Ref Raises $4M and New AI Model Claims 1,600x Cheaper Than GPT-5.6 Sol

Key Takeaways

1. Ref — a shared space for teams to decide what agents build before writing code, launched publicly with $4M in funding; its creator warns that letting agents make critical decisions is dangerous and coins "Velocity Sickness." — via 1
2. A new frontier classification model, rd-signal-2, claims to be 1,600x cheaper than GPT 5.6 Sol, with free trial available via @raindrop_ai. — via 1
3. Anthropic's invisible watermark for Claude-generated text raises concerns that AI providers could end up owning all AI-generated software; the coming watermark-removal arms race may be legally questionable. — via 1 2
4. China's July auto exports surged 88% year-over-year to 918,000 units (41% of production), while domestic passenger car retail sales fell 21% and fuel vehicle sales plunged 41%. — via 1
5. Arvid Kahl published "Hidden Multipliers" on growing a company with existing budget and team, including free AI skill resources. — via 1

1. AI Models and Safety

1. A new frontier classification model, rd-signal-2, claims to be 1,600x cheaper than GPT 5.6 Sol and is now available for free trial through @raindrop_ai. The offering includes training/hosting APIs for custom classifiers with zero data retention. This could substantially lower the cost barrier for teams that need specialized classification beyond general-purpose LLMs. — via 1
2. Anthropic's plan to embed invisible watermarks into Claude-generated text has Arvid Kahl worried that if ownership claims are attached to AI output, providers would effectively own all software produced with their models. He also questions the legality of adding a de-watermarking/diffusion step (especially for streaming APIs) and expects watermark-removal tools are already in development, calling it an inevitable and foolish cat-and-mouse game. — via 1 2
3. In the age of agentic programming, Arvid Kahl argues we need to rethink how strict safety guardrails should be; current tooling is insufficient to "guarantee" safety, and best practices have not yet integrated such approaches. This suggests that organizations adopting AI agents must actively define their own safety boundaries. — via 1

2. Product Launches and Funding

1. Ref, a shared space where teams decide what agents should build before writing code, opened public testing with $4M in funding. Its promoter Ben Tossell described "Velocity Sickness" as the phenomenon of engineering teams feeling fast but not impactful, and warned that putting agents in charge of critical decisions is the most dangerous outcome because humans lose control of both the system and the product. This positions Ref as an early attempt to bring structured planning to agent workflows. — via 1
2. Kaku v0.18.0 is now available for AI coding, introducing the ability to drag tabs and all their panes into standalone windows, keep windows below the menu bar when dragging across monitors, and add confirmations before menu restarts and Homebrew upgrades. These UX improvements aim to make the terminal more reliable for multitasking during AI-assisted development. — via 1
3. Arvid Kahl released his new book "Hidden Multipliers", which explores how companies can grow using existing budgets and teams. The launch includes purchase channels and a linked set of free AI skill resources, making it a practical playbook rather than just a theoretical read. — via 1

3. Market Data and Economic Signals

1. Citing China Passenger Car Association data, China's July auto exports jumped 88% year-over-year to 918,000 units, accounting for 41% of monthly production. In contrast, domestic passenger car retail sales fell 21%, while fuel vehicle sales plunged 41%. The divergence highlights how much the Chinese auto industry now depends on overseas markets as domestic demand weakens. — via 1
2. An analysis of July CPI data shows that beyond gold jewelry, computers, and phones (driven by gold prices and chip costs), most categories lost momentum once subsidies were removed—home appliance growth returned to zero, and pork prices were still down 13.3%. This suggests that the rebound in consumer prices is fragile and highly dependent on one-off policy support. — via 1

4. Engineering and Startup Insights

1. Jack Ellis advises not to use a dedicated database for caching workloads, even when you have 400GB of free memory. He argues that dedicated cache infrastructure adds operational complexity and cost, and simpler application-level caching is often sufficient. — via 1
2. Tibo shared that before selling Tweet Hunter and Taplio, he had two startup failures: a gaming company that raised $500K with a 20-person team, and a recent attempt to restart using Fable 5 that also failed. He plans to tell the full story and lessons in tomorrow's newsletter, offering a candid reminder that funding and team size are not proxies for product-market fit. — via 1
3. Tibo highlighted a (non-sponsored) user testimonial for Outrank: three months in, the user's Google Search Console clicks tripled and content is now cited by large language models. This provides an independent data point for the tool's effectiveness in improving search visibility and AI discoverability. — via 1
4. Arvid Kahl observes that many software professionals no longer feel the need to read through AI-generated code because their test suites are sufficiently reliable. This points to the growing importance of strong testing infrastructure as a prerequisite for safely adopting AI-generated code. — via 1