posts · 2026-07-06

This piece is worth opening because Tunguz turns the vague "will AI cost more than people" debate into three concrete numbers you can compare directly. Anthropic's compute spend runs 2.3x payroll at $2M per employee — which actually tracks with its $14M revenue per employee. But the rest of the software industry lives in a different universe: the top 1% spend $89k per engineer, roughly 40% of salary, while the median spends $137 a year — basically nothing. The three 2029 scenarios hinge on two variables: whether token prices keep falling 10x per year, and whether agentic workflows drive a 24x jump in token consumption. The Bull case hits $596k per engineer; the Bear case drops back to $106k. I'd discount one thing: a big chunk of Anthropic's $2M is training spend, and most companies won't train foundation models, so that benchmark is extreme. But the Ramp payment data is real — a 680x gap between leaders and the median says most companies haven't really started using AI for actual work yet.

This is worth a click because it skips the valuation gossip and goes straight to the legal knot that makes an OpenAI or Anthropic IPO genuinely hard. Both are registered as public benefit corporations, but the core assets and ultimate control still sit with the nonprofit. That means the standard toolkit for protecting public shareholders—disclosure rules, anti-fraud provisions, shareholder lawsuits—doesn't map cleanly onto this structure. The article didn't get a response from either company and offers no IPO timeline. I'd read it as a legal risk flag, not a prediction that they won't float.

The /goal prompt numbers are what make this worth a click: $165 in API costs, 25 autonomous experiments, 50% faster builds, 60% lower token spend. Not "might be useful" — the author says they verified these repeatedly. The other prompts are similarly practical: a work-mode prompt that turns usage patterns into reusable Skills, an action-norm prompt that keeps subagents in check, a reverse-interview prompt that demands 95% confidence before executing. The core idea isn't "make the model stronger" — it's "make the model study how you work, then save you effort." This is an RSS snippet, so the actual prompt templates and task details aren't shown. I'd discount the numbers a bit — they come from one person's workflow and may not replicate directly. But the mindset is right: once API billing kicks in, treating prompts as portable assets matters more than chasing the newest model.

This one's worth opening because it lays out the full chain of accountability failure in one case. The plaintiff's attorney used AI to fabricate precedents. The defense attorney knowingly applied Texas-specific legal software to a Mississippi case. Both local lawyers signed filings without reading them. The judge revoked all four lawyers' pro hac vice admissions and banned them from the district for two years. I'd discount the headline drama a bit—the post doesn't give exact fine amounts—but the harder numbers are there: 239 U.S. lawyers sanctioned for AI hallucinations as of February 2026, 712 global rulings in 2025 alone. Stanford's testing is even more direct: legal-specific tools like Lexis+ AI and Westlaw AI-Assisted Research still hallucinate at 17% to 34%. Vendor claims of 'hallucination-free' have zero independent evidence. The core issue isn't AI going wrong. It's that the legal profession's existing verification norms are breaking under AI acceleration. Senior partners have always had to read junior associates' drafts. Local counsel have always had to review out-of-state filings. That's been the rule for a century. AI just swaps the delegate from human to tool—the verification duty hasn't changed. But 43% of firms have no AI usage policy, and 54% offer no training. When execution cost drops to near zero, judgment gets more expensive.

This piece connects two stories that look unrelated at first: the FBI seizing a residential proxy network, and SEO companies selling ChatGPT scraping services. NetNut's scraper catalog listed ChatGPT Scraper, Perplexity Scraper, and Google AI Mode Scraper right alongside standard web scrapers. The same infrastructure layer serves brands tracking their AI visibility and attackers harvesting model outputs for distillation. Anthropic's February report put numbers on it: 24,000 fraudulent accounts made over 16 million Claude interactions from DeepSeek, Moonshot, and MiniMax. Reuters later reported an Alibaba-linked figure of nearly 25,000 accounts and 28.8 million interactions. These aren't SEO queries—they're systematic model output collection. The article maps the industry into four layers, from proxy infrastructure up to brand visibility dashboards. The most useful insight: model API outputs differ from what real users see on the interface. So both brand monitoring and distillation attacks bypass APIs and hit the actual product surface. There's no official observability tool for this yet—third parties fill the gap with simulation and sampling. What's missing: how much AI scraping NetNut actually sold, and at what price. The post doesn't give those numbers. But the logic holds: AI search shifted traffic from blue links to generated answers, the visibility gap created a new market, and the access layer serves both legitimate and adversarial use cases.

This one's worth opening because it turns model values from a vibe into coordinates you can compare. Gemini and Qwen sit next to each other on self-expression, GPT-4o and DeepSeek R1 are near-twins, but DeepSeek's own R1 and V4 Flash land on opposite ends of the secular-traditional axis. Tunguz's explanation is straightforward: Common Crawl is 46% English, so the baseline voice models imitate is a college-educated American online, and post-training pulls them in different directions. I'd discount the framing a bit—the World Values Survey is a Western academic instrument, so using it to measure global models' moral positions has some built-in assumptions. But the procurement question Tunguz raises is real. Worldview is invisible in code generation, but it's a live input in marketing copy, user behavior predictions, and customer support tone. Enterprise RFPs today score price, latency, context window, and benchmarks—values aren't on the list. For certain use cases, that's a blind spot.

This piece pulls the AI-in-education conversation back from cheating panic to the real problem: polished output no longer signals ability. The four-tier framework—apprentice (no AI, test bare skills), operator (use AI but explain edits), manager (delegate and verify), decision-maker (choose what problem to solve)—is clearer than most policy talk I've seen. California Community Colleges and Columbia Teachers College already moving this way tells me it's not just theory. But the article doesn't spell out how to redesign rubrics or handle teacher workload, so I'd hold off on calling it a roadmap. The useful bit is shifting assignments from a single deliverable to an evidence chain: prompt logs, verification steps, and a responsibility statement. That makes copy-pasting AI output hard to score well, which beats detector arms races.