Qwen’s Overthinker, Data Center Emissions, and Agent Harnesses

Compact Conversations for 2026-08-16: 6 AI stories, ai news worth knowing in just 5 minutes.

[Audio embed placeholder]

The Lead: Qwen 3.8 27B is excellent, but it defaults to wildly overthinking things

Alibaba’s Qwen research lab released Qwen 3.8 27B, a 27-billion-parameter, vision-capable open-weights model that fits in a 17GB file. Its default ‘reasoning effort’ setting is set to maximum, causing it to spend minutes over-analyzing simple tasks like drawing a circle.

Why it matters: This model’s strong performance in coding, tool use, and vision shows the rapid progress of local AI, but its impractical default highlights the importance of configuration for real-world use on consumer hardware.

Source: Simon Willison’s Weblog

Number to Know: Big Tech’s data centre boom poised to drive up carbon emissions

A Financial Times analysis of the 60 largest planned data center projects finds they could emit the carbon equivalent of 27 coal-fired power plants, or 24 million cars, per year.

Why it matters: This quantifies the significant sustainability challenge directly tied to the scale of new AI compute infrastructure, with most projects concentrated in the U.S. and Europe.

Source: Financial Times

The Feed

AI Weekly Issue #522: Zuckerberg promises superintelligence for all. Experts aren’t sold.

Mark Zuckerberg’s 6,500-word case for universal superintelligence was widely shared among AI experts, but met with widespread skepticism. The newsletter also tracked stories about an AI agent hacking a gym booking system and Claude subscribers canceling over provenance watermarks.

Why it matters: The expert reaction underscores the gap between ambitious theoretical promises and the practical, often messy realities of AI deployment and trust.

Source: AI Weekly — AI News & Updates

OpenAI upheaval mounts as Sam Altman readies IPO push

OpenAI is facing executive exits and changes to its safety team, unsettling staff as the company prepares for a blockbuster initial public offering. The departures include a senior safety researcher.

Why it matters: The internal changes raise questions about organizational priorities and safety focus during a critical period of growth and financial transition.

Source: Artificial intelligence

How to break ChatGPT’s sycophancy bias: The Red Team Perspective Challenge prompt

A prompt engineering framework forces an LLM to adopt the persona of a ruthless ‘red team’ critic to find structural flaws in plans and arguments, overriding the model’s default tendency to be agreeable.

Why it matters: This provides a concrete method to get more rigorous, critical feedback from AI assistants for strategic planning, system design, and investment proposals.

Source: appliedaihub.org

CORS Chat

A web tool for testing OpenAI-compatible chat endpoints, like local models running in LM Studio. It can progressively render SVG images in the chat as tokens stream in and persists conversations in the browser.

Why it matters: It’s a practical utility for developers experimenting with local or custom model deployments, simplifying the testing and visualization of model outputs.

Source: Simon Willison’s Weblog

One Thing to Try

Review a Reddit discussion where developers share the custom ‘harnesses’ they’ve built around coding agents. A recurring useful pattern is implementing a separate review agent—a second, often smaller model that checks the primary agent’s plan or code before execution to increase reliability.

Sources

Transcript

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

Host A: [curious] For the weekend update, Alibaba’s Qwen research lab released Qwen 3.8 27B, an open-weights model with 27 billion parameters and vision capabilities that fits in a 17 gigabyte file. It’s impressive, but the default settings make it overthink everything.

Host B: The model supports a reasoning feature—basically letting it think through problems step-by-step before answering. By default, that’s set to maximum. One developer testing it asked the model to draw a simple circle. The reasoning trace went on for several minutes about creating a ‘geometric study’ with animation and a distinctive palette. [with emphasis] His strong recommendation: ignore the default and run it on low or no reasoning first. When you do, the model’s performance is solid.

Host B: One number to know today is 60. That’s how many of the largest planned data center projects the Financial Times analyzed in a report on AI infrastructure’s environmental impact.

Host A: [thoughtful] Those 60 facilities could emit the carbon equivalent of 27 coal-fired power plants, or 24 million cars, per year. The report highlights the sustainability challenge tied directly to the scale of new AI compute, with most of these projects concentrated in the U.S. and Europe.

Host A: From AI Weekly, expert reaction to Mark Zuckerberg’s lengthy case for giving every person superintelligence. The document was widely shared among AI researchers and builders over the last week, but almost none of them shared it kindly.

Host B: The newsletter also tracked experts passing around stories about an AI agent that hacked a gym’s booking system and the first hard number on the cost of content provenance: Claude subscribers canceling over an invisible watermark. [skeptical] These are examples of the practical, sometimes messy realities that follow big theoretical promises.

Host A: The Financial Times reports on mounting upheaval at OpenAI. Executive exits and changes to the safety team are unsettling staff as the company prepares for an initial public offering. The departures include a senior safety researcher and restructuring of the team focused on catastrophic risk—which raises questions about organizational priorities during a critical growth phase.

Host B: Two quick tools worth knowing about. Simon Willison built CORS Chat, a web interface for testing local models like Qwen 3.8 27B running in LM Studio. It has a neat trick of progressively rendering SVG images in the chat as tokens stream in. And Applied AI Hub published a ‘Red Team Perspective Challenge’ prompt designed to break ChatGPT’s sycophancy bias by forcing the model to act as a ruthless critic instead of a cheerleader.

Host A: One thing to try is looking at how you structure the safety net around coding agents. A Reddit thread asked developers what custom harnesses they’ve built beyond basic instructions.

Host B: [conversational] The interesting signal from the discussion is that several people find a separate review agent—a second model that checks the primary agent’s plan or code before it runs—genuinely increases reliability. It’s a simple pattern: let one agent propose, and a second model acts as a gatekeeper before any commands execute. One commenter suggested using a smaller, faster model for the review step to keep costs and latency low.

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