Outside tech still isn't seeing AI pay off
The most useful thing in today's news is a warning. Business spending on AI tokens grew just 0.2% in September, and some companies are cutting back because token costs are eating into margins. Buyers outside tech want to see a measurable gain, not a clever demo. Before you raise prices or add features, check what each customer costs you to serve.
- Outside tech still waiting on AI payoff
- Five Qwen models teaming up match Fable 5
- Mac notch shows your Claude Code agents
- Quick hits across the AI field
- Formic robot round and AI Index move
Top stories
All news →What happened: A note from Apollo's chief economist says companies outside tech are not yet getting a measurable return from AI spending.
- Business spending on AI tokens grew just 0.2% in September, down from 42% in July, based on Ramp data.
- 80.4% of the top 1% of AI spenders in the data cited are VC-backed.
- A St. Louis Fed review of about 490,000 earnings calls found executives mostly describe AI productivity gains as still to come.
Why it matters: The test for your company is whether a buyer outside tech sees a margin gain, not whether your model is capable. If your product uses many tokens per customer, expect cost pushback and price with that in mind.
What happened: A GitHub project reports that five Qwen3.8-27B models working together can match Fable 5 when run with the GVS5H Harness.
- Five Qwen3.8-27B models are combined using the GVS5H Harness.
- The post reached the Hacker News front page with 165 points and 57 comments.
- The builders' discussion is at news.ycombinator.com/item?id=50228167.
Why it matters: If the match holds up in your own tests, a group of Qwen models may be worth comparing against the model you use now. Run your real tasks through it before you change anything.
What happened: Agent Notch puts a live count of waiting Claude Code agents beside the Mac notch, so you can see who needs you without switching windows.
- A red count appears when an agent asks a question or waits for permission, and clicking a row jumps to that agent's app.
- The app works only with Claude Code on Mac and needs macOS 14 or later.
- It is released under the MIT License and makes no network requests except when you press Check for Updates.
Why it matters: Users running several agents want to see which one is working, which one is blocked, and how to jump to it. Any agent product that hides that state inside a terminal will feel slow to supervise.
Quick hits
- Cerebras says it now runs trillion-parameter models, using Kimi K3, about 3 trillion parameters, as its worked example. cerebras.ai →
- zCAD AI turns plain text into editable, dimension-driven 3D parts and exports them to STEP for any CAD tool. Product Hunt →
- A preprint of nine models finds AI text uses more inanimate subjects and rarer word patterns than human writing. apricitas.io →
- Canva's in-house Craft image model debuted fourth on DesignArena and comes with a public API for developers. Product Hunt →
- A 380-point Hacker News thread debates how much AI caused a loss of 200k jobs in media, film and the arts. arxiv.org →
Newsletters & podcasts today
All newsletters →
