AI rebuilds, Chinese models, and agent TTL: founder reads for Aug 28
A packed week of founder reading. AI rebuilds, token pricing wars, agent memory science, and a data set that reframes how to think about YC Demo Day caps. Here is what is worth your time.
Headlines & Trends
CEOs truly rebuilding for AI are maybe 40% done
There is a difference between bolting AI onto an existing product and actually rebuilding a company around it. The former is easy. The latter is what most CEOs claim to be doing and few are actually doing. The piece estimates most are roughly 40% of the way through a genuine rebuild. Fin’s four-year rebuild resulting in a $3.6 billion exit is cited as the benchmark. Companies executing real rebuilds are reportedly seeing growth rates of 10% to 30%, which is modest for the effort involved.
US startups are quietly switching to Chinese AI models
The cost gap is hard to argue with. DeepSeek charges $0.14 per million tokens. OpenAI charges $5.00. US startups are noticing, and Chinese models now dominate OpenRouter traffic, largely displacing US-developed systems. The tradeoff is real: US lawmakers are investigating potential cybersecurity and data privacy risks from these models. This is a decision every founder building on third-party AI infrastructure needs to make deliberately rather than by default.
Strategies & Tactics
Lazyweb hit 50k signups with $0 in marketing spend
Lazyweb reached 50,000 agent signups in a couple of months without spending on marketing. The founder built in public and focused specifically on reducing friction for agents using the product. The post is a retrospective on what worked and what did not.
How long should an AI agent live?
This is a product decision most teams are not making deliberately. The data cited in this essay is worth reading carefully: squeezing a conversation context to fit token limits dropped standing instructions in 30% to 59% of runs, depending on the model. Memory quality degrades as it grows across every frontier model tested. Memory carried between sessions is also a potential attack surface for prompt injection.
The proposed architecture: a coordinator agent with a one-day lifespan, specialist agents running for roughly 30 seconds, and state written to ordinary files rather than held in the agent’s context window.
⚒️ Tools & Resources
Salesforce put Claude at the center of its products
Claude is now the reasoning model behind Salesforce’s Atlas Reasoning Engine. It also powers Agentforce Vibes and Agentforce Coworker by default and is available in Agent Builder. Slack is now Claude-powered as well. On the other side of the integration: Salesforce for Claude is a plugin that gives sellers 37 prebuilt sales skills that run inside Anthropic’s interface rather than inside Salesforce’s.
Cohere Parse: $1.50 per 1,000 pages for document intelligence
Cohere launched Parse, a vision language model for processing large volumes of enterprise documents. It converts complex multimodal files into structured, machine-readable data, handles key visual elements, and works across nine major world languages. Pricing is $1.50 per 1,000 pages via the Cohere API.
Miscellaneous
The data case for investing in expensive YC startups
A fund analyzed 642 YC companies from the 2021 through 2024 batches and found a clear pattern. Companies priced above their batch median valuation cap reached a Series A or beyond 25% of the time. Companies priced below median made it only 8% of the time. Shutdown rates ran the other direction: 6% for above-median companies versus 16% for below-median. Median caps moved from $15M to $40M between W21 and P26. The pattern held inside each individual batch and among companies the fund passed on.
AI-era business formation is moving to the suburbs
Stripe Economics published data showing AI-era businesses are forming in outer suburbs rather than city centers. The companies directly responsible for AI development continue to cluster in major metro areas. But new AI-era business formation is broadly distributed across the US, with major metro areas seeing below-average formation relative to their population size.
⚡ Quick Links
- Focus and followthrough are the moat — The argument is simple: no one wants a half-finished product or just a shiny idea.
- Listen Labs used a cryptic billboard to hire engineers — The San Francisco billboard encoded AI-related tokens that only the right candidates could decode. Niche recruiting at its most literal.
- Listening to your market beats third-party data — Connecting directly with customers and tracking organic marketing signals lets you spot trends before they show up in syndicated research.
- Bank Statement Converter is watching its business shrink — The founder reports more cancellations and fewer new subscribers, and attributes the decline to people using AI directly to convert their bank statements.
I’m founder of BizStack at Cagri Sarigoz LLC and a passionate advocate for entrepreneurs.
With over 14 years in tech, marketing, and AI, including my role as Head of SEO at CitizenShipper and co-founder of TaleBot at Intale AI, I’m dedicated to sharing genuine, useful product insights and tips.
At BizStack, I aim to cut through the digital noise to provide clear, actionable advice.
And more than all else, I’m a father to a (always) little girl and a husband.
Contact me at [email protected] for assistance.
