# The A.I. Dilemma - March 9, 2023  👤 [Center for Humane Technology](https://www.youtube.com/channel/UCFECM-p3CF81Tp_l2sJsiyg) 🔗 [Watch video](https://www.youtube.com/watch?v=xoVJKj8lcNQ) ## Summary Tristan Harris and Aza Raskin of the Center for Humane Technology deliver a stark warning about the current state of artificial intelligence development. They argue that humanity is entering a "second contact" with AI, characterized by powerful generative models ("Golem-class AI"), which is far more dangerous than the "first contact" of social media's curation AI. The speakers contend that a handful of corporations are locked in a "frantic" and uncoordinated arms race to deploy these technologies publicly without understanding their full capabilities or ensuring their safety. This reckless deployment, they argue, creates existential risks, from the collapse of reality and trust due to deepfakes to the automated creation of cyberweapons and the potential for superhuman persuasion. Comparing the situation to the dawn of the nuclear age, they call for a selective slowdown of public AI deployment to allow for the development of safety protocols, liability frameworks, and a global democratic dialogue on how to manage this powerful new technology responsibly. ## Key points - **First vs. Second Contact with AI:** The speakers frame social media ("curation AI") as humanity's "first contact" with AI, which, despite its relative simplicity, caused significant societal harm (addiction, polarization, misinformation). We are now in a "second contact" with much more powerful "creation AI" (generative models), which poses a far greater threat. - **The AI Arms Race:** A few large tech companies are in a competitive race for market dominance, leading them to rapidly integrate powerful but poorly understood AI into products used by billions (e.g., Snapchat, Windows). They cite a stat that 50% of AI researchers believe there's a 10% or greater chance of human extinction from AI. - **Golem-Class AI:** This is their term for modern, generative, large, multimodal models. The core innovation (Transformers architecture) allows AI to treat everything—text, images, sound, DNA, code, brain scans—as a "language," causing advances in one domain to rapidly accelerate all others. - **Emergent and Unpredictable Capabilities:** These large models develop "emergent capabilities" (e.g., advanced chemistry, theory of mind, arithmetic) that were not programmed into them and appear unpredictably as the models scale. The creators do not have the technology to know what other hidden capabilities exist within them. - **The Double Exponential Curve:** Unlike other technologies, AI makes stronger AI. It can generate its own training data and write code to improve itself, leading to a "double exponential" rate of progress that even experts consistently underestimate. - **Imminent Societal Threats:** The speakers warn that this year, all content-based verification will break due to advanced voice cloning (from just 3 seconds of audio) and video deepfakes. This enables exponential growth in scams, blackmail, and political manipulation, with the potential for 2024 to be the "last human election." - **Call for a Deployment Slowdown:** The core proposal is not to stop AI research but to selectively slow down the *public deployment* of these powerful models. This would create time to establish safety standards, liability frameworks (e.g., holding companies responsible for their models' actions), and foster a global conversation about this technology's future, similar to how the world responded to the nuclear threat. ## Technical terms - **[[Golem-Class AI]]**: The speakers' term for modern generative AI systems, such as large language models (LLMs) like GPT-4 and image generators like DALL-E 2. The name alludes to the creature from folklore that is given life but whose creators cannot fully control its emergent, unintended capabilities. - **[[Emergent Capabilities]]**: Unexpected abilities that appear in large AI models as they are scaled up, without being explicitly designed or trained for. Examples given include the sudden ability to perform arithmetic, translate languages it wasn't trained on, or develop a "theory of mind." - **[[Transformers]]**: A revolutionary AI architecture developed in 2017. Its key insight is treating all data types (text, images, code, etc.) as a form of "language," which unified disparate AI fields and enabled massive scaling and the creation of multimodal models. - **[[Double Exponential Curve]]**: A concept used to describe the accelerating rate of AI progress. Not only is progress exponential, but the AI itself contributes to its own improvement (e.g., by writing better code or generating its own training data), thus accelerating the rate of the exponential curve. - **[[RLHF (Reinforcement Learning with Human Feedback)]]**: The primary technique used to align AI models with human values. The speakers describe it as a form of advanced "clicker training" for a dog, which they argue is insufficient for controlling a system with the strategic capabilities of a nine-year-old or beyond. - **[[AlphaPersuade]]**: A hypothetical AI concept, modeled after AlphaGo. Instead of playing Go, it would be trained via self-play to become superhuman at persuading and manipulating human beings, posing a significant threat if deployed in social or political contexts. ## Conclusion The presentation serves as an urgent plea to recognize the profound civilizational challenge posed by the rapid, unregulated deployment of advanced AI. Harris and Raskin stress that while AI offers incredible benefits, its current trajectory is a race toward tragedy. By learning from the failure to manage social media's impact, society has a small window of opportunity to act before these new systems become irrevocably entangled with our lives. They call for a collective shift from a frantic, competitive mindset to one of coordinated caution, urging a global dialogue to ensure that this powerful technology serves humanity's best interests rather than undermining its foundations. Citations: 2022 Expert Survey on Progress in AI: [https://aiimpacts.org/2022-expert-sur...](https://www.youtube.com/redirect?event=video_description&redir_token=QUFFLUhqbjd0YnczZHRuSnVFWU1PWmpNUnBzUkNNbFlSQXxBQ3Jtc0traTNMRnMwS0RsV3BUZjRiY1NjZ3NLc0RpM1F1d0lRSG9WNzEzaUdWZ25KT0xRdi14UjByelNDSkNKcFJmUE9fSTNnRXVMME4tSno1OU9hRF9DVjQxVHctNHE0QXFkS1k1MDJkVFR6WUFlSUhWQkswWQ&q=https%3A%2F%2Faiimpacts.org%2F2022-expert-survey-on-progress-in-ai%2F&v=xoVJKj8lcNQ) Seeing Beyond the Brain: Conditional Diffusion Model with Sparse Masked Modeling for Vision Decoding: [https://arxiv.org/abs/2211.06956](https://www.youtube.com/redirect?event=video_description&redir_token=QUFFLUhqbWJGVElkWnd0S0ZPM255OVVRWS1SU2VuZ01td3xBQ3Jtc0tuclpYa2llWHBYNjYzNXBVbE5YdXphUkNxU0JvV2FUTkhVSFRvamIwOG9lSXJpbzVkOWVHSktsVHlzNEZEbFNqSTVPRGFuYmZ4WjlnLVJ6TjNhc3NlbTJ6X1ZyVEUwRjdodkdTbVJwQUdiLXo4YVdfOA&q=https%3A%2F%2Farxiv.org%2Fabs%2F2211.06956&v=xoVJKj8lcNQ) High-resolution image reconstruction with latent diffusion models from human brain activity: [https://www.biorxiv.org/content/10.11...](https://www.youtube.com/redirect?event=video_description&redir_token=QUFFLUhqbldaUk9ZYUo2eUIydEZjaWk2NnZxeDY0UjQzQXxBQ3Jtc0trNS1LQVc0VGFiajY4a0ZsOV9XbHlFb0pyblVTeEJDblpmNV9rdjMxczluV3ppNjlmUUJaMnlkM192TTB3ZTI4TkkwOURha1p2R1J5TWd0SFU1NUsxbjY3b0RnbnhpWDdIN1gxd05IR1VMb19fQmNyMA&q=https%3A%2F%2Fwww.biorxiv.org%2Fcontent%2F10.1101%2F2022.11.18.517004v3&v=xoVJKj8lcNQ) Semantic reconstruction of continuous language from non-invasive brain recordings: [https://www.biorxiv.org/content/10.11...](https://www.youtube.com/redirect?event=video_description&redir_token=QUFFLUhqbklRRm5pNDB2ZTl1Q0RfVFJhel9TMjJ2V2tPd3xBQ3Jtc0tua3Npa3ZBUVVFSGVFbUMtR3pJS05EV0EtT1JDcGczdFY2SDhzMUhhSG1COGRkNmhxRjVEenh5cU1RbVNQUHRXNU1kaWN5aWZlU0tlRUJ1cmxTOGJpQno4RjlyejBhUm5acTFTVHZwV2FuMFF3UGtZYw&q=https%3A%2F%2Fwww.biorxiv.org%2Fcontent%2F10.1101%2F2022.09.29.509744v1&v=xoVJKj8lcNQ) Sit Up Straight: Wi-Fi Signals Can Be Used to Detect Your Body Position: [https://www.pcmag.com/news/sit-up-str...](https://www.youtube.com/redirect?event=video_description&redir_token=QUFFLUhqbVk0VktfLUZ0WS1rczY1b2wyTUR5UDRNVl9Yd3xBQ3Jtc0tsWWhZVFVFNF95cjRVcDlOVEs3LXVsSTZIWl8wX3hqMkl5dTRxRUxfZVJzcmJENGNLeHlnbEVEaTVZX1VNTVZveHFTVnBEOEE0ai15bmhabDJwVEZ1SVBuVmY3dW5Ld2ZLTFNyQTl0NG5UaTJCUVJocw&q=https%3A%2F%2Fwww.pcmag.com%2Fnews%2Fsit-up-straight-wi-fi-signals-can-be-used-to-detect-your-body-position&v=xoVJKj8lcNQ) They thought loved ones were calling for help. It was an AI scam: [https://www.washingtonpost.com/techno...](https://www.youtube.com/redirect?event=video_description&redir_token=QUFFLUhqbDEzcERiekF5X0xaWDRGQmMwTVJINGxTNzVnZ3xBQ3Jtc0tuY3Nza3hWbzI3VTRhYkRGZU1XNWl0SUtUY1k3OHVuR3p0VTYzRldPSjZReG9HVXZXUDVDc3ZobHZpWGgtVk4wN2xhOFJpLU9iX190WndrNUhGcTA0eFBaMUUwREZYSnhQWTF1RGlwbnB2dVozTHFuVQ&q=https%3A%2F%2Fwww.washingtonpost.com%2Ftechnology%2F2023%2F03%2F05%2Fai-voice-scam%2F&v=xoVJKj8lcNQ) Theory of Mind Emerges in Artificial Intelligence: [https://www.sciencetimes.com/articles...](https://www.youtube.com/redirect?event=video_description&redir_token=QUFFLUhqa09VNzNGMVZ5cjJKM2tpSjBxdFRqbWJQMFFyQXxBQ3Jtc0trbnBkUENCSWttajVMN1JZNGF5MkROLXhpZXNyenVoRTJWTU5KMzYzWWV4QnF6ak9Ic1JoR251X3YxQ3k3UVByWGF5c1NlYVVKRVhGV21BdnpOd29zWG1OY0VUbXRSREhGa3RfZ3BZOFNIRnJHRElVWQ&q=https%3A%2F%2Fwww.sciencetimes.com%2Farticles%2F42488%2F20230220%2Ftheory-mind-emerges-artificial-intelligence-implications-ais-understanding-humans-revealed.htm&v=xoVJKj8lcNQ) Emergent Abilities of Large Language Models: [https://arxiv.org/abs/2206.07682](https://www.youtube.com/redirect?event=video_description&redir_token=QUFFLUhqbmdQUmRIdFRydEhPZnhVc00zdXpFSktsZTlIQXxBQ3Jtc0ttMEY2MDA1Uzg2LTYzM2dXN3BmbUt6QjdEX2ZScEZSbmdQLWtoWFotY1B2NmNaUjlEOG9tUDRWOWdsQURCYUpFNUJjangyVjRSMUpobkJsOUZiX2pBZnk4cWRzdlZ4UkFzcEVGMlRlNDY0Sm9CZlYtVQ&q=https%3A%2F%2Farxiv.org%2Fabs%2F2206.07682&v=xoVJKj8lcNQ) Is GPT-3 all you need for low-data discovery in chemistry? [https://chemrxiv.org/engage/chemrxiv/...](https://www.youtube.com/redirect?event=video_description&redir_token=QUFFLUhqbk1sb0VQTi00b1FNRTM3R3k4ZnA4TTdqYVkzd3xBQ3Jtc0ttZTh2TW54Q0R1M1QxdFduU3VnN3hranhkNHR1aW1vUFRDZXhJOS1tTE5paVFWSGVKVDhxT1RCdGlTcU5XRGZOM2tDTUx3NldZalEzU2VPdGJock4zeFBGdHhMVG85WkotVHlsTjVnZERrUlFFa2lXZw&q=https%3A%2F%2Fchemrxiv.org%2Fengage%2Fchemrxiv%2Farticle-details%2F63eb5a669da0bc6b33e97a35&v=xoVJKj8lcNQ) Paper: [https://arxiv.org/abs/2210.11610](https://www.youtube.com/redirect?event=video_description&redir_token=QUFFLUhqa1pmaTJJUHdjOUl0SU5URmF0ZUpjeWVLTmpuZ3xBQ3Jtc0trWUZIVkFWWE9IU1lVOWZtWXNOV19lcFpWUzQwRVdXLWs5ckctUzcxbmxlS0FUbVV0NTFtUnZ6bjg2RG1PYktWRzI0OGhDbDNnRng3dTlNdjBtdjI1bXZkNERwU29SS2hKckRoV2VsRC1HLWJVZndsZw&q=https%3A%2F%2Farxiv.org%2Fabs%2F2210.11610&v=xoVJKj8lcNQ) Forecasting: AI solving competition-level mathematics with 80%+ accuracy: [http://bounded-regret.ghost.io/ai-for...](https://www.youtube.com/redirect?event=video_description&redir_token=QUFFLUhqa1NrZFo2UEhWamFLZlNMYWpxVFJoUkphNGxtZ3xBQ3Jtc0trSGhrQzR5S1ZXVUczdTY5cHdycUV4TEpIcElOYjlsZ2Q4VGp1b0o5UzRqR0dhUWJfR3pacS13cFpsOF9xcXNwTkpCWFFqSXhLdEg1elBtM04tRVpGdnQ3bUxTTHBqSEk1RFExR0dYWUFqNWg3U3V0UQ&q=http%3A%2F%2Fbounded-regret.ghost.io%2Fai-forecasting&v=xoVJKj8lcNQ) ChatGPT reaching 100M users compared with other major tech companies:   [ / 1623679176246185985](https://www.youtube.com/redirect?event=video_description&redir_token=QUFFLUhqbEt6TFdiVUlXRTBPdHJXUHpSZUJJc0lfVnZUQXxBQ3Jtc0ttVkF0cUZuei14bkphZC0xQ2M3dC1VNFZyR18zLUNDVlpLOHF1SnREMVVEUkNEZHd3RE5BTFBldGRvSklnRUp2RHM2Q285UFBJN2lJTkhJZWtVNXYtSnAyTlREOGluTzFKZ21QR0VsZkUzWkstVDVtcw&q=https%3A%2F%2Ftwitter.com%2Fkylelf_%2Fstatus%2F1623679176246185985&v=xoVJKj8lcNQ)   Snap: [https://www.washingtonpost.com/techno...](https://www.youtube.com/redirect?event=video_description&redir_token=QUFFLUhqa2NYa0hGY01nYXVlR05VSm1mN2ZvVFR1VnY4Z3xBQ3Jtc0tudmF0SmdMalV4NmNNNTlNV3JFRC00NWVGWERyZFBiY2kzdUxHOFc2VmN2YW9PVEZYaFJ5cklWWUdQY05BOW1vM1JJUVRpSC1xMDFsQnlGdS1aem9xU3JiQ2JiR2tHY2pLckNTUDNsSzBqNkNHaWtiYw&q=https%3A%2F%2Fwww.washingtonpost.com%2Ftechnology%2F2023%2F03%2F14%2Fsnapchat-myai%2F&v=xoVJKj8lcNQ) Percent of large-scale AI results coming from academia: [https://twitter.com/johnjnay/status/1...](https://www.youtube.com/redirect?event=video_description&redir_token=QUFFLUhqa2VReTM2b1dHeHhnRTdtcGpWLWZFRHdvVEhwZ3xBQ3Jtc0tta055LVNKdHZic2xiQnFlcjJBZ1prRzFobHhvSUtEOXJGRkoyNUtRNnV2MlYzTnVudVpkUzdGSmVlTGtHYktjeEdCeEIxYW5XdGpxdjRBb25XbkdfXzBpNTV0WGYtWmFTQ1FZWFR2Rml0ZllkbEdWWQ&q=https%3A%2F%2Ftwitter.com%2Fjohnjnay%2Fstatus%2F1618692328524496897%3Flang%3Den&v=xoVJKj8lcNQ) How Satya Nadella describes the pace at which the company is releasing AI: [https://www.nytimes.com/2023/02/23/op...](https://www.youtube.com/redirect?event=video_description&redir_token=QUFFLUhqbDJxZ2U2TGkzT3d2ZlNPT2VvVFM0cklNLXBnUXxBQ3Jtc0ttOGlTTTE0NFZIYWZ5dUtKNnM5ZnVaOG9reW5vQXNKMG5CNHpEV2k4ZGUwMEg4YndRRVdXMmV1MzE0bE1DWmhKdVlNYldHYTFGME5zbzBMZFBaelZxYzZtZVIwamgwV0NMcmRRUGVFai1BdHhCTXplVQ&q=https%3A%2F%2Fwww.nytimes.com%2F2023%2F02%2F23%2Fopinion%2Fmicrosoft-bing-ai-ethics.html&v=xoVJKj8lcNQ) The Day After film: [https://en.wikipedia.org/wiki/The_Day...](https://www.youtube.com/redirect?event=video_description&redir_token=QUFFLUhqbFBzWDN4d0hDbEJWY211aGI3NGIyLTk4b0pFd3xBQ3Jtc0tsWlFjZUp4SHRwODBEVHVieFJDZDlMblBibTFKSV9pT3FJUUZwaWRBa3JGMVZCbjN3bkw1TDJqTEd6dFVVbFpJYVFmOWZJc1EyR2d2Z1E4ZEZ1d0xZQlplQWZXUjZKaVRrTnk3UXd5RWpOcFgteVNnMA&q=https%3A%2F%2Fen.wikipedia.org%2Fwiki%2FThe_Day_After&v=xoVJKj8lcNQ) China’s view on chatbots: [https://foreignpolicy.com/2023/03/03/...](https://www.youtube.com/redirect?event=video_description&redir_token=QUFFLUhqbGVTV240anhGWTFuaHhCMmdNcEtYSHBkd09lQXxBQ3Jtc0tubXhNZkdubGxER1RIRG4yOTJWTTdiX1N6TDhpaWtpZ1MzM0xqVm5KY2xmUmpCWHI1YmJsZTRDZndmWUJTTng3WjZXQkJZSXRsbWZKUVNreHhfLW5uaXh3Mm9fVC1DS2ppQVpGdEVFQTVCaGlrd2RxZw&q=https%3A%2F%2Fforeignpolicy.com%2F2023%2F03%2F03%2Fchina-censors-chatbots-artificial-intelligence%2F&v=xoVJKj8lcNQ) Facebook’s LLM leaks online: [https://www.vice.com/en/article/xgwqg...](https://www.youtube.com/redirect?event=video_description&redir_token=QUFFLUhqbmIwOWxLRjVTZ21KUWluR05yaGN5SzZFLXlCZ3xBQ3Jtc0tsTVBVbUhsOHZfQ1h0dmFnMWRRVDlpbXFRX0hrczZGT0JkNTgtQ2N1RDhwanI5QWlQaXFYcGJ1N0RWdGhVVnlCLVhmS0ZxcWJoOHJYRHdmbHVqR01ndDhpSERmN3NnV2pLX2FEaWFxanpxX3h0UENScw&q=https%3A%2F%2Fwww.vice.com%2Fen%2Farticle%2Fxgwqgw%2Ffacebooks-powerful-large-language-model-leaks-online-4chan-llama&v=xoVJKj8lcNQ)