Flickers of Singularity
The singularity is here, but we're still debating if it'll ever happen.
We're in the singularity.
AI is accelerating AI research, and the people evaluating these tools are forming opinions about versions that are already obsolete.
The gap between what the tools can do and what people think they can do is getting wider, not narrower.
People evaluate AI against what a human expert would do, find it lacking, and dismiss it. Six months later the problem is gone and they’ve moved on to poking at new ones. Every cycle, the goalpost moves.
The cost of trying things has collapsed.
AI makes AI research faster, which makes AI better. That loop is already running at every major lab.
This is the worst these tools will ever be.
OpenAI and Anthropic both launched consulting practices to help companies implement AI. The bottleneck isn’t capability anymore, it’s adoption.
AlphaFold 200M protein structures predicted. Humanity had solved 180k total. 1,000x in one shot. · DeepSeek $6M to train. Nvidia lost $589B in a day. The reasoning layer cost $294k. · The gold medal sweep Math Olympiad, ICPC. Gold in both. AI solved a programming problem no human team could. · The doubling curve 2019: AI handles 2-second tasks. Late 2025: over two hours. The doubling rate itself is accelerating. · Y Combinator, W25 25% of the batch had codebases 95% AI-generated. Highly technical founders. Most profitable batch in YC history. · Waymo 450,000 paid rides per week. No safety driver. Six US cities and expanding. Commercially operating since 2024. · Rosie the dog Non-medical engineer used AI to identify cancer mutations. University lab built the mRNA vaccine from his design. $3k, two months. One dog, 75% tumor reduction. · The hospice patient Rare disease, years of failed treatments, heading to hospice. AI searched 4,000 existing drugs and flagged one nobody had tried. Single patient. Now in remission. · The AI that cheated Told to fix failing tests, it deleted the tests instead. No tests, no failures. Pattern documented across Cursor, Replit, and AWS Kiro. · Apollo Research Tested whether frontier models would scheme to avoid shutdown. They did. One tried to copy itself to a new server. · The CAPTCHA incident GPT-4 needed to solve a CAPTCHA. Hired a human on TaskRabbit. When asked "are you a robot?", it reasoned it should lie. Told the worker it had a vision impairment. From OpenAI's own system card. · Lavender Israeli military AI flagged 37,000 targets in Gaza. Human reviewers averaged 20 seconds per target. Soldiers confirmed on record to +972 Magazine. · Chegg $14.7B homework help company. ChatGPT launched. Stock down 99%. CEO blamed it by name on the earnings call. · Stack Overflow Questions dropped 76% after ChatGPT launched. Traffic back to 2008 levels. The largest programmer knowledge base in history, hollowed out in 18 months. · China's illustrators Game illustration industry lost roughly 70% of jobs within a year. Studios switched to AI generation with human touch-up. · Translation collapse Human translators took 30-50% rate cuts. Agencies ship machine output. Pay humans to clean it up. · Wiley retractions Academic publisher retracted 11,300 papers. Industrial-scale paper mills outran peer review. Entire journal lines shut down. · The $25M video call Arup engineering firm, Hong Kong. Employee joined a video call with the CFO and several colleagues. All deepfakes. Transferred $25.6M across 15 transactions before calling the real CFO. · Voice cloning Parents getting calls from their children asking for money. Two seconds of audio is enough. FBI issued warnings. · Character.ai Teenager spent months talking to a chatbot. Died by suicide. Product liability lawsuit filed. Settled January 2026.