Other (17)

• Theo highlights Anthropic's recent "self-inflicted wounds," including performance issues with Opus 4.7, flawed system prompts, a poorly received desktop app, and mysterious user bannings.
• Ben notes Figma's stock dropped significantly after a brief collaboration with Anthropic, falling 85% since its IPO and an additional 15 percentage points after the announcement.
• Theo observes Anthropic's "psychosis around safety," where Opus 4.7 refused basic cryptography prompts and recommended Sonnet 4, actively degrading product quality and service reliability.
• Theo suggests Anthropic "lobotomizes" models within first-party Claude products, while Opus 4.7 via third-party APIs completed complex security-adjacent tasks without issue.
• Theo identifies intentional model "gimping" (Opus 4.7 performs worse than 4.6 on safety benchmarks) and flawed system prompts as blocking layers; Ben adds a "babysitter" model on Claude.ai to enforce restrictions.
• Theo and Ben agree Anthropic employees use different, presumably superior, internal versions of products and models compared to what external users receive, impacting user experience.
• Theo created T3 Code, an open-source GUI wrapper, which users were banned for despite adhering to Anthropic's terms; Ben notes Anthropic previously adjusted API billing based on keywords like "OpenClaw."
• Ben's OpenClaw heartbeats incurred $120/month due to inefficient caching; Theo explains LLMs recompute context for each token, making proper caching crucial for cost efficiency during Anthropic's "compute crisis."
• An AMD AI head's audit showed Claude Code's degradation: thinking redaction, 173 "stop hook" violations after March 8th, doubled user frustration, and a massive cost increase from $26 to $42,000.
• Anthropic confirmed Sonnet 4 requests were incorrectly routed to a "dumber" 1 million context window version (up to 16% of requests) instead of the 200,000 token version, confirming an intelligence difference.
• Theo theorizes Anthropic forces heavy users to the dumber 1 million context version (now default and free) on Amazon/Google TPUs, reserving superior Nvidia GPUs for researchers during a compute crisis.
• Theo suggests Anthropic's obfuscation of thinking traces, combined with poor engineering, means crucial reasoning data isn't properly available to models, leading to dumber decisions and reduced performance.
• Anthropic's new tokenizer for Opus 4.7 uses 1.47x more tokens (vs. their stated 1.35x), an unprecedented dot-update change. They also reduced cache TTL from an hour to five minutes.
• Theo reports T3 Chat's caching, despite following recommendations, cost over 50% of their $30,000 monthly bill ($15,000) just for cache writes, suggesting poor economic viability.
• Theo attributes Anthropic's issues to a cultural problem stemming from its founders' anti-engineering bias, resulting in "garbage software," lack of transparency, and a "holier than thou" attitude.
• Theo asserts open-source benchmarks like SWEBench are "polluted" because models can recreate commits by hash, making them unreliable for accurately testing new model performance.
• Theo proposes tracking obtuse metrics across AI labs: Anthropic's product releases (4 since transparency promise), Google's important departures (4 since Gemini 3.1 Pro), OpenAI's rate limit resets, and XAI's account bans.