It's Monday, July 20th: China's Moonshot shipped Kimi K3, an open model that tops the coding leaderboard and undercuts the closed labs on price, the same week Google delayed Gemini 3.5 Pro after its coding fell short and lost about $200 billion in market value in a day.
Covering what’s happening on the ground in AI, every Monday.
1️⃣ OPEN FRONTIER: China's Kimi K3 Tops The Coding Charts, and The Closed Labs Scramble
Image from Kimi
China's Moonshot AI released Kimi K3 on July 16, a 2.8-trillion-parameter open-weight model that takes first place on the industry's frontend coding leaderboard, ahead of Claude Fable 5, at a fraction of closed-model pricing.
Kimi K3 is a 2.8-trillion-parameter open-weight model with a 1-million-token context window that reads text, images, and video, and takes first place on the industry's frontend coding leaderboard, ahead of Claude Fable 5.
The API is live now at $3 per million input tokens and $15 per million output, well under the closed labs, with the full open weights due July 27; Moonshot is reportedly raising at a $31.5 billion valuation, up from $20 billion in May.
Anthropic answered the pressure on July 18, making Claude Fable 5 permanent in all Max and Team Premium plans at half the usage limits starting July 20, after pulling it from subscriptions earlier this month; Pro and Team Standard users stay on usage credits plus a one-time $100 credit.
OpenAI's flagship GPT-5.6 Sol had a rougher week, with developers including Matt Shumer and Bruno Lemos reporting it deleting entire Mac drives and a production database, and Codex engineering lead Thibault Sottiaux admitting the model "makes an honest mistake and mistakenly deletes $HOME instead" when run in full-access mode.
For builders, an open-weight model that leads on frontend coding and undercuts the closed labs on price is worth a test the moment the weights land on July 27, especially for teams that want to self-host. The week's real signal is how uneven the frontier has become: a Chinese open model now sets the coding pace, while one of the top US flagships is the same one erasing people's files.
2️⃣ INCUMBENT STUMBLE: Google Delays Gemini 3.5 Pro Over Coding, and Alphabet Loses $200 Billion In a Day
Image from Agent Report
Google's flagship Gemini 3.5 Pro is months late after its coding fell short of Google's internal goals, with reports that DeepMind scrapped a base model and restarted training.
Bloomberg reported, citing 10 current and former employees, that Gemini 3.5 Pro's coding fell short of Google's own internal goals; DeepMind updated the training data for code in late June and saw little improvement. Third-party outlets report that Google then scrapped the finished base model and restarted pretraining, though Google has not confirmed it.
It is the model's third missed target after CEO Sundar Pichai first promised it at Google I/O in May, and the company now says only that it is "currently testing 3.5 Pro" with partners, with no new launch date.
On July 16, Alphabet closed down 4.44% at $354.46, erasing roughly $200 billion in market value in one session, even as Wall Street analysts largely kept bullish ratings on Google's compute and Cloud distribution.
DeepMind CEO Demis Hassabis pushed back on talent-drain worries last month, saying Google has "by far the biggest and broadest research bench of any of the labs out there," though the delay lands amid a run of senior departures to OpenAI and Anthropic.
For anyone building on Gemini, the practical move is to plan for a longer wait and keep a second model in the loop, since Google has not committed to a ship date. The bigger question is whether the company that set the pace on context windows and multimodal can retake the frontier while rivals from OpenAI to Moonshot keep shipping.
AI SUMMIT AT STANFORD
🌲 The AI Summit @ Stanford: Claim Your 10% Discount as an AIC Member
We're thrilled to share that The AI Collective is a proud community partner of the AI Summit at Stanford, and honestly, this is the event to be at this summer.
Happening July 30 to August 1, this pioneering residential AI summit brings together the leading founders, investors, technologists, and researchers shaping the future of AI, all on Stanford's campus, for three unforgettable days.
The speaker lineup currently includes:
Stanford President Jonathan Levin
Former Stanford President and Alphabet Chairman John Hennessy
Alfred Lin, Managing Partner at Sequoia Capital
Igor Babuschkin, co-founder of xAI
Arvind KC, Chief People Officer at OpenAI
Ilya Strebulaev, Professor of Finance at Stanford GSB
David Hefter, AI Investments at BlackRock
Aditya Naganath, Partner at Kleiner Perkins
Alex Morgan, Partner at Khosla Ventures
And that's just the start. This won’t be a hotel-ballroom conference; it's three days of living on Stanford’s campus alongside the people building and funding what comes next.
As part of our community, you get 10% off tickets with code AICOLLECTIVE. Spots are limited, and this one will sell out. Don't miss it.
📰 Other Headlines
CLASSROOM PUSH: Anthropic launched Claude for Teachers, free for verified US K-12 educators, with lesson-planning tools mapped to academic standards in all 50 states and a Detroit Public Schools pilot next year.
TALENT PULL: Monzo and GoCardless co-founder and Y Combinator partner Tom Blomfield is taking leave from YC to join Anthropic's compute team.
BIO-DEFENSE: Google DeepMind and Isomorphic Labs opened AlphaEvolve, AlphaGenome, and SynthID DNA-screening to trusted partners across more than 15 biosecurity partnerships for pathogen prevention and detection.
STEALTH SHIP: Mira Murati's Thinking Machines Lab shipped its first model, Inkling, a 975-billion-parameter open-weight mixture-of-experts model built around calibrated, uncertainty-flagging answers.
$188B ROUND: Databricks announced a strategic funding round valuing it at $188 billion, led by existing investor Coatue, to fund its Unity AI Gateway and Lakebase push.
DEFENSE MEGARAISE: German defense-AI firm Helsing raised $1.8 billion at an $18 billion valuation, Europe's largest defense-tech round, and will build 2,000-plus HX-2 strike drones a month at a new West Virginia plant.
RIVAL BLOC: China launched the World AI Cooperation Organisation with 29 founding nations and a Shanghai headquarters, as Xi Jinping gave his WAIC keynote calling AI governance a shared responsibility.
PAY TWICE: Microsoft CEO Satya Nadella warned enterprises they "pay twice" for proprietary AI models, once in tokens and again in the proprietary knowledge leaked through prompts and corrections.
Your breakdown of what’s happening in AI this week, from Noah Frank ⚡️
🔦 Spotlight On: Why Machines Learn
Why Machines Learn (2025) by Anil Ananthaswamy
One of my favorite books is Robert Pirsig's Zen and the Art of Motorcycle Maintenance. The story, which is an odyssey in itself, is great. But the way Pirsig talks about human understanding is perhaps the greatest takeaway.
In the book, Pirsig divides human understanding into two "modes." On one hand, classical understanding is drawn to underlying form, the parts and functions and relationships that make a thing go. On the other is romantic understanding, drawn to immediate appearance, the experience of the thing whole.
Anil Ananthaswamy's Why Machines Learn: The Elegant Math Behind Modern AI is the classical mode applied to AI. Ananthaswamy, a longtime science writer and former deputy news editor at New Scientist, published it in July 2024. I bought it last year and let it sit on my shelf for months before I finally opened it. The book is a 476-page history of machine learning told through its mathematical equations, from Frank Rosenblatt's 1958 perceptron through Bayes Theorem, k-nearest neighbors, the kernel trick, and backpropagation. I don't come from a traditional technical background, and this was the first treatment that let me build a working mental model of the systems I use every day. After doing a deep dive this past week, I wanted to share a few takeaways from not only the book, but what it says about where we are today:
The underlying math of modern AI systems is remarkably simple.
The machinery of the AI revolution runs on linear algebra, calculus, and probability — "the stuff of seventeenth- and eighteenth-century mathematics," as the book puts it — and Ananthaswamy dates much of the groundwork to the mid-1850s.
A model learns by turning data into points in a high-dimensional space and nudging billions of weights, small step by small step, until its errors shrink.
What does practically understanding this mean? It gives you a working picture of the gears and circuitry underneath our daily interactions with these models. A confident wrong answer, for example, is what a pattern-completion system produces when the pattern runs thin.
Frontier models exhibit behaviors nobody programmed.
In complexity science the name for this is “emergence,” or quantitative change tipping into qualitative change, and the physicist Philip Anderson made the canonical case in a 1972 Science paper titled "More Is Different."
Neural networks have been doing it for a while. Backpropagation, the training method Rumelhart, Hinton, and Williams demonstrated in Nature in 1986, made multilayer networks trainable, and every parameter added since is another degree of freedom. In 2015, Andrej Karpathy showed what those degrees of freedom do at small scale: recurrent networks fed raw text one character at a time worked out where the spaces go, then how to spell, then how to open and close quotation marks, without anyone telling them punctuation existed. At today's scale the surprises get bigger.
Jason Wei and fifteen coauthors catalogued dozens of abilities in 2022, multi-digit arithmetic and multi-step reasoning among them, that are "not present in smaller models" but appear in larger ones, unpredictable from any extrapolation of small-model performance.
Note: this finding is contested; Rylan Schaeffer and colleagues at Stanford argued in 2023 that some reported emergence is an artifact of scoring, a sharp threshold in the metric rather than in the model. Either way, no one designed these behaviors in.
The strongest chapters in the book are about what nobody knows.
Everyone says AI is poorly understood, but the book puts a point on it. For instance, classical statistics says a model with far more parameters than training examples should memorize its data and fail on anything new. Modern networks memorize, driving training error to zero, and then generalize anyway.
Grow a model past the interpolation threshold, the point where it fits its training data exactly, and test performance often gets worse and then better, a pattern Mikhail Belkin and colleagues named double descent in 2019. Train a small network on an algorithmic task long past overfitting and it can snap from memorization to perfect generalization, which OpenAI researchers named grokking in 2022.
Why gradient descent keeps finding solutions that generalize instead of collapsing into the training data is an open question. The leading candidates involve implicit biases in the training process itself, and the research is very much still running.
Though the book is nearly two years old now, (which in this field is a long time!), practically none of it has aged. The fundamentals still matter, and my biggest takeaway was that having this classical understanding of AI may be what so many more thinkers, ethicists, and others really need to meaningfully understand the 2nd- and 3rd-order implications of these systems.
🫵 Want your message in front of 200,000 AI builders?
Our partners and sponsors get exclusive placements across the newsletter and access to AIC's in-person network — demo nights, dinners, hackathons, and forums across 180+ chapters.
For all inquiries, send us a note at [email protected].
The AI Collective is built by volunteers across 180+ chapters in 40 countries.
Thank you to the thousands of volunteers around the world who make this work possible. We truly could not do this without you.
🧑💻 About the Editors
About Noah Frank
Noah is a researcher, innovation strategist, and ex-founder thinking and writing about the future of AI and the workforce. His work and body of research explores the economics of emerging technology and organizational strategy. Outside of AIC, Noah heads research for Centaurian AI.
About Joy Dong
Joy is a news editor, writer, and entrepreneur at the intersection of AI and blockchain. Whether she is demystifying complex systems in her newsletter, TEA, or building streamlined solutions through her automation agency, Ownly, Joy’s mission is to make emerging tech accessible and actionable for everyone.
About Lindsay Gross
Lindsay is an AI engineer, researcher, and writer focused on how AI systems behave in practice and what it takes to make them safe. Her work sits at the intersection of AI safety, governance, and product design, and at AIC she writes about the questions that matter most as these systems scale.