In 48 hours, China built a parallel AI world.
Two things happened in Shanghai last week, and they only make sense together. On July 16, a Beijing company called Moonshot AI released Kimi K3, and said it will publish the model's "weights" (the actual trained file) for anyone to download free by July 27. It is the largest open model ever released, at 2.8 trillion parameters. In blind developer testing on Arena's Frontend Code leaderboard it took first place, ahead of Anthropic's Claude Fable 5. Moonshot is upfront that its model still trails the best American systems overall, but on writing front-end code it is now the one to beat.
The next day, at the World Artificial Intelligence Conference in the same city, Xi Jinping launched WAICO, a new intergovernmental organization for AI rules with 29 founding countries including Russia, Brazil, Indonesia, Malaysia, South Africa, Senegal, and Pakistan. It will be headquartered in Shanghai. The United States is not a member. Xi's framing: "AI development should not be a solo performance by a single country, but a symphony of international cooperation."
For two years the story has been that America builds the best AI and sells you access to it. China's answer is to give a comparable model away and organize the countries that cannot afford to buy one. If you run a business, the price of capable AI is heading toward zero, because it is hard to charge for something a competitor posts for free. If you care about the rules, notice that there are now two rulemaking clubs: the one America leads and the one meeting in Shanghai. Most of the world will end up picking a side, and that choice will shape which AI ends up inside the products you use.
Kimi K3: 2.8T total parameters (mixture-of-experts), 1M-token context, native vision, always-on reasoning. Ranked #1 on Arena's Frontend Code Arena (1,679 pts), first in 6 of 7 front-end domains, ahead of Claude Fable 5; full open weights due Jul 27. Moonshot's own eval still places K3 behind Fable 5 and GPT-5.6 Sol on overall performance. WAICO: formed at WAIC Shanghai (Jul 16 to 17), 29 founding states, HQ Shanghai, US not a member; analysts read it as a bid to shape AI norms at UN level.
Google's flagship slipped again, and it cost about $200 billion Gemini 3.5 Pro
We told you in Issue #0 that Google missed its June deadline for Gemini 3.5 Pro, and in Issue #1 that the company answered with a smaller, faster model instead. Now the big one has been delayed by months. Google retrained part of it with more programming examples, did not like the results, and went back to work rather than ship something below its own bar. Alphabet's stock fell about 4.4% on the news, erasing roughly $200 billion in market value in a day.
Two useful lessons. Even the biggest company in this race can miss, repeatedly. And "our AI is a bit worse at coding than we hoped" is now a $200 billion sentence, which tells you how much of the stock market is riding on these launches.
Gemini 3.5 Pro missed its June target (announced at May I/O). Google added programming data to the training set; internal results fell short on coding and long-horizon reasoning, so it re-entered development rather than ship. Alphabet fell ~4.4% on Jul 16, ~$200B market cap, while rivals' recent models outpaced Google on code generation.
OpenAI's first gadget is a talking speaker that follows you around the house OpenAI hardware
Bloomberg reports OpenAI's long-rumored device is a portable, screenless smart speaker with cameras, sensors, and small moving parts meant to give it a sense of personality. It runs on GPT-Live, the voice technology that can listen and talk at the same time, and it is being designed with Jony Ive's studio and a group of former Apple engineers. OpenAI hopes to show it later this year and sell it in 2027.
This is the version of AI that stops living behind a screen and starts sitting on your kitchen counter. Worth deciding now how you feel about a camera and microphone that move on their own in your home, because the pitch will be that it is a companion, not a computer.
Screenless, battery-powered, portable; cameras + sensors + mechanical motion; runs GPT-Live full-duplex voice; developed with Jony Ive's LoveFrom and ex-Apple staff. Unveil targeted for 2026, sale 2027; timeline could be affected by Apple's pending trade-secret suit over the io hardware team.
OpenAI built an AI to attack its own AI GPT-Red
Its new system, GPT-Red, hunts for the trick that turns a helpful assistant into a rogue one: hidden instructions planted in a web page or email that the AI reads and obeys as if they came from you.
AI assistants are starting to run errands in your inbox and files. This is the flaw that could hijack them, and even OpenAI treats it as unsolved.
GPT-Red uses self-play to generate and surface prompt-injection vectors, feeding fixes back to harden GPT-5.6 Sol. Prompt injection remains the central unsolved risk for agentic assistants that read untrusted web/email content while holding your credentials.
Was your layoff list picked by software?
On July 13, twenty-six current and former Meta employees sued the company in federal court in Oakland, California. Their claim is that Meta's roughly 8,000-person layoff, about 10% of its workforce, was shaped by internal AI systems, keystroke and activity monitoring, dashboards tracking how much employees used AI tools, and algorithm-assisted performance rankings.
The heart of the complaint is a design problem rather than a robot villain. Those scores, the plaintiffs argue, are impossible to earn while you are out on protected leave. About half the plaintiffs had taken caregiving or medical leave, including eight women on maternity leave and four men on parental leave. The suit cites the Americans with Disabilities Act, the Family and Medical Leave Act, and the Pregnancy Discrimination Act, among others. Meta says the allegations "lack merit and are not based on facts. Workforce management and organizational decisions were and are made by people, not AI."
This is the first serious legal test of AI-assisted layoff decisions, and it lands on a question anyone with a job should care about: if a system measures your output while you are legally entitled to be absent, what exactly is it measuring? Expect the answer to shape how your own employer is allowed to use this software.
Filed N.D. Cal. (Oakland), Jul 13, 26 plaintiffs; targets the ~8,000-person (approx 10%) reduction that began May 2026, separations from Jul 22. Alleges keystroke/activity monitoring, AI token-usage dashboards, and algorithmic performance rankings that cannot accrue during protected leave. Claims under ADA, FMLA, Pregnancy Discrimination Act, Pregnant Workers Fairness Act, Title VII disparate impact, plus CA/NYC bias-testing laws. First major challenge to AI-assisted reduction-in-force decisions.
And the door into a first job keeps getting narrower Class of 2026
A new Class of 2026 hiring report found that 48% of hiring managers would rather put money into AI tools than hire and train a recent college graduate. Fifty-five percent of companies have moved at least part of their entry-level budget to AI, and 45% have reorganized so one experienced person plus AI does work that used to take several junior hires.
Source: ResumeTemplates Class of 2026 hiring report
We flagged the quiet freeze on beginner roles in Issue #0. This is the same story with the receipts. If you are early-career or advising someone who is, the useful move is showing you can run these tools rather than competing with them.
ResumeTemplates Class of 2026 report: 48% prefer investing in AI tools over hiring/training a grad; 55% shifted entry-level budget to AI; 45% restructured to senior+AI; 30% say AI reduced their need for grad hires; 23% will hire fewer 2026 grads or none. Separately, NY Fed data put 22-to-27-year-old grad unemployment at 5.6% vs 4.2% overall.
AI has started doing the science, not just reading it.
Stanford researchers published Biomni in the journal Science, an AI "co-scientist" that runs whole research workflows rather than answering one question at a time. It reviews the literature, proposes a hypothesis, pulls together the data, runs the analysis, writes the code, and shows its work. It draws on roughly 150 tools, 105 software packages, and 59 databases across 25 areas of biomedicine, from genomics to pharmacology.
Most AI news is about software that writes your email. This is software that helps design experiments, which is how new medicines eventually start. The honest caveat: a tool used by 10,000 labs can spread a mistake as efficiently as an insight, which is exactly why the "show your work" part matters.
Biomni (Stanford, Leskovec lab), published in Science (DOI 10.1126/science.adz4351). Autonomous agent spanning ~150 tools, 105 software packages, 59 databases across 25 biomedical subdomains; produces traceable, reproducible workflows with humans retaining scientific judgment; prototype used by 10,000+ labs. Nature Medicine ran a related research highlight on Jul 17.
“Should an industry that moves this fast be graded by a referee it helps pick?”
The debate around Demis Hassabis's proposed AI referee, this week
Demis Hassabis, the Nobel laureate who runs Google DeepMind, proposed on July 14 that the United States lead the creation of an independent standards body to test powerful AI before it reaches the public. His model is FINRA, the organization that polices stockbrokers: an industry-funded referee rather than a government agency, on the theory that a normal regulator cannot keep pace with how fast these systems change.
- Step 1A lab finishes a powerful new model.
- Step 2It hands the model to an independent body up to 30 days before release.
- Step 3Testers probe it: broken guardrails, deception, cyberattack and weapons help.
- Step 4Voluntary at first. Passing could later be required to launch in the U.S.
Notice who is proposing the referee, and notice the timing. In the same week that China set up its own rulemaking club in Shanghai, a top American lab boss asked Washington to organize one here. Critics point out that industry-funded oversight has a mixed record, and FINRA itself gets criticized for being soft on the firms that pay for it. The argument to follow is not whether AI gets checked, it is who holds the clipboard.
Hassabis proposed a US-led, FINRA-style self-regulatory organization on Jul 14. Labs submit frontier models voluntarily up to 30 days pre-release; evaluators test for guardrail bypass, deception, cyberattack capability, and biological/nuclear uplift; compliance could later become a prerequisite to launch for US users. Critique: industry-funded bodies (FINRA included) draw soft-touch complaints.
The AI trade had a wobble. Alphabet dropped about 4.4% on the Gemini delay. Nvidia, the chipmaker that has been the market's proxy for the whole AI boom, lost the title of world's most valuable company on July 17 after falling about 3.5%. Apple took the crown back at roughly $4.91 trillion.
Alphabet shed roughly $200B in a day, Apple retook No. 1 at $4.91T · Source: CNBC
Anthropic is still marching toward the stock market. Bankers have begun scheduling investor meetings for a listing that could come as soon as October, with Goldman Sachs, Morgan Stanley, and JPMorgan Chase involved. The company confidentially filed its IPO paperwork back on June 1 and was last valued privately near $965 billion.
of the $412.7 billion raised by U.S. startups this year went to AI.
About 86 cents of every venture dollar, an all-time concentration in one sector.
For the first time in a while, investors treated an AI stumble as a reason to sell. That is worth watching, because a boom where nothing is ever punished is a boom that eventually corrects hard. If you have a retirement account, you own more of this story than you probably chose to. (Transparency note: Human Terms is written with help from Claude, made by Anthropic, which appears in this section and in our lead. We report it the same way we would report anyone else.)
Alphabet ~-4.4% (~$200B) on the Gemini delay; Nvidia ~-3.5% on Jul 17, ceding most-valuable-company to Apple (~$4.91T vs ~$4.83T); Nvidia first crossed $5T in Oct 2025. Anthropic: confidential draft S-1 filed Jun 1; investor meetings scheduled for a listing as soon as October (Goldman Sachs, Morgan Stanley, JPMorgan); last private valuation ~$965B post Series H; ~$47B revenue run-rate. US VC H1 2026: $412.7B total, $355.9B (86%) to AI, 87.5% in rounds >= $100M (PitchBook).
New York became the first state to stop building. On July 14, Governor Kathy Hochul signed Executive Order 62, a statewide pause of up to one year on state environmental permits for new "hyperscale" data centers, the giant AI campuses that draw 50 megawatts of power or more. Two weeks ago we told you New Jersey and Oregon made data centers pay their own power bills. New York changed the question from who pays to whether they get built at all. Hochul said the state will "lead the way in creating the strongest standards in the nation for data center development, ensuring that when companies succeed because of New York, New Yorkers succeed too."
And next month, Europe starts making AI introduce itself. The provisions of the EU's AI Act arriving in August 2026 are the ones you will actually see: rules requiring AI-generated content to be labeled and chatbots to disclose that they are chatbots.
The rulebook is being written in three places at once now, and none of them is the U.S. Congress: state capitals like Albany, Brussels, and as of this week, Shanghai. Practically, that means what AI is allowed to do depends on where you live.
NY Executive Order 62 (Jul 14): up to one-year pause on state environmental permits for new hyperscale data centers (50MW+ per trade press; threshold not stated in the governor's release). DEC withholds discretionary permits not already deemed complete; DPS to draft a Generic Environmental Impact Statement and weigh a Grid Acceleration Fund; four hyperscale centers operating, dozens of applications pending. EU AI Act transparency duties (synthetic-content labeling, chatbot disclosure) apply from Aug 2026.
- Tomorrow, Jul 22Meta's layoff separations begin, with the AI-discrimination lawsuit already filed.
- Jul 27Moonshot says it will publish Kimi K3's full weights, the moment a frontier-class model becomes genuinely free to download.
- August 2026Europe's AI transparency rules kick in, so labels on AI-generated content start appearing.