Gen AI's two FATAL FLAWS that even human babies can do & the industry knows it

Generative AI companies are trying to convince the world—and your boss—that large language models are ready to replace human workers. But there are two fundamental, innate human abilities that generative AIs simply cannot do: they cannot Learn, and they cannot Prioritize.

In this video, I break down why LLMs are frozen in time, why context windows are not a substitute for learning, and how the inability to prioritize commands leads directly to severe safety failures, jailbreaks, and runaway scripts.

Patreon: https://www.patreon.com/cw/InternetOfBugs

// Sources and Resources
// AI company pushes AI therapist
https://fortune.com/2025/05/16/microsoft-ai-copilot-mustafa-suleyman-gen-z-therapist/

// @3blue1brown on how LLMs work
https://www.youtube.com/watch?v=LPZh9BOjkQs

// Dr Simon Prince on @MachineLearningStreetTalk
https://www.youtube.com/watch?v=sJXn4Cl4oww

// Meta AI researcher has AI delete her inbox:
https://www.pcmag.com/news/meta-security-researchers-openclaw-ai-agent-accidentally-deleted-her-emails
https://x.com/summeryue0/status/2025774069124399363

// LLMs can't keep track of long contexts
https://arxiv.org/abs/2307.03172
https://arxiv.org/abs/2510.10276v1
https://arxiv.org/abs/2412.10079
https://arxiv.org/abs/2410.14641
https://www.alphaxiv.org/abs/2502.17204v2
https://arxiv.org/abs/2411.07037
https://tianpan.co/blog/2026-04-14-the-instruction-position-problem
https://arxiv.org/abs/2407.03651v1

// Many shot Jailbreaking
https://www.anthropic.com/news/many-shot-jailbreaking
https://arxiv.org/html/2404.01833v3
https://arxiv.org/html/2505.19773v1
https://repello.ai/glossary/many-shot-jailbreaking

// The bad things that happen when AI's try to learn in public
https://aiinstitute.hbs.edu/platform-digit/submission/garbage-in-flaming-dumpster-fire-out-the-cautionary-tale-of-tay/
https://ai.meta.com/blog/blenderbot-3-a-175b-parameter-publicly-available-chatbot-that-improves-its-skills-and-safety-over-time/
https://www.bbc.com/news/technology-62497674
https://www.cnn.com/2022/08/11/tech/meta-chatbot-blenderbot
https://www.cnn.com/2025/07/12/tech/xai-apology-antisemitic-grok-social-media-posts
https://www.bbc.com/news/world-asia-china-40815024
https://www.npr.org/sections/alltechconsidered/2016/03/24/471757178/microsoft-chatbot-snafu-shows-our-robot-overlords-arent-ready-yet
https://www.seattletimes.com/business/microsoft/microsoft-takes-chatbot-offline-after-it-repeats-offensive-remarks/
https://www.wbur.org/endlessthread/2022/12/02/bots-tay-microsoft
https://hai.stanford.edu/news/be-careful-what-you-tell-your-ai-chatbot


00:00 - The Two Things AI Will Never Do
01:30 - Who is Carl? (Internet of Bugs)
02:08 - Flaw #1: Generative AI Cannot Learn
02:54 - How Context Actually Works (and Forgets)
04:16 - Why Public Models Must Stay Frozen
05:35 - The Limits and Dangers of Context Windows
07:00 - Flaw #2: Generative AI Cannot Prioritize
07:16 - Real-World Failure: "Confirm Before Acting"
08:40 - Why LLMs Cannot Follow Hierarchies or Rules
09:34 - Compounding Risks: The Multi-Shot Jailbreak Danger
10:54 - The Industry Knows This (Selling the Next Breakthrough)
11:44 - Philosophy, Programs, and Prompts Podcast

#generativeai #ArtificialIntelligence #SoftwareEngineering #TechHype #InternetOfBugs #AIEthics #CyberSecurity