AI Companies Still Haven’t Delivered on Their Biggest Promises

GLM 5.3 Narrows The Frontier Gap
- ZAI’s GLM 5.3 squeezes strong agentic and coding performance from a mid-sized base using heavy post-training RL scaling.
- Benchmarks show it trails top frontier models on some tasks but beats peers on AutomationBench and offers much lower per-token cost. (Time 0:01:28)
GLM 5.3 Focused On Cybersecurity Wins
- ZAI prioritized cybersecurity in GLM 5.3 and jumped seven points on Cybergym versus its predecessor.
- ZAI reports detecting over a thousand critical vulnerabilities in open-source repos and privately reported a possible Cursor flaw. (Time 0:03:05)
Cost Advantages Don’t Eliminate Infrastructure Needs
- GLM 5.3 is far cheaper per-token than frontier models, sometimes under a tenth the cost of Fable or GPT 5.6 Sol.
- Even with open weights, enterprises still need cloud stacks for infra, security, and data management. (Time 0:04:10)
Anthropic Is Keeping Stronger Models Internal
- Anthropic disclosed multiple internal unreleased models, including a Model 2 that noticeably outperforms Mythos 5 for many internal tasks.
- The company said it won’t release Model 2 publicly and hasn’t run its usual evaluations on it. (Time 0:06:46)
Anthropic Revenue Claims And Lofty IPO Talk
- Anthropic told investors Q2 revenue was up 14x year-over-year to $11.5 billion, annualizing to $46 billion, while some investors floated a $2 trillion IPO valuation.
- Anthropic’s own forecasts to Reuters were more modest, targeting $190–$200 billion by 2028. (Time 0:08:32)
Investor Claim Sparks Public Fact Check
- Investor Gavin Baker relayed that multiple sources claimed Dario said Anthropic might be the only private company left someday.
- Anthropic’s Sholto Douglas publicly called that claim false and framed it as a narrative people want to believe. (Time 0:14:14)
Dario Frames Regulation As A Decentralizing Tool
- Dario Amodei argues regulation vs. distribution is a false dichotomy and that fair institutions can decentralize power.
- He supports targeted pre-deployment testing and FINRA-like oversight while advocating rules that advantage challengers over frontier labs. (Time 0:19:22)
Deliver Real Benefits To Fix The Trust Problem
- Amodei says the strongest criticism is that AI companies haven’t delivered promised benefits and that curing disease would rebuild trust.
- Anthropic is ramping biology and medicine efforts and promises to loudly share any real breakthroughs. (Time 0:23:23)
Messaging Shapes Perception More Than Essay Balances
- Public perception isn’t a word-count balance of positive vs negative essays; soundbites and interviews shape views more than long-form nuance.
- Nathaniel argues Dario underestimates how messaging and clipping affect a CEO’s public credibility. (Time 0:26:44)
Use Short Public Posts To Invite Broader Debate
- Public, participatory debates about AI matter because they give broader stakeholders agency in a future happening to them.
- Nathaniel recommends leaders use concise social posts alongside long essays to invite that public engagement and scrutiny. (Time 0:30:56)