AI Daily Newsletter: The Demise of Amazon Mechanical Turk and the AI Inference Price Storm

🧠 GLM 5.2 Initiates the Era of Margin Collapse for AI Inference

Zhipu AI releases GLM 5.2 – the first open-weight model considered a true competitor to Claude Opus and GPT for agentic tasks, but at only 15-20% of the inference cost. Concurrently, they launch ZCode, a coding agent platform directly competing with Claude Code and OpenAI Codex, offering a free 5-day trial. An analysis by Martin Alderson calls this the “beginning of the AI margin collapse” as open-weight models have caught up with closed-source models in terms of agentic work quality.

Source: HN discussion | The Decoder | Martin Alderson

🛑 Amazon Discontinues Mechanical Turk After Nearly 20 Years

Amazon announces that it will stop accepting new customers for Amazon Mechanical Turk (MTurk) – the legendary “Artificial Artificial Intelligence” platform that was once the primary tool for labeling data for numerous first-generation AI models. The Decoder notes that the end of MTurk marks the complete shift of AI companies to using LLMs for data labeling instead of humans, creating an “AI training AI” loop.

Source: TechCrunch | The Decoder

☁️ Cloudflare Replaces Blanket AI Bot Block with Granular Controls

Cloudflare changes its AI bot blocking policy: instead of blocking all AI crawlers, it now categorizes them into three groups – search crawlers, training crawlers, and agent crawlers – each with its own policy. This move is seen as a way for Cloudflare to protect publishers while not blocking legitimate use cases (such as Google Search). This step is also in line with the new policy requiring AI companies to pay for publisher content.

Source: The Decoder | TechCrunch

⏳ Epoch AI: Top Models Now Last Only 7 Weeks Instead of 1 Year

Research from Epoch AI shows that the rate of replacement of top models has increased dramatically: GPT-4 once held the top position for a year, while today’s state-of-the-art models only last an average of 7 weeks before being surpassed. The main reason is the explosion of lab rooms (from the US, China, Europe) and the fact that open-weight models have narrowed the quality gap with closed-source models to almost zero.

Source: The Decoder

🔀 Vercel CEO: The Battle to Separate Models from Agents

Guillermo Rauch, CEO of Vercel, shares his thoughts on the architectural battle happening in the AI industry: whether models and agents should be separated or not. His viewpoint is that clearly separating the model layer! (reasoning capability) and the agent layer (action capability) is key to building trustworthy AI applications. This debate reflects the emerging trend of agent frameworks (such as LangChain, Vercel AI SDK) becoming increasingly separate from model providers.

Source: TechCrunch