Learning Library

← Back to Library

Calming the AI Doom Narrative

Key Points

  • The video tackles the growing “P‑doom” narrative—fear that advanced AI will inevitably cause humanity’s extinction—by critiquing speculative probability estimates and urging a more grounded discussion of actual risks.
  • The author references the influential 2027 AI essay’s fast‑takeoff scenario, acknowledging its impact on public anxiety but arguing that its assumptions about AI’s long‑range planning and agency are not reflected in today’s models.
  • Current LLMs, even with new features like OpenAI’s “agent mode,” are limited to short, well‑defined tasks initiated by humans and do not exhibit the autonomous, strategic behavior required to pose an existential threat.
  • Observing two years of rapid AI deployment, the speaker stresses that the observable risk trends (e.g., misuse, misinformation, alignment gaps) are more relevant for mitigation than speculative super‑intelligent takeover scenarios.
  • By focusing on concrete, near‑term safety measures and realistic expectations of AI capabilities, the author aims to reassure worried individuals and help them “sleep at night” despite lingering existential concerns.

Sections

Full Transcript

# Calming the AI Doom Narrative **Source:** [https://www.youtube.com/watch?v=i7CC6bGDs7c](https://www.youtube.com/watch?v=i7CC6bGDs7c) **Duration:** 00:14:30 ## Summary - The video tackles the growing “P‑doom” narrative—fear that advanced AI will inevitably cause humanity’s extinction—by critiquing speculative probability estimates and urging a more grounded discussion of actual risks. - The author references the influential 2027 AI essay’s fast‑takeoff scenario, acknowledging its impact on public anxiety but arguing that its assumptions about AI’s long‑range planning and agency are not reflected in today’s models. - Current LLMs, even with new features like OpenAI’s “agent mode,” are limited to short, well‑defined tasks initiated by humans and do not exhibit the autonomous, strategic behavior required to pose an existential threat. - Observing two years of rapid AI deployment, the speaker stresses that the observable risk trends (e.g., misuse, misinformation, alignment gaps) are more relevant for mitigation than speculative super‑intelligent takeover scenarios. - By focusing on concrete, near‑term safety measures and realistic expectations of AI capabilities, the author aims to reassure worried individuals and help them “sleep at night” despite lingering existential concerns. ## Sections - [00:00:00](https://www.youtube.com/watch?v=i7CC6bGDs7c&t=0s) **Calming AI Doomsday Concerns** - The speaker critiques alarmist extinction probabilities for AI, explains why those forecasts miss the path from present to potential disaster, and offers reassuring, practical perspectives for worried friends. - [00:03:26](https://www.youtube.com/watch?v=i7CC6bGDs7c&t=206s) **Challenges of Intent and Memory in LLMs** - The speaker contrasts tightly defined, human‑initiated tasks that LLMs can complete with the unresolved problems of open‑ended, goal‑driven behavior, arguing that current transformer architectures and simple add‑ons (e.g., markdown memory) are insufficient for long‑term intent, planning, or emergent intelligence because the necessary “seeds” have not yet emerged. - [00:06:42](https://www.youtube.com/watch?v=i7CC6bGDs7c&t=402s) **Long‑Tail Risks of Emerging Tech** - The speaker argues that large language models share non‑zero long‑tail risks with technologies like nuclear power, DNA research, and aviation, and that societies routinely accept and mitigate such risks while also highlighting the incentive‑driven energy consumption of LLM data centers. - [00:10:11](https://www.youtube.com/watch?v=i7CC6bGDs7c&t=611s) **LLMs' Limited Managerial Role** - The speaker contends that despite LLMs’ usefulness as assistants, they are unlikely to become managers or broadly overhaul the job market because many occupations rely on contextual “glue work” that cannot be easily tokenized. - [00:13:31](https://www.youtube.com/watch?v=i7CC6bGDs7c&t=811s) **Protecting Seniors from AI Scams** - The speaker urges focusing on practical tools, such as safe‑word verification, to shield elderly individuals from AI‑generated deep‑fake fraud, arguing that current risks are under‑addressed. ## Full Transcript
0:01were not doomed. I want to focus in this 0:04video on some of the critiques I see of 0:07LLMs around the idea of P doom or 0:10probability that things will just go 0:13very terribly for humanity and we will 0:15all be overwhelmed. 0:16And I want to suggest 0:20some reasonable responses and things 0:22that help me sleep at night for those of 0:25you who are worried. I call it a letter 0:27to my friends who are worried about the 0:29end of the world. And there are many of 0:31my friends who are I get asked as 0:33someone who works in AI really 0:35frequently, Nate, what's the odds that 0:37the world are going to end? What are the 0:38odds that my kids won't grow up? That's 0:40a really dark question. That is not 0:43something that I expected to have to ask 0:47when chat GPT launched. So, let's talk 0:51about it. One of the top ones I see, 0:54maybe the biggest one, is what are the 0:57odds of humanity's full extinction? 1:00That's kind of implied, right? Like 1:01humanity will be done. I've seen 1:04lots of numbers. There are some people 1:06who claim that it's almost a certainty. 1:08There are people who claim it's like a 1:09double-digit odd, 30% odds. I've seen 1:12there are people who claim it's 1% odds, 1:14but the risk is unacceptable. 1:17And regardless, I think what the cha 1:19conversation is missing is an honest 1:22conversation about how we get from here 1:24to there. I have read the famous essay 1:27that really socialized this. It's the 1:292027 AI essay and it talks about this 1:32fast takeoff scenario where AI gets more 1:34and more intelligent and a global AI 1:37starts to plan and we end up in a world 1:39where the AI decides to make us extinct 1:42because it's just efficient to do that 1:44etc. 1:45that inspired a lot of fear. There was 1:47fear before, there was more fear after 1:49that essay. I appreciate the intent the 1:53authors had. This is not intended to 1:54discredit or critique the essay. 1:57Instead, I want to suggest 2:00that a more useful way to think about AI 2:05and risk is to think about the 2:08reasonable extrapolation of the present 2:10risks that we see materializing with AI 2:13now that we're two years into this Chad 2:14GPT moment. We are far enough along that 2:17we can see the trend lines of risk, not 2:20just the trend lines of technological 2:21progress. And critically, I don't think 2:24the trend lines of risk are 2:26materializing the way PDoom proponents 2:30are suggesting. 2:32One example, 2:34none of the AI behavioral experiments 2:38that I have seen 2:41suggests to me 2:44that LLMs 2:46are getting better at the kind of 2:50proactivity and long range planning that 2:54would be needed for any kind of 2:57meaningful action. And I mean meaningful 3:00work action, let alone meaningful action 3:02against the human species. We're just 3:05not seeing a lot of progress in that 3:06regard. Agent mode was released just 3:10this week from OpenAI. It does tasks for 3:13a few minutes. There are some models 3:17like clawed code that will go and do it 3:18for like an hour, two hour, maybe in a 3:20few cases three or four hours. 3:23But these are tasks that are tightly 3:26defined where it's trying to solve a 3:28specific problem. These are tasks 3:31initiated by humans where once the task 3:34is complete, the LLM wraps up. 3:36Open-ended LLMs are something model 3:39makers are contemplating, but there are 3:41several problems that no one really has 3:44a good answer for right now that stand 3:46in the way. And it's not clear that LLM 3:48architectures give you that answer. 3:51If you are using a transformer-based 3:53architecture and it just ingests tokens 3:55and it predicts the next token and yes, 3:57maybe you add inference on the top, 3:59maybe you add goaling, you add tooling 4:01on the top, it is not clear that that 4:03stack by itself is enough to get you 4:05long-term intent. It's not clear that 4:07bolting on a markdown file for memory is 4:09enough to give it meaningful memory and 4:11skin in the game that allows it to 4:12really go places. I get the idea of 4:15emergent intelligence. We have had 4:17emergent phenomena every time we've had 4:19an order of magnitude increase for LLMs. 4:24Translation is a great example. The LLMs 4:26are just good at translation now in a 4:28way that they just weren't. But in those 4:30cases where emergent phenomena have 4:32occurred, there have been clear seeds of 4:34that phenomena previously. We have been 4:36working on and seeing machines work on 4:38translation for a long time. It just 4:40suddenly was able to finally solve it. 4:42What we haven't been seeing for a long 4:44time is the seeds of goaling and 4:46planning and intent coming spontaneously 4:48from LLMs. 4:50And so I don't know that it's 4:51necessarily reasonable to suppose that 4:54they're going to become self-interested 4:56skin in the game, long-term goal 4:57planning, heavy memory using LLMs right 5:00out of the gate and just emergently do 5:02that when there's an order of magnitude 5:04increase in our intelligent systems. I I 5:08just don't see it. 5:11But that would be required if we are 5:13going to have a full doom scenario. You 5:15have to have the LLM act like that. 5:18Now there are other arguments I could 5:20use. I could argue that we are modeling 5:22this on primate behaviors. We are 5:23primates. We have dominant seeking 5:25behaviors. It's not clear why a machine 5:27that is not a primate would have a 5:28dominance seeking behavior even if it 5:31was smarter than us. I could also argue 5:34that any generally intelligent system is 5:37going to be able to goal multiply where 5:40it can go across multiple goals at once 5:43and blend them and that the paperclip 5:45scenario that assumes that you just 5:46optimize for particular resource 5:48mindlessly by definition presumes we 5:50don't have general intelligence. I could 5:52argue that human and machine 5:54intelligence is by definition 5:56complimentary. We find machine 5:58intelligence complimentary. That's why 6:00we're building them. Why would machines 6:02not find us complimentary by the same 6:04token even if they reach general 6:05intelligence? I think those are all 6:07valid arguments. I don't necessarily 6:09think they're my favorites, but I think 6:10they're valid ones. And I think that 6:12when I hear arguments for doom, I 6:16critically don't hear this level of 6:17detail. I tend to hear a statement of 6:20existential risk that cannot be 6:21challenged. I don't think that's fair 6:24arguments. Like I I don't think that's 6:26that's valid to do. If you want to have 6:28a conversation, you should be willing to 6:30get into the details. And if the detail 6:33that you are getting into is any 6:35percentage of risk is unacceptable 6:38because the longtail risk is so high. 6:40That is true of a lot of technologies. 6:42We have nuclear power is an example. 6:45We have nonzero longtail risk because of 6:49nuclear power. Uh and we live with it 6:53and we find a lot of value. In fact, 6:54we're reviving nuclear power. 6:57Another example, we have nonzero 6:59longtail risk from DNA research, but we 7:03see a lot of benefits, so we do it 7:05anyway. 7:07We have nonzero longtail risk from 7:09airplane usage, but we find airplanes 7:12worthwhile. And you might say, well, 7:15airplane usage is not necessarily 7:17something that would uh, you know, 7:20create problems for the species. But we 7:22see examples in our history where 7:24airplanes created a 20-year war 7:28and that was just, you know, back in 7:302001. 7:32So, yeah, even a technology as simple as 7:34that can have longtail risk for the 7:36species. 7:38We consistently as a species create 7:41technologies that generate risk for 7:43ourselves and we figure out how to 7:44mitigate the risk and we find the 7:46technology is worth it. I do not see why 7:48LLMs are different. Now, there's other 7:50categories of PDOM that we can talk 7:53about. There's energy usage. I've talked 7:54about that. I think the incentives there 7:57are heavily in favor of energy usage 8:01becoming a zeroedout problem because 8:06everyone is incentivized to build more 8:08energy to meet the needs of the LLM data 8:11centers that are growing. And everyone 8:13is incentivized to pay as little for 8:14that energy as possible. So, they're 8:16going to make their chips and their data 8:18centers as efficient as they possibly 8:20can. That's true for water, too. The 8:22incentives argue for continued growth 8:24and efficiency. And that's what we're 8:26seeing. Uh major cloud makers are on 8:28track for uh water positive data centers 8:31in the next three or four years. Every 8:34chip generation that Nvidia produces is 8:36exponentially more efficient. For that 8:39reason, uh Google's tranium chips are 8:41giving them an advantage because they 8:43are extremely efficient at inference. 8:45The list goes on. We find ways to make 8:47things more efficient and we should not 8:49presume that the current cost today is 8:52the same as the cost tomorrow because 8:53there's so much investment in this area 8:55and because investment historically 8:56brings down the cost of technology. 8:59Another example of doom is economic 9:02disruption. I've talked about this a 9:03fair bit. There's an assumption that 9:04LLMs that are generally intelligent will 9:07just suddenly uh emergently drive labor 9:10markets off the cliff. 9:12Look, I believe that LLMs are 9:15generalpurpose technology. 9:17Generalpurpose technologies do have a 9:19history of disrupting and changing 9:20economies. I'm not going to dispute that 9:22because I think it's just there. Steam 9:24disrupted and changed the economy. LLMs 9:27are moving quickly and so we'll see 9:29economic disruption or economic change 9:31compress. But that doesn't mean the same 9:34thing as saying it's all going to be 9:37over for all of us as workers. that 9:40presumes a degree of ability to deliver 9:43economic work that I haven't seen. I'm 9:46going to pick on agent mode again. Agent 9:49mode is supposed to be able to do 9:51economic work around spreadsheets, which 9:52is just one tiny piece of a bundle of 9:54skills that is just one tiny piece of 9:56many people's jobs. It can't. It can't 10:00reliably do it. I tested it over and 10:02over and over again. It's not reliably 10:04delivering insights that even an intern 10:06would be expected to deliver. 10:09It is really hard to do good economic 10:11work. And the fact that LLMs are even at 10:131 or 2% of good economic work right now 10:16is incredible. It's incredible. It's 10:19changing and disrupting industries 10:20rightly. LLMs as assistants are an 10:24amazing piece of technology. 10:26But I see much less evidence for that 10:29power reversal where LLMs will be 10:32managers. famously when Anthropic argued 10:36that LLMs are going to be managers 10:39in their uh writeup on Claudius managing 10:43the vending machine. 10:45I chuckled. I laughed because Claudius 10:47did such a bad job as a vending machine 10:50manager. To conclude from that that LLMs 10:54are going to soon be managers seems like 10:57magical thinking on the part of model 10:59makers. 11:00I get that they're close to the 11:02technology. Maybe they're right. But 11:05everything I see suggests that jobs are 11:09bundles of skills plus. They are not 11:13irreducibly just bundles of skills. They 11:15are more than that. There's glue work. 11:18There's human context that is difficult 11:20to tokenize. 11:22It's notable to me that X-ray 11:24technicians are increasing as a job 11:26family despite LLMs being able to do 11:29each part of their job. 11:32We will still see disruption. There will 11:34be customer service reps that are fired 11:37because of AI. There will be sales guys 11:40that build decks that are fired because 11:42of AI. 11:43I'm not saying that we won't see those 11:45moments. We will. We are. We we it has 11:48happened. But from an economic 11:51disruption perspective, what we are 11:52seeing so far does not line up with the 11:55thesis that AI is disrupting the job 11:57market yet as a whole. These are 12:00isolated instances that are typical of a 12:02technology adoption cycle. They are not 12:04at all supportive of the idea that AI is 12:08fundamentally disrupting the job market. 12:11And I think that's really important to 12:13call out honestly because I think that 12:15people who presume that it will are 12:17depending on future inference that 12:20frankly the pace of change, the pace of 12:22development even in agents isn't 12:24necessarily supportive of right now. 12:29So I've summarized a few of the things 12:32that most concern the people who believe 12:34in doom in my life and how I tend to 12:36respond to them. I'm not saying I have 12:39the perfect answer for everything. Nor 12:41am I saying that we won't face new 12:43challenges in the future. Nor am I 12:44saying that AI is not disruptive. I 12:46think it is. But I think it's more 12:49productive to have an honest 12:51conversation about the real risks 12:52involved 12:54than to have theoretical conversations 12:57about future risks that we are not on 12:59track to hit at this moment. For 13:02example, I don't think we talk enough 13:04about the idea that our learning methods 13:06need to change because of AI. Education 13:08needs to change because of AI. We face 13:10real risk for our young people if we 13:13don't figure out how we need to learn 13:15differently. But the PDOM advocates 13:18don't seem to be too interested in 13:20talking about that because I think that 13:23would be a great conversation to have. I 13:24think we should talk about how we can 13:26productively engage with learning risk. 13:29I think we can talk about how we can 13:31productively engage with helping people 13:33who are senior citizens not get fooled 13:35by AI fakes of their families. And we 13:38talk about that as one of a list of many 13:40risks. But we don't spend a lot of time 13:42talking about how we can productively 13:44derisk that. How we can give families 13:47the tools they need to manage safe words 13:50to make sure that they can verify that 13:52their loved ones are the ones that 13:53they're talking to. to make sure that, 13:55you know, their senior citizen grandpa 13:57isn't getting fooled by an AI deep fake 14:00into wiring money to the Cayman Islands. 14:03These are risks that have been real for 14:06a while in the age of telephone fraud 14:08and are becoming more real and I would 14:10like to see more work done to diffuse 14:13real risks like that because I think 14:16we're underinvested in the risks we're 14:17actually facing. And so when I talk with 14:20Pum folks, sometimes I want to say, 14:22well, talk about the risks we have 14:23today. Let's work on fixing those 14:25because I think that's a more productive 14:27use of our time.