DeepLearning.AI, and why we partner with experts to provide high-quality, technically deep courses.
“We have a simple mantra that drives the order for whose interests we prioritize: ‘Learners first, Partners second, Ourselves last.’ When you lead a team, you get to pick a small number of mantras to repeat ad infinitum, to the point where the whole team is sick of hearing them. For DeepLearning.AI, I decided to make this one of them.”
Read Andrew’s letter here.
Other top AI news and research covered in depth:
- OpenAI previewed three GPT-5.6 Models (Sol, Terra, and Luna), with a wider release coming soon.
- Sakana debuted dedicated orchestrator models, Fugu and Fugu-Ultra, that spawn Claude, Gemini, and GPT agents task by task.
- Microsoft revealed MAI-Thinking-1, a Claude Sonnet 4.6-sized reasoning model developed without distillation.
- Inside RoboReward, a family of vision-language reward models that train robots to take action.
A special offer for our community
In case you missed it, DeepLearning.AI launched our first-ever subscription plan for our entire course catalog! As a Pro Member, you’ll immediately enjoy access to:
- Nearly 200 AI short and long courses from Andrew Ng and industry experts
- Labs and quizzes to test your knowledge
- Projects to share with employers
- Certificates to testify to your new skills
- A community to help you advance at the speed of AI
Enroll now to lock in a year of full access for $25 per month paid upfront, or opt for month-to-month payments at just $30 per month. Both payment options begin with a one-week free trial. Explore Pro’s benefits and start building today!
Data Points is produced by human editors with AI assistance.