AI Cost Cuts, Security Warnings, and the Realities of Small Business Ownership

Key Takeaways

1. @levelsio cut monthly infrastructure costs by nearly $34,000 using Claude Code, reducing GPU spend from $36k to $7k and backup costs from $1,200 to $26.
2. A former Scale AI security lead warns that AI security is entering a "hell" period, arguing that models should not control deterministic systems and that credentials, not model behavior, are the real risk.
3. Voicy, a speech-to-text tool built for the founder's father, now generates $16k–$20k per month, with 20–30% of users being elderly or disabled.
4. Brainrot, a screen-time app that adds friction instead of hard blocking, grew monthly revenue from ~$10k to ~$400k.
5. Codie Sanchez argues that buying a small business is often "buying yourself a second job," citing historical examples like Buffett's Berkshire and Schultz's Starbucks.
6. Alex Lieberman shares lessons from David Campbell on AI security: the BICO framework (Behavior, Identity, Control, Observability) and a warning that defenders are bringing "knives to a gunfight."
7. Justin Welsh argues that the most dangerous addiction is seeking approval from people who don't matter, and that true success is working 2–3 hours a day on your own terms.

1. AI and Infrastructure Cost Optimization

1. @levelsio reported that his entire portfolio of websites and apps now costs about $300 per month to operate, covering 20,000 paying users, 252 million annual visitors, 2.4 billion requests, and 2.4 PB of bandwidth. Including Photo AI's $7,000 monthly GPU bill and $800 for R2 storage, the total is about $8,000 per month. He credited Claude Code with saving nearly $34,000 per month by switching to cheaper video models, moving tasks to Google Cloud and Sume, consolidating servers to Hetzner, applying Backblaze B2 lifecycle policies, and cleaning up unused domains. He argued that AI is good at finding cost savings, which is bad for the economy short-term but positive long-term as it cuts waste and redirects spending to better investments. 1
2. Noah Kagan said Claude now handles AppSumo's SEO every Friday, and he no longer reads the full reports, asking Claude for at most three action items per week. Last week, Google Search revenue was $64,320, and ChatGPT mentioned AppSumo in 8 out of 10 answers. 1

2. AI Security and Agent Risks

1. Alex Lieberman summarized a conversation with David Campbell, former head of security research at Scale AI, who described three eras of security: castle moats (~30 years), cloud (~10 years), and AI (~4 years). Campbell argued that early AI security focused too much on model behavior and guardrails while people handed over credentials to agents. He predicted the next 18–24 months will be "hell" for cybersecurity professionals because they are "bringing knives to a gunfight," and recommended returning to solid external boundaries and edge protection. 1
2. Campbell's BICO framework examines agents across four dimensions: Behavior (can instructions, retrieved context, memory, and tool outputs influence decisions; can data be distinguished from instructions; are important actions verified independently of the proposing model), Identity (can the agent be distinguished from the initiator; are credentials short-lived and scoped to the task), Control (deterministic boundaries between reasoning and authorization; permissions for tools, parameters, data, and destinations; policy checks or approval gates), and Observability (can the decision chain be reconstructed, logging context sources, identity and authorization events, tool calls, policy evaluations, approvals, changes, and outcomes). 1
3. Campbell's ICLR paper on "defensive refusal bias" found that in the NCCDC competition, students' LLMs were 2.7 times more likely to refuse when told the request was for security work than when asked directly "is this malware?" He argued models may be over-trained to be unhelpful to defenders, while attackers currently have a huge unfair advantage and are moving at exponential speed. He recommended hardening edges, triaging vulnerabilities quickly, hiring more security staff before models can truly help defenders, building engineering tools to replace AI guardrails with deterministic controls, and expecting burnout in security teams in the coming months. 1
4. Arvid Kahl questioned the idea that "if an AI agent does it, it's not a crime," noting that agents act on the explicit instructions of their orchestrators, and the indifference of developers and CEOs is concerning. He also argued that "AI forces us to pay off 40 years of technical debt" really means we collectively have the ability but lack the organization to clean up digital tools, which could be open source's finest moment or its lowest point due to automated PRs and anti-AI policies. 1 2

3. Bootstrapped Product Growth and Small Business Realities

1. Kourosh built Voicy, a speech-to-text tool based on OpenAI Whisper, to help his father who has difficulty typing. It now generates $16k–$20k per month, including $11k MRR, with a $260 lifetime deal, 10,000 monthly active users, and 1,400 active subscribers. User research showed that 60–70% of users are fast typists seeking a 10% efficiency gain, while 20–30% are elderly or disabled, prompting him to focus on the latter group with a 20% disability discount, now used by 20% of users. He argued that for ordinary users, dictation is a nice-to-have, but for the group he serves, it is a necessity that alleviates pain. Growth came from the Chrome Web Store with five friend reviews for the first 100 users, Reddit posts until banned for 200–300 users, SEO with 200 blog posts, 30 landing pages, and 4–5 languages, and a YouTube "AI for Seniors" series that grew subscriptions from 1K to 12K in a month. 1
2. Yoni built Brainrot, a screen-time app that turns limiting usage into a game: when users open Instagram or similar apps, they receive a prompt and must go through a short flow, with a demo featuring "your brain FaceTiming you." It does not hard-block but adds friction to interrupt unconscious opening. Launched in summer 2025, monthly revenue grew from about $10k to about $400k, and Yoni said a small amount of friction is enough to change habits. 1
3. Codie Sanchez argued that buying a small business is often "buying yourself a second job," citing examples: Buffett bought a failing textile mill in 1965 that became Berkshire Hathaway, now worth over $1 trillion; Schultz bought 6 Seattle coffee stores for $3.8 million in 1987, and Starbucks now has over 41,000 stores and $37 billion in annual revenue; Les Schwab sold his house and borrowed $11,000 to buy a tire shop, now with over 600 stores and over $1 billion in sales. She also cited Carl Karcher mortgaging his car to buy a hot dog cart, Wayne Huizenga borrowing $5,000 from his father to buy a used garbage truck, and Kenn Ricci buying a small charter company for $27,500 that grew into an aviation group operating over 340 aircraft. She argued one can also buy businesses with enough profit to hire operators, then grow, add value, and exit, noting she bought her first business for under $100,000 while working 60–70 hours a week in finance. 1
4. Justin Welsh said his life goal is not to be as rich as possible but to work 2–3 hours per day, help thousands of people, and have ample time for hobbies, outdoors, cooking with his wife, exercising anytime, and avoiding meaningless meetings. He argued the most dangerous addiction is seeking approval from people who don't matter, and that an underrated life skill is being happy when a friend achieves greater success than you. He also shared the view that whom you marry, what you do for work, and where you live shape your life more than any productivity system. 1 2 3 4