Compact Conversations for November 15, 2026

Compact Conversations for 2026-11-15: 6 AI stories, ai news worth knowing in just 5 minutes.

[Audio embed placeholder]

The Lead: Meituan releases a 1.6-trillion-parameter agent model that never touched an Nvidia GPU

Meituan released the weights for LongCat-2.5, a 1.6-trillion-parameter mixture-of-experts model built for AI agents and trained entirely on Chinese hardware. It scores 88.4 on Terminal Bench, close to Claude Opus 5 at 89.1. A sample deployment runs on 194 Huawei Ascend chips with no Nvidia guidance. The model is available under the MIT license and supports OpenAI and Anthropic API formats.

Why it matters: This shows a major Chinese company can train a competitive frontier model without any Nvidia hardware, which has implications for global AI supply chains and hardware procurement.

Source: The Decoder

The Feed

Gemini 4 Carbon takes the MathArena crown but costs 15 times more than GPT-6.1 Sol

Google’s Gemini 4 Carbon topped MathArena, knocking OpenAI’s GPT-6.1 Sol from first place for the first time in a while. Anthropic’s Fable 5.2 came second. However, with the reasoning settings used in testing, Carbon costs about 15 times more than GPT-6.1 Sol, and Fable 5.2 about 40 times more. Carbon’s most common failure is exceeding the time limit, and it almost never gives up on a task.

Why it matters: The cost gap highlights the trade-off between raw benchmark performance and practical affordability for enterprise math and reasoning tasks.

Source: The Decoder

Anthropic says outdated prompting habits are slowing down Claude Opus 5.5

Anthropic made Claude Opus 5.5 the default on all paid plans and published new prompting guidance. Key advice: drop phrases like ‘think hard’ because the model reasons on its own, start at Medium effort (which matches Opus 5 on High in internal tests), and describe the desired result in a single message rather than step-by-step instructions.

Why it matters: Teams paying for Opus 5.5 may be wasting money with outdated prompts; following this guidance could improve speed and reduce costs.

Source: The Decoder

Beware the New AI-Fueled Ransomware Boom

Researchers report a 20 to 40 percent increase in ransomware attacks since spring, with 111 distinct groups claiming attacks between July and September. AI tools are being used for phishing, negotiation, and code checking, lowering the barrier to entry. One new group run by an inexperienced teenager made over $800,000 in under two weeks.

Why it matters: The fragmentation of ransomware groups and AI-enabled tactics mean more potential threats for enterprises, even if individual attacks are smaller.

Source: Wired

US and allies sign accord to protect AI supply chains from China

The US and its allies signed an accord to protect AI supply chains from China, ahead of President Trump’s meeting with Xi Jinping at the APEC forum in Shenzhen.

Why it matters: This agreement could affect procurement and vendor risk for enterprises relying on AI hardware and components.

Source: Financial Times

Hackathon project for MariaDB’s Wikipedia Through Time challenge: matching sentences across seven Wikipedia editions at any past date

A team at the Junction hackathon built Parallax, a tool that lines up Wikipedia articles across seven language editions statement by statement for any past date. It stores over 875,000 revisions, uses sentence embeddings and an LLM judge to flag where editions disagree, and can also check chatbot answers against Wikipedia history.

Why it matters: This demonstrates a practical use of embeddings and LLM judges for cross-lingual fact-checking and temporal comparison.

Source: 135.181.63.238.sslip.io

One Thing to Try

Claude Code v2.1.323 introduces Code Mode, an experimental feature where Claude writes JavaScript to make several tool calls in one step. This is useful for changes that repeat across many files, searches that need filtering before reading, or several independent commands that can run concurrently. To try it, start Claude Code with Code Mode enabled, confirm, and give it a task with many tool calls. Note: it’s terminal-only, requires a Claude subscription, and code stops after ten minutes unless run in the background.

Sources

Transcript

Host A: Welcome to Compact Conversations, the show that compresses the day’s AI news into 5 minutes.

Host A: [conversational] It’s the weekend update for Sunday, November 15th, so we’re covering the last couple days. Today’s lead: Meituan, the Chinese food delivery giant, has released the weights for LongCat 2.5, a 1.6 trillion parameter model built for AI agents, and the company says it was trained entirely on Chinese hardware. On Terminal Bench, a benchmark that measures how well agents handle command line tasks, it scores 88.4 points. Meituan’s own table puts Claude Opus 5 at 89.1, so it’s close behind. Host B: [curious] And the hardware detail is the interesting part. The self-hosting guide describes a sample deployment on 194 Huawei Ascend accelerators, and there’s no guidance for Nvidia GPUs at all. The model is a mixture-of-experts design, so only about 48 billion of those parameters are active per token, and it’s out under the MIT license. [with a small lift] The API even supports both the OpenAI and Anthropic formats, so it plugs into agent harnesses like Claude Code.

Host B: One number to know today: 111. That’s how many distinct ransomware groups researchers tracked claiming attacks between July and September, according to a Wired report over the weekend. [thoughtful] The ecosystem is fragmenting, with more groups but smaller average hauls per incident. One new group run by an inexperienced teenager reportedly made over 800,000 dollars in under two weeks, which shows how low the barrier to entry has become.

Host A: In other news, Google’s Gemini 4 Carbon has taken the top spot on MathArena, an AI math benchmark, knocking OpenAI’s GPT 6.1 Sol out of first place for the first time in a while. Anthropic’s Fable 5.2 came in second. Host B: [skeptical] The catch is cost. With the reasoning settings used in testing, Carbon runs roughly 15 times the cost of GPT 6.1 Sol, and Fable 5.2 about 40 times. The reporting also notes Carbon’s most common failure is simply exceeding the time limit, and it almost never gives up. In one case it launched a script that ran for more than eleven hours without checking on it. Host A: Next, Anthropic has some prompting advice for Claude Opus 5.5, which is now the default model on all paid plans. The company says old habits are holding it back, starting with phrases like think hard or think carefully. The model reasons on its own now, and in Anthropic’s tests, responses without those phrases started faster with no quality drop. [with emphasis] They also recommend starting at Medium effort, saying it matched the previous Opus at High in internal tests. Host B: [conversational] On the policy side, the Financial Times reports the US and its allies have signed an accord to protect AI supply chains from China. The pact comes ahead of President Trump’s meeting with Xi Jinping at the APEC forum in Shenzhen. Host A: [lighter] And one from the hackathon circuit: a team at Junction in Espoo built Parallax, a tool that lines up Wikipedia articles across seven language editions, statement by statement, at any past date. It stores over 875,000 revisions and uses an LLM judge to flag where editions disagree.

Host A: [conversational] One Thing to Try: Claude Code’s new experimental Code Mode. Instead of one tool call per turn, Claude writes a bit of JavaScript that makes several tool calls in a single step. So a change that repeats across many files, or a search that needs filtering before Claude reads the results, happens in one pass instead of a long back-and-forth. [with a small lift] It’s in version 2.1.323 and later, you flip it on at startup and confirm, and it works in an interactive terminal session. The catch: it’s terminal-only for now, it doesn’t work through cloud providers or gateways, and code stops after ten minutes unless it runs in the background. Give it a task with lots of tool calls and read the code it writes. Host B: [lighter] That’s the show. See you tomorrow.

Host A: That’s Compact Conversations for Sunday. More AI news tomorrow. Until then, happy prompting.