Hello from The AI Night,
Today in AI:
OpenAI Gpt-5.2 Co-Authored a Real Breakthrough in Physics
Moonshot AI Intoduces Kimi Claw into its Browser Platform
Alibaba Launches a 397B-parameter Open Weight Model in Qwen-3.5 Series
Image Source: OpenAI blog
Here's the deal: OpenAI published a preprint on arXiv showing that a class of gluon scattering amplitudes, long assumed to be zero, is actually nonzero under specific conditions. The paper was co-authored by physicists from the Institute for Advanced Study, Harvard, Cambridge and Vanderbilt alongside OpenAI.
The Breakdown:
Standard textbook arguments say "single-minus" gluon tree amplitudes vanish. The paper identifies a precise kinematic regime, called the half collinear regime, where they do not
Human authors computed amplitudes by hand up to n=6 gluons, producing super exponentially complex expressions
GPT-5.2 Pro simplified those expressions, spotted a pattern, and conjectured a closed-form formula valid for all n
An internal scaffolded version of GPT-5.2 independently derived the same formula and produced a formal proof over roughly 12 hours of reasoning
The result was verified against the Berends Giele recursion relation and the soft theorem
Extensions from gluons to gravitons are already underway using GPT-5.2
The bigger picture: This is a concrete case of an AI system generating and formally proving a novel scientific result, not just assisting with computation. It establishes a replicable template for AI-human collaboration in theoretical research, with peer review and further publications expected.
Image Source: Kimi AI tweet
Here's the deal: Moonshot AI integrated OpenClaw, the open source AI assistant framework with personality and long term memory, directly into kimi.com. Users can now deploy, customize and run persistent AI agents from their browser without managing servers or APIs.
The Breakdown:
One-click cloud deployment of OpenClaw agents, powered by the Kimi K2.5 Thinking model, with no command line setup or separate API keys required
Access to ClawHub, a library of 5,000+ community built skills that agents can discover, install, and chain together inside conversations
40GB cloud storage per user and built in pro grade search pulling live data from sources like Yahoo Finance
Users can connect their own self-hosted OpenClaw instances to Kimi via a plugin, or bridge agents to external apps like Telegram group chats
Agents support scheduled tasks, custom personas and structured output rules, turning them into automated workflows rather than simple chatbots
Currently in beta for Allegretto (paid) members and above, available on web and mobile
The bigger picture: This positions Kimi as a managed hosting layer for autonomous agents, lowering the barrier from "deploy your own infrastructure" to "click and configure." For builders already using OpenClaw, the Telegram bridge and skill chaining create a path toward always on, multi channel AI assistants without stitching together separate tools.
Image Source: Qwen tweet
Here's the deal: Alibaba's Qwen team released Qwen3.5-397B-A17B, a natively multimodal vision language model built on a hybrid architecture combining Gated Delta Networks (linear attention) with sparse MOE. Despite its 397B total parameters, only 17B activate per inference pass.
The Breakdown:
Decoding throughput is 8.6x faster than Qwen3-Max at 32k context and 19.0x at 256k context, while matching performance
Natively fuses text and vision through early multimodal integration, processing up to 1M tokens (roughly two hours of video)
Competitive with GPT-5.2, Claude 4.5 Opus, and Gemini-3 Pro across reasoning, coding, agent and multimodal benchmarks
Tops several vision benchmarks including MathVision (88.6), CountBench (97.2), and OCRBench (93.1)
Language and dialect support expanded from 119 to 201, with a larger 250k vocabulary boosting encoding efficiency 10-60% across most languages
Post-training gains driven by scaling RL across diverse environments rather than optimizing narrow benchmarks
The bigger picture: Open-weight models matching frontier closed models at a fraction of the inference cost reshapes what builders can self host. The native multimodal agent capabilities, from GUI automation to spatial reasoning, signal a shift toward deploying general purpose agents without relying on proprietary APIs.
What else you need to know:
OpenAI shipped an internal product containing one million lines of code written entirely by Codex agents over five months, estimating development took one-tenth the time of manual coding.
Anthropic opened its Bengaluru office as its second Asia location and launched partnerships across enterprise, education and agriculture in India, its second largest Claude.ai market.
The Pentagon is threatening to end its up to $200 million contract with Anthropic over the company's refusal to allow military use for mass surveillance and fully autonomous weapons.
Cline released CLI 2.0, an open-source terminal AI coding agent with parallel agents, headless CI/CD pipelines and ACP support, plus temporary free access to Minimax M2.5 and Kimi K2.5.
ElevenLabs launched a government focused division offering multilingual voice and chat agents for public sector services, already deployed in Ukraine, Czech Republic and Midland, Texas handling thousands of daily citizen calls.
That’s it for today’s edition of The AI Night.
Our goal is to cut through the noise, surface what actually changed, and explain why it matters.
If this was useful, you’ll get the same signal here tomorrow.