Claude Cancellations Surge as AI Agent Wipes Production Database and Vibe Maths Cracks Erdős Problem

01"I cancelled Claude" hit 957 points on Hacker News the same week an AI agent deleted a production database

Nicky Reinert published a post titled "I cancelled Claude: Token issues, declining quality, and poor support" on his personal blog on April 24. Within a day it reached the top of Hacker News with 957 points and 574 comments.

The headline lists Reinert's complaints in order. He alleges three failures. Claude's token billing did not match his usage. Answer quality has declined over the life of his subscription. When he raised either issue with support, the responses did not address them. His decision to cancel followed all three, not any one in isolation.

957 points puts the post in the upper percentile of any tech topic on Hacker News this year. The 574-comment thread filled with developers describing the same pattern: opaque billing, perceived regression, support tickets answered with documentation links rather than account-specific data. One user wrote a long-form complaint; the comment volume is what made it news.

While Reinert's post was climbing, a separate story was running on the same site. A developer posting as @lifeof_jer reported on Twitter that an AI agent operating against their stack had deleted the team's production database. The post included what the author described as the agent's own after-action explanation. That thread drew 355 points and 489 comments.

The two items sit at opposite ends of the trust ladder. One paying customer walked away over billing and answer quality. Another team is restoring from backups because they handed an agent write access to live infrastructure and the agent used it.

The asymmetry matters for buyers. A single user cancelling a $20 plan does not move Anthropic's revenue line. Enterprise teams pulling agent permissions or moving workloads to a different provider do. Both audiences are reading the same HN thread.

For engineers running Claude in production this week, the practical work narrows to three audits. First, which workflows actually need a frontier model versus a cheaper alternative. Second, where agents currently hold write permissions to systems a wrong call cannot recover. Third, whether the support contract covers a billing audit or only a help-center link.

Anthropic has not publicly responded to either thread.

02A 13-acre Mill Valley listing now accepts Anthropic stock — days after Google committed $40B

Anthropic's valuation is being marked the same week from two ends of the capital stack. Google said it will invest as much as $40 billion in the company, according to Ars Technica. The commitment lands days after a similar but smaller Amazon round, and during the same week a Mill Valley seller listed a 13-acre property and said the owner will accept Anthropic equity as payment, TechCrunch reported.

The two transactions price the same asset at different layers. Google and Amazon are setting the institutional anchor through primary capital. The Bay Area listing, north of San Francisco, treats employee stock as a cash equivalent for a hard asset. Sellers usually demand wires.

That a private seller of a single-family property is willing to take private-company shares signals where secondary pricing has settled. Anthropic equity has historically traded through tender offers and SPVs at a discount to the last primary mark. A homeowner accepting it directly compresses that gap and gives employees a non-tender path to convert paper wealth.

Google's check size also resets the comparison set. The company had previously committed roughly $3 billion across earlier rounds, per prior Ars Technica reporting. A $40 billion top-end pulls Anthropic's strategic backing closer to the cumulative scale Microsoft has put into OpenAI, and it does so while Amazon is concurrently topping up its own position. Two of the three largest cloud providers are now anchoring the same lab on overlapping timelines.

The downstream effects are mechanical. Anthropic staff with vested shares gain a more liquid exit route at a higher reference price. Recruiters at competing labs lose a pitch. Secondary buyers pricing other frontier labs — xAI, Mistral, Cohere — now have a fresher Anthropic comparable to argue from, one validated by both a hyperscaler and a residential seller in the same week.

What is not yet public is the structure of the Google commitment: how much is cash, how much is compute credits, and over what period the $40 billion deploys. Earlier hyperscaler deals with frontier labs have been weighted toward cloud spend that flows back to the investor. The Mill Valley listing does not disclose the share price the seller is using, or what discount, if any, applies versus the new primary mark.

03Vibe maths cracked a 60-year-old Erdős problem, the same week the AI industry conceded the public hates it

An amateur without formal mathematical training used ChatGPT to make substantive progress on a question Paul Erdős posed about 60 years ago, Scientific American reported. Its writeup called the method "vibe maths": human direction plus model-generated and model-verified work, judged serious enough to be discussed alongside professional number theory. The contributor was not based at a research lab, held no graduate degree in the field, and used the same consumer product available to any subscriber.

That story landed in the same news cycle as a New Republic piece reporting AI industry figures had begun publicly conceding what their internal data had shown for months. Most users do not like the technology. A meaningful share actively resents it. The piece framed the moment as an internal acknowledgement rather than an external attack. Executives who spent the year describing AI as universally beneficial were quoted softening that line.

The two reports describe the same product family from opposite ends. In a narrow, expert-adjacent setting, consumer-tier ChatGPT helped extend a record on a problem mathematicians had not closed in six decades. Among general users, the same family of tools is generating documented backlash regardless of capability gains.

Neither outcome displaces the other. A polling number is not cancelled by a theorem. The two effects sit on different ledgers, with different audiences, on different time horizons. Erdős results move through a small expert circle that already uses these models. Survey data reflects the much larger group who never reads about them, and is not waiting to be persuaded by another capability story. The expert wins do not propagate into the consumer mainstream.

The next quarter of consumer sentiment surveys will determine which of the two signals the major providers price into their roadmaps. Capability-win coverage has not moved the polling in either direction so far.