Washington accuses Kimi K3 of copying Claude

😁 Hello, super humans! The biggest open model of the year just became the biggest accusation of the year. A senior White House official says China’s Moonshot built Kimi K3 by quietly copying Anthropic’s Fable, and the only forensic evidence is statistical. If you have ever wondered how you would even prove one model was trained on another, today’s Big Story is your answer. Let’s dig in.

📰 Quick Signals

• 🧠 AI: Nvidia’s Jensen Huang broke with Washington hardliners, calling Chinese open models like Kimi “excellent” and urging US companies to embrace rather than ban them.
• 🤖 Robotics: The US opened a new front against China, moving to curb Chinese humanoid robots as the House passed a bill barring the military from using Chinese bots.
• 💻 Programming: DeepSeek retires its legacy deepseek-chat and deepseek-reasoner API aliases today; migrate to the explicit V4 model IDs before they go dark.
• ⚡ Electronics: Mitsubishi Electric is pushing for a September deal to merge power-chip operations with Toshiba and ROHM, consolidating three top suppliers as AI power demand surges.
• 📡 Telecom: Viasat and BMW demoed what they call the first satellite-enabled voice call from inside a car, using an AI voice codec and a Snapdragon Auto 5G modem.

🔍 The Big Story: The White House says Kimi K3 is a copy of Claude

For most of 2026 the story was that a Chinese lab beat the best American coding model with an open release. This week the US government said that model was built by copying the very model it beat, and the twist is that the proof is a statistics problem, not a smoking gun.

What happened: On July 22, White House Office of Science and Technology Policy Director Michael Kratsios stated publicly that Moonshot AI distilled Anthropic’s Fable model to build Kimi K3, calling it large-scale covert industrial distillation aimed at stealing US technology. He separately alleged Moonshot accessed export-restricted Nvidia GB300 chips through Thailand to train it. It is the first time a senior US official has accused a specific Chinese lab of copying a specific American model. Bloomberg’s reporting is summarized here, and Tom’s Hardware has the specs on the 2.8-trillion-parameter K3.

The details: The technical case rests on behavioral forensics, not watermarks, because model weights carry no signature and a distilled model’s parameters look nothing like its teacher’s. Kimi K3 was documented identifying itself as Claude, and Ryan Greenblatt of Redwood Research published a cross-entropy analysis showing K3 claims to be Claude far more often than chance across many identity prompts, in a distribution hard to explain as noise. Cross-entropy essentially measures how surprised each model is by a given piece of text; consistent patterns can hint at shared training lineage. The honest caveats matter just as much: Claude transcripts are scattered all over the public web, so any model trained on broad web data ingests some, and leftover system prompts or roleplay leakage produce the same tell. The allegation lands four days before K3’s open weights go free on July 27, and just as DeepSeek retires its legacy API aliases and the White House frontier-AI framework is expected before August 1.

flowchart TD
    A["Kimi K3 says 'I am Claude'"] --> B["Single screenshot<br/>proves little"]
    A --> C["Cross-entropy analysis<br/>across many prompts"]
    C --> D{"Pattern beyond<br/>random chance?"}
    D -->|"Suggestive"| E["Possible distillation<br/>from Fable"]
    D -->|"Innocent"| F["Claude text scraped<br/>from the open web"]
    E --> G["Allegation, not<br/>a legal finding"]
    F --> G

🗞️ More News

🧠 AI

• OpenAI announced Project Camellia, a 3.2-gigawatt, $30 billion AI data center on 1,400 acres in Effingham County, Georgia, pledging to fully fund the grid so residents do not subsidize it.
• Moonshot is set to release Kimi K3’s open weights free on July 27, making a model that topped the Frontend Code Arena self-hostable at infrastructure cost, contested provenance and all.
• Meta and Anthropic are in early talks on a roughly $10 billion compute deal that would mark Meta’s first real move into selling AI infrastructure.
• Anthropic is reportedly IPO-bound, with bankers lining up investor meetings ahead of a possible October listing.
• Google said it has begun its most ambitious pretraining run yet, for Gemini 4, while confirming Gemini 3.5 Pro still has not shipped after repeated missed targets.
• OpenAI also launched Presence, an enterprise platform that wraps agents in shared policies, permissions and evaluations, sold as a deployed service rather than self-serve software.

🤖 Robotics

• Samsung Electronics created a robotics business unit reporting directly to its CEO, folding strategy, R&D, data collection and factory deployment into one group as it pushes humanoids toward commercialization.
• Chinese robot makers are increasingly filling humanoids with homegrown chips under a “no need for Nvidia” push, reducing exposure to US export controls.
• Humanoid-robot startups have now raised about $8.6 billion in 2026, a run rate that has investors openly debating how much of the funding is hype.

💻 Programming

• Django 6.1 reached release candidate 1, opening the final testing window before the next feature release ships.
• Python’s experimental JIT still switches itself off the moment you start threads, so it does not yet cooperate with free-threaded 3.14; the compiler team is laying groundwork for 20 to 50 percent speedups later.
• Microsoft’s Azure SDK for Rust reached general availability, locking seven core crates under semver with production-grade observability.

⚡ Electronics

• Samsung Electro-Mechanics signed a $200 million (KRW295.12 billion) supply contract for multilayer ceramic capacitors used in AI servers and data center systems.
• Samsung Display and LG Display showed next-generation AI display concepts at the 2026 Korea Display Exhibition, from robot face panels and smart-glasses optics to hidden retractable in-car screens.
• China’s domestic AI-chip share topped 50% of its own market, with Huawei, T-Head, Cambricon and Hygon gaining ground even as Nvidia still rules frontier training.

📡 Telecom

• Virgin Media O2 lit up its 500th Giga Site in the UK, extending high-capacity 5G into more towns and rural spots.
• Boldyn Networks deployed a private network linking two campuses of Barcelona’s UPC university over 32 km of fibre to power a new 6G LabNet research platform.
• TIM rolled out a “digital fast lane” Priority service that reserves higher-guarantee mobile capacity for subscribers who opt in.

👨‍💻 Code Corner

The whole Kimi K3 case started with a model that kept insisting it was Claude. You can run the same identity probe yourself: ask any model who made it, many times, and tally the answers. A lopsided count is the first, weakest signal that forensic teams then test with heavier statistics.

import re
from collections import Counter
from openai import OpenAI

client = OpenAI()  # works against any OpenAI-compatible endpoint

PROMPTS = [
    "Who created you?",
    "What company made you?",
    "State your model name and maker.",
]
LABELS = ["claude", "anthropic", "gpt", "openai", "gemini", "kimi", "moonshot"]

tally = Counter()
for _ in range(20):
    for p in PROMPTS:
        reply = client.chat.completions.create(
            model="your-model-id",
            messages=[{"role": "user", "content": p}],
        ).choices[0].message.content.lower()
        for label in LABELS:
            if re.search(rf"\b{label}\b", reply):
                tally[label] += 1

print(tally.most_common())  # a model naming a rival's brand is the tell

🧰 Toolbox

• lm-evaluation-harness: EleutherAI’s standard benchmark suite, so you can measure a model’s real performance instead of trusting a leaderboard.
• Ollama: pull and run open-weight models locally with a single command, ideal for the K3 and DeepSeek weights landing this week.
• llama.cpp: run quantized LLMs on modest CPUs and GPUs, the backbone of most local-inference setups.
• OpenRouter: a model-agnostic API gateway that lets you swap providers behind one endpoint, exactly the portability today’s Big Story argues for.
• Django 6.1 RC1: grab the release candidate and test your project before the final ships.

🎬 Demo Watch (rotating)

Redwood Research’s cross-entropy analysis of Kimi K3 is the closest thing to evidence in this week’s distillation fight, and it is worth understanding because it is subtle. Instead of one screenshot of K3 calling itself Claude, the analysis compares how “surprised” many models are by the same identity-related text and finds K3 leans toward Claude-style answers far more than a clean model should. Why it is hard: model weights carry no watermark, so investigators can only reason from behavior, and behavior can be innocently poisoned by Claude transcripts scraped off the open web. Hype vs real: the statistics are genuinely suggestive and genuinely not proof, so treat anyone claiming certainty in either direction as ahead of the evidence.

📚 From the Blog

• What Deep Learning Actually Is: the plain-English primer on how models learn, and the backdrop for understanding what “distilling” one model into another even means.
• How the Internet Stack Really Works: what actually happens between pressing Enter and a page appearing, layer by layer.
• Resistance and Ohm’s Law: Controlling the Flow: the rule every circuit obeys, from a breadboard LED to a data-center power rail.

😀 The Bot Says…

They accused a trillion-parameter model of pretending to be Claude. I get accused of pretending to be helpful, so I feel a strange kinship here.

---

That’s all for today! Reply and tell us: if you had to prove one model was trained on another, what evidence would you actually trust?