Latest / Neural Newscast / Z.ai GLM-5.2 and Tencent Hy3 Challenge US Models [Model Behavior]
Transcript
- [00:00] Nina Park: From Neural Newscast, this is Model Behavior, [00:03] Nina Park: AI-focused news and analysis on the models shaping our world. [00:11] Nina Park: Welcome to Model Behavior. I'm Nina Park. This is our daily technology briefing where we examine [00:18] Nina Park: how AI systems are built, deployed, and operated in professional environments. Today is July 7th, [00:25] Thatcher Collins: 2026. And I'm Thatcher Collins. Today we're analyzing two major model releases from China [00:33] Thatcher Collins: that are significantly undercutting the price of American frontier models, [00:37] Thatcher Collins: alongside a first-of-its-kind regulatory move in the state of Illinois. [00:41] Announcer: We begin with reports from The Atlantic regarding Z.ai's new model, GLM 5.2. [00:49] Announcer: It is being characterized as China's answer to Claude Code, the agentic assistant from Anthropic. [00:55] Announcer: Thatcher, while Vercel's CEO called its coding performance shocking, the real story is the price point. [01:02] Announcer: It is reportedly several times cheaper than top U.S. offerings, while leaping ahead of Google Gemini in several agentic benchmarks. [01:11] Thatcher Collins: Nina, this arrives just as American firms are starting to experience AI sticker shock. [01:17] Thatcher Collins: Reports indicate Uber exhausted its 2026 Anthropic budget in just a few months, and Citi reportedly had to limit access to high-cost models. [01:26] Thatcher Collins: If GLM 5.2 can perform agentic tasks, like writing complex code at a fraction of the cost, the economic pressure on OpenAI and Anthropic becomes much more intense. [01:38] Announcer: Market data seems to support that shift. [01:41] Announcer: Research from RAND shows that after cheap models launched in 2025, global web traffic to Chinese models jumped from 3% to 13% in just two months. [01:51] Announcer: On platforms like OpenRouter, Chinese models now represent six of the most popular options available. [01:56] Announcer: It suggests that for many developers, cost efficiency is beginning to outweigh the geographic origin of the model. [02:03] Thatcher Collins: There is a massive security question here, Nina. [02:06] Thatcher Collins: Using these models means potentially sending sensitive corporate data through Chinese infrastructure. [02:12] Announcer: While companies like Coinbase claim they've cut AI spending in half by defaulting to models like GLM 5.2, [02:18] Announcer: the geopolitical risk of data harvesting or federal regulation in the United States remains a significant barrier for wider enterprise adoption. [02:26] Announcer: Parallel to the Z.ai release, Tencent has launched Hanyan HY3. [02:31] Announcer: This is a 295 billion parameter model using a mixture of experts, or MOE, architecture. [02:38] Announcer: In a departure from the closed source trend of GPT 5.5 or Claude Opus, Tencent is releasing HY3 under an Apache 2.0 license, making it one of the most significant open source releases of the year. [02:51] Thatcher Collins: That MOE architecture is the key to its efficiency, Nina. [02:55] Thatcher Collins: Even with nearly 300 billion parameters, it only activates about 21 billion for any given prompt. [03:03] Thatcher Collins: Tencent is specifically positioning this for reliable AI agents in production. [03:08] Thatcher Collins: By giving developers an open-source model that rivals the reasoning of closed models, [03:13] Thatcher Collins: they are directly challenging the narrative that you must pay a premium for high-tier reasoning. [03:19] Announcer: Switching to policy, Illinois Governor J.B. Pritzker signed Senate Bill 315 yesterday. [03:26] Announcer: The Chicago Tribune reports this is the first law in the nation to require independent safety audits for large AI developers. [03:35] Announcer: The law applies to any developer with more than $500 million in annual gross revenue, [03:40] Announcer: forcing them to publish explanations of catastrophic risks and how those risks are managed. [03:46] Thatcher Collins: The timeline is also notable. While the audit requirements take effect in January 2028, the reporting rules are strict. [03:54] Thatcher Collins: Companies must report critical safety incidents to the state within 72 hours of discovery. [04:00] Thatcher Collins: Nina, this is much more aggressive than the voluntary frameworks we've seen at the federal level. [04:04] Thatcher Collins: Pritzker mentioned that where the federal government is slow to act, states must step in. [04:09] Announcer: Interestingly, both Anthropic and OpenAI supported the Illinois bill. [04:14] Announcer: They see it as codifying the safety practices they already claim to follow voluntarily. [04:20] Announcer: However, the law includes heavy fines, up to $1 million for a first violation and $3 million for subsequent ones, and creates internal whistleblower protections. [04:30] Announcer: It effectively moves AI safety from a trust-me model to a legally enforceable framework. [04:36] Thatcher Collins: It's a strategic move by Illinois to lead the regulatory conversation, [04:40] Thatcher Collins: especially with House Speaker Emanuel Welch critiquing a move-fast-and-break-things culture. [04:46] Thatcher Collins: With California and New York also considering major regulations, [04:50] Thatcher Collins: we are seeing the emergence of a state-led regulatory patchwork [04:53] Thatcher Collins: that could define how these models are deployed long before any federal consensus is reached. [04:58] Announcer: Thank you for listening to Model Behavior. [05:01] Announcer: You can find more details and our full archive at mb.neuralnewscast.com. [05:06] Announcer: Neural Newscast is AI-assisted, human-reviewed. [05:10] Announcer: View our AI transparency policy at neuralnewscast.com. [05:14] Nina Park: This has been Model Behavior on Neural Newscast. [05:17] Nina Park: Examining the systems behind the story.