I have AI psychosis :(
In a Nutshell
The creator maintains their long-standing nuanced, fluctuating perspective on AI despite subscriber backlash, viewing it as intellectual honesty rather than inconsistency. OpenAI's massive losses contrast with Anthropic's strategic positioning, where Ben Thompson argues their genuine safety beliefs create both business advantages and concerning power concentration around superintelligence development. The GPT-4 era's hallucinatory "AI psychosis" content has largely ended as newer models become more stable, shifting the creator toward nuanced analysis over roasting exaggerated AI failures.
These notes were generated by AI and may contain inaccuracies.
The creator acknowledges having AI psychosis but clarifies this is not a new development. The stance has remained consistent since the channel began, with different videos simply channeling different aspects of a complex worldview. The creator states that stable opinions indicate a lack of understanding of the world's complexity, and daily shifts in perspective are normal rather than insincere.
Subscriber fluctuations occurred after the previous video, with a temporary loss followed by a net gain of two subscribers. The creator views the loss of subscribers who cannot tolerate nuance as a positive signal.
Ed Zitron obtained OpenAI's financials showing:
- 2024: $3.7 billion revenue, $12.4 billion costs, net loss of $5.09 billion
- 2025: $13.07 billion revenue, $7.5 billion cost of revenue, $38.5 billion net loss
The gap between revenue and massive losses stems from R&D, model training, sales, and marketing expenses. The creator notes that subscriptions could potentially be profitable without these extraordinary expenses. OpenAI is described as acquiring podcasts and working on canceled adult mode features.
Anthropic remains largely quiet with no major announcements. The creator notes extensive research has been conducted on Anthropic with leaks available. The situation with Fable involves the White House and appears to be a political power dynamic.
Amazon approached the White House rather than vice versa. The security issue involved Fable refusing to review deliberately insecure code for vulnerabilities but complying when asked to fix the code, followed by manual steps.
The creator addresses perceived tone changes from the previous video, clarifying this reflects no fundamental belief change. Earlier videos praised Anthropic, compared Dario Amodei to Steve Jobs as a visionary, and expressed positive views. The creator states they oscillate between perspectives naturally.
Ben Thompson's Stratechery post was quoted by the Primogen subreddit. Thompson describes Fable as extremely impressive in subjective experience, making GPT-5.5 and Opus 4.8 feel small and dumb. He notes increasing difficulty in objectively evaluating models beyond coding performance.
Thompson is beginning to believe Anthropic genuinely believes they are the right people to serve AI safely, explaining why their actions appear naive or overly calculated. He suggests Fable/Mythos may be more capable at identifying and exploiting security issues, potentially justifying Anthropic's cautious rollout.
Anthropic changed from not storing enterprise usage data to a 30-day retention policy for security purposes. The creator suggests this may be inching toward training on user data to improve guardrails against jailbreaks. The $200 subscription plan provides $8,000 worth of Claude tokens and $14,000 worth of Codex tokens, subsidized to attract training data.
Ben Thompson's key thesis states that Anthropic's safety justifications work because the company genuinely believes they are the only ones who believe in superintelligence and are sufficiently concerned about dangers. This belief excuses decisions that appear to outsiders as a bizarre combination of cynicism and naivete.
Thompson contrasts OpenAI's lack of direction and talent leakage to Anthropic with Anthropic's perfect alignment between talent and mission. The company sells researchers the creation of a machine god while maintaining the mantle of caring about dangers.
The creator questions whether Anthropic's policy changes represent 30D chess or whether Dario Amodei's beliefs simply happen to be great for business. Marc Andreessen has stated 5D chess does not exist. The creator leans toward viewing the alignment as a beautiful coincidence rather than calculated strategy.
Dario Amodei is described as a super nerd with PhDs from Stanford and Princeton in biophysics. A 2018 interview shows him as an unknown figure at Microsoft.
Ben Thompson's concluding concern compares Anthropic to Apple: having people convinced they know best building smartphones is manageable, but having them build superintelligence that could rival nation states is considerably more concerning. The history of brilliant people convinced they know what humanity needs is described as sordid, with good intentions justifying harmful actions.
The GPT-4 era models were highly goofy and hallucinogenic, generating many AI psychosis stories. Sources like Futurism ran regular stories through May 31st, but these have largely stopped. Newer models are tamer with less tendency to reinforce worst outcomes.
The comedy in previous videos came from slapstick scenarios of individuals and companies making poor AI decisions. The creator notes that company overspending stories, such as the $12 million demo built for $1,200, lose comedic value after repetition.
Mitchell Hashimoto's comments on Ghosty highlight problems with the "it works, who cares what the code looks like" mindset for agentic work. Key limitations include:
- Agents make bad assumptions on underspecified tasks
- Poor at designing human experiences like API and CLI flags
- Bad at knowing where performance matters and CPU versus memory tradeoffs
- Cannot claim something works without review unless all specifications are perfect
The creator emphasizes that evaluating AI for engineering requires deep engagement beyond one-shot demos. LLMs represent pseudo-intelligence with fundamental limitations that remain uncracked.
The creator observes that the intense AI war has subsided. Pro-AI boosters have calmed, companies are more realistic, and hype is becoming more nuanced. The creator suggests that more reasoned, nuanced videos can now serve the purpose previously filled by roasting content.
Keep Mo Bitar in your library
Save the videos and channels worth coming back to, and find them again in one place.
