🍄 DAILY SHROOMS

Zhipu Just Turned AI Hype Into a $4 Billion War Chest

Here's the deal: Chinese AI model maker Zhipu went public in Hong Kong back in January, and since then its stock has done something absolutely bonkers — up nearly 1,500%. That's not a typo. That's the kind of number that makes crypto's craziest bull runs look tame. Today, Zhipu announced it's capitalizing on the hype with a $4 billion share sale. Smart move, honestly — when the market's throwing a party this good, you sell tickets at the door. This comes on the same day Iluvatar CoreX (a Shanghai chipmaker) is trying to raise $850 million off its own post-IPO surge. Plot twist: while everyone's watching OpenAI, Anthropic, and Google slug it out, China's AI stock market is quietly minting fortunes. The bottom line: investors aren't just betting on which model is smartest — they're betting on entire national AI ecosystems, and right now China's public markets are on fire.

• POW!France’s competition authority ordered Meta to resume talks with French media groups and propose a payment plan over unpaid publisher fees. No more “we’ll negotiate when the vibes improve.”
• ZOOM!Apple is spending more than $30B with Broadcom on U.S.-made wireless chips through 2031. Not core AI chips, but very relevant to the bigger U.S. silicon reshoring story.
• BAM!DeepSeek is reportedly developing its own inference chip to reduce reliance on Nvidia and Huawei. The AI labs are learning the same lesson: renting the future from someone else’s hardware stack gets expensive fast.
• WHOA!Alibaba had its best single-day share-price gain in 10 months as investors rotated back into Chinese tech. The China AI/tech trade is very much alive.

Akashic / MemAttention

Akashic is a new research system for making long-running AI agents less wasteful. Instead of forcing the model to reread an entire conversation or workflow history every time, it organizes context into chunks and models relationships between them. The reported result: better accuracy, better throughput, and higher sustainable request rates versus prior memory baselines. Builder takeaway: if your agent is slow, forgetful, or expensive, the answer may not be “use a bigger model.” The answer may be “stop making the model reread its entire life story every turn.”

CSTutorBench: Benchmarking Small Language Models as Tutors for Block-Based Programming

Researchers introduced CSTutorBench, a benchmark for testing small and large language models as computer-science tutors in VEX VR’s block-based robotics environment. The interesting part: models did fine on tone and vocabulary, but struggled with deeper tutoring behavior — especially avoiding answer leakage and understanding students’ debugging history. Translation: sounding like a good tutor is easy. Actually teaching is still hard.

Read the paper → · via arXiv AI Papers

Douglas Elliman

Douglas Elliman is pushing deeper into AI with Elius, an AI subsidiary powered by Google Cloud technology that is meant to overhaul the brokerage’s operations and tech stack. This is the kind of “boring industry, real AI adoption” story that matters. Not every AI win looks like a new model release. Sometimes it looks like an old-school real estate brokerage quietly rebuilding its operating system around AI.

The Engineering Story: Memory Is Becoming Infrastructure

No flashy API pricing war today, but the research feed is giving builders a very clear message: long-running agents need better memory systems, not just bigger context windows. The “Memory in the Loop” paper argues that moving memory in-process can reduce retrieval from tens or hundreds of milliseconds to roughly 100 microseconds, making per-step agent memory practical instead of painfully slow. Pair that with Akashic’s MemAttention approach, which organizes conversation history into bounded semantic chunks instead of replaying the whole history every turn, and the trend is obvious: 2026 agent performance is not just about model choice. It is about memory architecture.

AI Capital Is Still Flowing — But The Bar Is Getting Meaner

Zhipu seeking around $4B is the cleanest money signal today. Capital is still flooding into AI, especially in China, but investors are getting more selective. The old trade was “AI exposure = buy.” The new trade is “AI exposure, but show me durable economics, compute access, and distribution.” That is why today’s market story is so interesting: OpenAI is dealing with rollout controls, Anthropic is dealing with usage-credit optics, and Zhipu is raising war-chest money. The market is not done believing in AI. It is just done believing for free.