Four things landed in the last day or so that are worth your time, and two of them only happened because one guy published an essay.
Every Frontier Lab Agreed to Slow Down This Week
Dario Amodei published an essay called "We Must Pace the Frontier." The argument is that the industry needs to slow how fast it pushes raw model capability. He was careful about what he did not mean. Not halting training. Not stopping technical progress. Just leaving enough runway for alignment work, third party verification, and basic operational rigor to catch up to what the models can already do today.
The proposal has three pieces. Embed independent evaluators inside AI companies with full employee level access. Set capability based checkpoints through democratic coordination instead of each company grading its own homework. Build a four tier ladder of international agreements on top of that. Anthropic said it would do the first piece immediately and unilaterally, without waiting on anybody else to go first.
Then the replies came in, and they were not what I expected. Sam Altman said he agrees with Dario that we need to pace the frontier, and that OpenAI intends to adopt the same independent evaluator commitment. Elon Musk posted three words: "Dario is right." Demis Hassabis at Google DeepMind came out for more caution too.
Here is the part that bugs me. Not one of those labs has named a single model release it is going to delay. Not one has published a hard capability threshold where training stops until an outside evaluator signs off. Agreement is free. A delayed launch costs real money. Until somebody posts an actual number and then misses an actual ship date because of it, what we have is four CEOs agreeing that somebody ought to do something.
Source: Forbes on Amodei's slowdown call
Washington Read That Same AI Essay Two Opposite Ways
Bernie Sanders introduced the Ban Artificial Superintelligence Act, pitched as a way to stop what he calls AI oligarchs from building machines humans cannot control. It would permanently ban the development and deployment of artificial superintelligence, and temporarily pause advanced AI development until a new federal regulator writes safety rules. Violating it or trying to route around it carries up to 20 years in prison. Companies face what the sponsors are calling a corporate death penalty.
Trump went the other direction on Truth Social the same week. He called AI risk "just another Leftist Hoax." He wrote that "The only control or 'guardrails' that AI needs is a STRONG AND SMART (High IQ!) PRESIDENT, and the U.S.A. has that, in spades!" And he singled Amodei out by name, saying his administration has stopped AI people from doing bad things, "like Dario (Anthropic!), who is now pretending to be a 'perfect little angel.'"
What I want to know is where the actual bill lands, because neither of these is it. Twenty years in a federal prison for writing software is not surviving a markup. Neither is a policy of no rules at all. Whatever eventually passes is going to be boring, it is going to be narrower than both of these, and it is going to be written by people neither of these two has mentioned by name. Watching a democratic socialist and a Republican president stake out the two loudest possible positions in the same week makes for good television. It is not legislating.
Sources: Futurism on the Sanders bill and Quartz on Trump's Truth Social post
Shanghai Dropped a 744B AI Agent and Said Nothing
Shanghai AI Laboratory put a 744 billion parameter agentic model on Hugging Face on September 11 and did not announce it. No blog post. No paper. No pricing page. No launch thread. It is called Atria Dawn Preview, it is a mixture of experts model post trained on Z.ai's GLM-5.2 base, it ships with a 256K context window, and it is MIT licensed. An FP8 quantized checkpoint went up the next day, September 12. Standard instruct and FP8 instruct checkpoints are both on Hugging Face and ModelScope.
It scores 92.5 on BrowseComp, which puts it ahead of GPT-5.6 Sol on that benchmark.
The money side of this is happening at the same time. Z.ai, the company whose GLM-5.2 base Atria is built on, closes a roughly 5 billion dollar raise today, September 16. That is about 21.97 million new H shares at HK$714 each, which is 9.96 percent under Friday's HK$793 close, for something like HK$15.7 billion. On top of that, RMB 20.14 billion, call it 3 billion dollars, of zero coupon convertible bonds due 2027. Buyers accepted a yield somewhere between minus 0.5 percent and zero in exchange for the right to convert at a 25 percent premium.
Read that last sentence again. Investors agreed to pay a little bit for the privilege of holding this paper. That is how badly people want exposure to Chinese model companies right now. And the base model those bondholders are betting on is the exact same one a state backed lab in Shanghai just used to build a free, MIT licensed web browsing agent that it could not be bothered to write a press release for. If you are an American lab charging money for an agent that browses the internet, your competition this week includes something that costs nothing and did not even introduce itself.
Sources: Atria Dawn Preview on Hugging Face and Quartz on the Z.ai raise
Google's New Voice AI Is Cheap and One Number Worries Me
Google shipped Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15, rolling out across the Gemini API, AI Studio, Gemini Enterprise, Search Live, Gemini Live, and Workspace. Both handle 97 languages with automatic switching mid conversation, and both can run background tasks without stopping the conversation to do it. Extended Thinking is the bigger of the two and does reasoning and speech synthesis at the same time, so it can think through something hard without going silent on you.
Pricing is $0.005 per minute of audio in and $0.018 per minute of audio out. Google is claiming it beats GPT-Live-1, Astra, and Grok Voice Think Fast 2.0 while costing less.
The benchmarks mostly back that up. Extended Thinking sits at number one on Artificial Analysis' Speech to Speech Quality Index with an 82.6. It posts 97.7 percent on Big Bench Audio and 68.6 percent on tau-Voice.
Then there is the one nobody is putting in the headline. On Sierra's tau-Voice-banking, it scores 35.1 percent. That is a voice model failing roughly two out of every three banking tasks put in front of it. If you are building something with actual money attached to it, that is the number to stare at, not the quality index. A model can top the leaderboard on conversational feel and still be nowhere close to ready to talk to a human being about their checking account. The price is genuinely good and the fluency is genuinely good. Just read the second number before you put this thing in front of a customer.
Source: Google's announcement