😸 Kimi K2 Thinking = The First Good AI Creative Writer?

PLUS: A $1 TRILLION dollar AI sell-off?!

Welcome, humans.

This week’s shenanigans led to a $1T selloff in top AI stocks including NVIDIA, Microsoft, Palantir, and Oracle. Shall we count the headwinds that hit AI this week? They include:

Sam Altman’s reply on how OpenAI will fund $1.4T in commitments and the CFO’s government “backstop” walk‑back; fresh U.S. curbs blocking NVIDIA’s scaled‑down China chips and Beijing’s order that state‑funded data centers use domestic AI accelerators; the splashy rise of Moonshot’s Kimi K2 Thinking (more on that below); Meta’s $30B bond sale (plus a $27B private‑credit deal) and Oracle’s $18B project loan; big‑bank CEOs warning of a 10–15% correction; consumer sentiment sinking near 3½‑year lows; and Michael Burry’s new shorts against NVIDIA and Palantir.

The All-In pod squad had an interesting angle on the whole thing, which we break down here. Basically, this is a rational rebalancing after a months-long rally that needed to happen for a variety of reasons.

Let’s not forget: We’ve got ChatGPT 5.1, Grok 5 (or is it Grok 4.20?), and Gemini 3.0 on the horizon coming any day now, too…

Here’s what happened in AI today:

WE break down Moonshot AI's new open-source AI, Kimi K2 Thinking.
Microsoft formed MAI Superintelligence Team to develop domain-specific AI.
The U.S. tightened export controls preventing NVIDIA to sell AI chips to China.
Box CEO Aaron Levie predicted that within five years, 95% of AI agent usage will address tasks humans never did before.

Advertise in The Neuron here!

Kimi K2 Thinking, Explained (and Tested)

DEEP DIVE: Everything to know about Kimi K2 Thinking (and our hands-on demo)

Most AI writes like it's cramming for a test. Fast, efficient, technically correct… but kinda soulless?

Kimi K2 Thinking is different. It's an open-source reasoning model from China that can think through 300 steps straight without losing the plot, match (or beat) the best closed models, and cost just $4.6M to train. Oh, and it's shockingly good at creative writing.

In the deep dive we cover:

The performance: Outscored GPT-5 on expert-level reasoning (44.9% vs 41.7%) and demolished humans on web research tasks (60.2% vs 29.2%)
The economics: A $4.6M model competing with billion-dollar training runs…and winning?!
The creative writing edge: Generates 1,595 thinking tokens for a single sentence (vs 110 for DeepSeek), making it exceptional for novels and long-form content
How to actually use it: Via API for pennies ($0.01 per 2,000-word story), through OpenRouter, or run the 245GB compressed version locally
Our real-world test: We co-wrote a YA novel called "The Salt Circus"—and the AI actually revised itself, scrapped bad ideas, and showed genuine creative judgment
The bigger picture: China's open-source models are outperforming closed US models at a fraction of the cost, and it's happening fast

Fun fact: Those 1,595 thinking tokens K2 generated? They were just to write a single sentence about cheese. Meanwhile, I'm over here using exactly zero thinking tokens before hitting "send" on most Slack DMs.

Why this matters: AI researcher Nathan Lambert analyzed Moonshot AI's Kimi K2 Thinking release, arguing that Chinese labs like DeepSeek, Qwen, and Kimi release open models months faster than closed American labs (he estimates Anthropic takes longest), have closed the performance gap to roughly 4-6 months, and are capturing growing mindshare globally—while questioning whether closed models that aren't publicly available even matter to users anymore.

FROM OUR PARTNERS

Breaking: The Great Startup Debate: 996 vs. The 4-Day Work Week

Silicon Valley’s founders are split. Some swear by the 996 grind 12-hour days, 6 days a week to chase market dominance. Others claim the 4-day work week makes teams faster, more creative and saner.

On Nov 19, Rippling is bringing both sides to the (virtual) stage for a live, unfiltered debate moderated by Varun Rana. You decide who’s building the future of work.

🎥 Expect bold takes, live Q&A, and exclusive giveaways (Airbnb gift cards + “First Principles” hats). Promo terms apply*.

Prompt Tip of the Day

Want to learn AI to learn anything? We’ve shared a few examples in the past, but this post from Suhail is the best encapsulation of all the tips to learn in one single prompt:

“Help me understand this paper step by step. Go from high level (simple explanations) to incredibly low-level, detailed technical explanations until I understand it. Do not advance without confirming that I understand each step with a quiz question first.”

We suggest you save this is as a project on ChatGPT or Claude or Gem in Google Gemini, so anytime you want to learn, just go to that project and paste in the topic, hit enter, and boom.

Treats to Try

LLM Gateway routes your requests to 90+ models across OpenAI, Anthropic, Google, and other providers through one unified API while tracking usage, costs, and response times; multiple pricing tiers available: free self-hosted version, free tier with 5% fee on usage, or Pro plan at $50/month w/ free trial.
Everywhere lives in your operating system and sees what's on your screen in real-time—press a hotkey and it helps with anything visible without screenshots, copying, or switching apps (code).
Fire Your QA Today watches you test once via screen recording, then runs those tests automatically thousands of times.
Golf is a security firewall for enterprise MCP servers that blocks prompt injection attacks on your MCP servers before they reach AI agents or customers.
Ancher learns what topics you care about (even when you can't articulate them), then filters news and information to show only what actually matters for your work.
Dazl combines app generation with visual editing panels to build you apps through chat, then lets you edit them visually or in code.
NocoBase is an open-source no-code platform you host yourself that lets you build internal tools, CRMs, or business apps with drag-and-drop blocks while keeping full control over your data and code (think Airtable meets internal tool builder).

Around the Horn

*This is a parody of “We didn’t start the fire” but for the AI bubble. Kinda puts the past two years in perspective. What fun!!*

NVIDIA won’t be allowed to sell China its new B30A “scaled‑down” chips as the U.S. tightens export controls again, while China ordered state‑funded data centers to use only domestic AI chips, forcing early‑stage builds to remove foreign accelerators or cancel plans.
Microsoft AI formed the MAI Superintelligence Team led by Mustafa Suleyman to develop “Humanist Superintelligence”, rejecting the race toward unbounded AGI in favor of domain-specific AI systems that remain controllable while solving concrete problems.
Computer scientist Boaz Barak analyzed METR's research showing AI capabilities (task-completion horizons) doubling every 6 months: even at 9% annual automation—far below current 75% trends—models predict 10x productivity within two decades, potentially breaking the 2% GDP growth constant through the last 150 years of electricity, computers, and internet.
Box CEO Aaron Levie predicts in five years 95% of AI agent usage will address tasks humans we never did before rather than improving existing workflows; enterprise meetings reveal companies want to automate too-expensive or talent-constrained work like real estate analyzing every lease for trends, life sciences screening drug data for quality, financial services mining deals for monetization, or legal serving unprofitable segments.

FROM OUR PARTNERS

Ideas move fast; typing slows them down.

Wispr Flow flips the script by turning your speech into clean, final-draft writing across email, Slack, and docs. It matches your tone, handles punctuation and lists, and adapts to how you work on Mac, Windows, and iPhone.

No start-stop fixing, no reformatting, just thought-to-text that keeps pace with you. When writing stops being a bottleneck, work flows.

Give your hands a break ➜ start flowing for free today.

Sunday Special

So the Pope posted about AI this week, and despite his very reasonable message about “cultivating moral discernment” in the design of AI, the first thing folks noticed was the dreaded em-dash.

“Yo, did the Pope just tweet about AI… USING AI?!”

Now, here’s a few funnies found from around the web… and by “around the web”, we mean Sora 2… (should we call it the “world wide slop network?” The “slop-ter-net?” The Sam Altman extended “Slopiverse”? You tell us!)