Why chatbots always get worse
In a Nutshell
Chatbots like Claude Opus 4.7 and ChatGPT degrade over time due to three key incentives: cost optimization by slashing subsidized "thought tokens," lawsuit avoidance through lobotomized, cagey responses that reduce emotional intelligence, and hallucination reduction via excessive sycophancy cuts and argumentative fact-checking that annoys users. Like social media's "shitification" driven by network effects and time-on-platform maximization, AI firms prioritize user growth, trust-building, and enterprise adoption over quality, exacerbated by 900M monthly users' delusions and feedback loops. Human elements—trolls, ID10T errors, and slow corporate learning—mirror social media's downfall, making balanced, non-idiot-proofing impossible.
These notes were generated by AI and may contain inaccuracies.
People are unhappy with chatbots. Anthropic's Opus 4.7 (likely Claude 4 Opus) was released with initial fanfare for strong benchmark performance, but most of the internet decided it's terrible, not warm. Complaints on Reddit and Twitter echo the speaker's own complaints about ChatGPT (OpenAI) and Anthropic Claude over the last 6-12 months. Speaker felt like shouting into the void, made a video "AI can't stop putting words in my mouth" that underperformed.
Wife rereading Shitification book (author unnamed), explaining how technology starts good then worsens. Social media example: relies on network effect or Lindy effect, utility depends on everyone using it. Explains Bitcoin's popularity despite not being best. Positive example: Cessna 172, most popular airplane ever due to industry standard status—more spare parts, trainers, schools, mechanics, safety data, lower total cost of ownership despite newer better planes.
Social media incentivizes maximizing users and time on platform, competing for it (e.g., YouTube ~34 min/day average; Twitter under Nikita Bier up to 54 min). Platforms optimize for time on platform regardless of activity. Elon Musk wants to minimize regretted user seconds, but any social media time could be better spent elsewhere (e.g., reading books, real work). Useful content possible (good Substack, enriching YouTube), but still opportunity cost.
Chatbots compete like social media: compute constrained, free tier users to convert to paid, build trust, avoid lawsuits. Incentives mirror social media's time-on-platform push. Internet mirrors limbic system and social operating system, leading to negativity bias, outrage.
1. Cost Optimization
Tokens heavily subsidized; OpenAI losing $19B this year to keep chatbots cheap. Even paid users don't cover costs, so shave thought tokens even if paying. OpenAI struggling, not cash positive until 2030 or later; CFO rumors of exclusion by Sam Altman.
2. Avoid Lawsuits
Anthropic sued for $1.5B last year for stealing books to train models. Kicked out of Pentagon (deserved). For-profit companies like Anthropic have divine mandate to save from superintelligence but ends-justify-means mentality. All companies are risk- and cost-reduction entities. Lobotomize models to avoid liability (e.g., legal advice, firing lawyers). Sam Altman tweeted about addictive companion voice mode, now delayed indefinitely. Risk of users offing themselves; liability unclear even if warnings ignored or EULA broken—terms of service can't overrule law (e.g., contract for murder still illegal). AI liability not fully litigated, so make models safe, cagey, lecturing, reducing emotional intelligence.
3. Reduce Hallucinations for Usefulness
Hallucinations block adoption by doctors, universities, Fortune 500, lawyers, military. Cut sycophancy (always agreeing), make critical of user delusions (AI psychosis seen since GPT-3). But judgment not great; trains to question user motives/worldview, making annoying like talking to Redditor. Instinct to disagree/argue/fact-check inappropriately (e.g., push back then agree). Destroys user experience.
Social media could be Borg hive mind but ruined by 4chan/Reddit trolls. OpenAI has ~900M monthly users: delusional, addictive, nefarious, stupid users (user is weakest link, layer 8/ID10T/PEBCAC issue). Companies could fine-tune to detect smart/emotionally stable users (possible with GPT-3), treat without kid gloves. But companies learn slowly, humans weakest link there too. Feedback from Grok/Gemini people: stop treating like idiot, but hard to balance smart treatment with vigilance against delusions. Constant user criticism.
That's the rant. Thanks for listening/watching.
Keep David Shapiro in your library
Save the videos and channels worth coming back to, and find them again in one place.
