Qwen 3.7 Open-Weights Release Imminent — Pricing Info Appears on OpenRouter
Reddit's LocalLLaMA community has surfaced evidence that an open-weights release of Qwen 3.7 is near: pricing information for "Qwen3.7-flash" has appeared on OpenRouter.
Since Qwen 3.6 called its 35B-A3B model "Qwen 3.6 flash," the 3.7-flash is likewise believed to be a small-scale MoE (Mixture of Experts) model. What stands out is the significantly lower price compared to the 3.6-flash, along with native support for a 1M (one-million-token) context window. The Qwen series has a solid track record as open-weight models, and the arrival of 3.7 will be a major milestone for the local LLM community.
There has been no official announcement yet, but the OpenRouter listing is a strong signal that release preparations are underway.
Over 30% of New Podcasts Are AI-Generated — Listen Notes Publishes the Numbers
Podcast search engine Listen Notes has published a startling statistic: **more than 30% of newly created podcast shows are AI-generated content (AI slop)**.
Unlike text, podcasts are audio media that is hard to vet carefully, making it easy for plausible-sounding but fact-free AI-generated content to slip in. The statistic shows that AI content pollution is spreading across audio media as a whole, not limited to specific platforms.
Around the same time, The Atlantic analyzed the phenomenon of AI-generated novels becoming bestsellers in a piece titled "What Comes After AI Slop." The line between human and AI in the content marketplace is getting blurrier by the day.
Big Tech Credit Risk Soars on Overheated AI Investment; Korea's KOSPI Plunges Nearly 5%
According to the Financial Times, credit risk at giant IT companies is rising rapidly on the back of massive AI investment. The scale of AI infrastructure spending is reaching levels that raise concerns about these companies' financial resilience.
The worry is spilling over into stock markets. Reuters reported that AI-related concerns triggered a plunge in chipmaker stocks, sending South Korea's KOSPI index down nearly 5%. While this doesn't point to a collapse of the AI bubble, it looks like a sign that investor psychology is shifting from "buy everything" to "selectivity."
With the outlook for recouping AI investments unclear, the effects are beginning to appear in Asian markets, including semiconductor stocks. The coming earnings season will put the reality of AI revenue to the test.
Claude Code Ships 8 Releases in 2 Weeks — /fork Rebuilt and the Pitfalls of Parallel Use
Claude Code went from v2.1.186 to v2.1.212 — 8 releases in just two weeks. The changelog entries add up to 290, but three changes matter in practice.
First, **the `/fork` command now duplicates the entire conversation**. It used to create only a partial branch; now it takes a complete snapshot of the conversation, which serves as a safety net when trying out experimental changes.
Second, **subagents and MCP servers now run in the background by default**. Auxiliary work proceeds without blocking the main conversation.
Third, it turned out that **worktree isolation in v2.1.210 and earlier had a security flaw**, which has been fixed. If you run Claude Code in parallel, updating to the latest version is recommended.
"Prompt Repetition": Paste the Prompt Twice and Accuracy Goes Up
The "Prompt Repetition" technique, reported by Google Research in late 2025, is once again drawing attention in Japan's AI community.
The technique is remarkably simple: **just state the same instruction verbatim, twice in a row, in the prompt**. In settings without CoT (Chain of Thought), accuracy improved significantly in 47 of 70 tested conditions, and not a single condition got worse.
Why does it work? When an LLM processes instructions sequentially, a phenomenon known as "instruction dilution" kicks in: the later a constraint appears in the prompt, the weaker the attention it receives. Having the model re-read the same instruction as a second pass redirects attention to the constraints that were overlooked the first time.
Unlike Self-Consistency (generating multiple responses to the same prompt and taking a majority vote), it delivers gains in a single generation. That's what makes it practical. It's a simple yet powerful way to boost accuracy without added cost — one worth knowing.
Claude Grew a Tomato for 100 Days — New Territory for Autonomous Agents
A project in which Anthropic's Claude autonomously grew a tomato from seed to fruit over more than 100 days has been drawing attention.
What makes this experiment notable is that the feedback loop lived in the physical world. Claude not only handled watering, light adjustment, and temperature control in real time, but also **designed its own circuits and extended the hardware** as needed. Unlike "task execution" on the chatbot continuum, it stands out as a case demonstrating long-horizon autonomy in a physical environment.
While it shows the potential of autonomous agents, it also surfaces reliability and safety challenges in long-running execution. The 100-day duration may well become a benchmark for how far LLM-based agents can carry sustained real-world tasks.
---