Activity
39K followers
-
Arvind Narayanan posted thisI use AI heavily for validating ideas and prototyping things, but I’m much more cautious about AI use for anything that I need to ultimately publish and stand behind. One unfortunate consequence of speeding up prototyping more than execution is that I now have a much bigger pile of interesting, validated ideas that I simply don’t have time to properly get started on. This has been bothering me but I’ve learned to come to terms with it. What would be even worse is starting tons of projects with AI, making partial progress, but never completing and publishing them. Project management skills like ranking projects by priority and making explicit go/no-go decisions were always a good idea, but have now become essential in order to avoid losing our sanity.
-
Arvind Narayanan posted thisMy best rough estimate of the fraction of the scientific literature that’s wrong — flat-out wrong, you’d-be-nuts-to-rely-on-this wrong — is 80%. The problem is that the normal process of scientific self-correction takes time and happens through follow-up work that fails to reproduce experiments, finds computational errors, challenges assumptions, and so on. This is a big opportunity for AI-for-science tools. Agentic interfaces to the scientific literature should be designed so that when they look up a paper they should explicitly look for later work that challenges it. A harder but potentially far more impactful intervention is to use agents to auto-annotate papers in the pre-training corpus based on what other later sources have said about them, as well as re-weighting and other tricks. Of course AI developers shouldn’t become the arbiters of scientific truth but the fact that online information is of widely varying quality is already central to LLM pretraining recipes and agentic retrieval; my point is that there are responsible ways to apply this insight to the scientific corpus as well.
-
Arvind Narayanan posted thisThis week I had the honor of speaking to Princeton’s entire incoming undergraduate class to address their AI anxieties. I had three messages for them — good news, bad news, and a note of optimism. Here’s a condensed version. The good news We have enough evidence now to conclude that the shrill predictions of rapid, massive job loss were misplaced. Even in a field like software engineering where AI has been rapidly adopted, its effect has been to shift, not replace the role of the human (see the “decide-execute-deliver” framework https://lnkd.in/egNDhhMe) Similarly, the panic about what to major in is also misplaced. There will be enduring demand for computer science, philosophy, and just about everything else. (In fact, AI companies hiring philosophers has been a big recent trend.) The bad news AI seems to help senior people much more than juniors. I can use AI for coding because I spent 25 years learning how to code, which lets me supervise coding agents effectively. (See my post on the “growth cycle” vs the “dependence spiral” https://lnkd.in/ep3XcYpg) You are in a bind — you can’t offload your skill-building to AI, but you’ll graduate into a market where employers will expect you to get work done with AI. We never faced this dilemma. As a result we haven’t figured out how to revamp our classes to help you do both. You’ll have to help us figure it out. And you’ll need to somehow resist the constant temptation to turn to the shortcut machine. The hope My point is not that AI is bad for learning. It’s an incredibly flexible tool. Is the internet good or bad for learning? Depends — are you using it to find research papers or waste time scrolling? I use AI every day for learning. The key is to use it to increase, not decrease your cognitive load. To learn deeper, not faster. There is no learning without the cognitive sweat. I try to make sure I’m mentally exhausted at the end of the day. I do feel that AI lets me push myself harder than I ever could before, and I have a vision that as AI continues to advance it will enable human-AI “co-superintelligence“. (I talked about this at the end of my ICML keynote. https://lnkd.in/eC_ErqJ2) === It was great to take the stage with my colleagues Tania Lombrozo, Matt Jones, and Michael Gordin. The energy of the incoming class was wonderful and infectious, and I'm excited for what they'll accomplish.
-
Arvind Narayanan shared thisICML has released talk videos from the conference! Here's my keynote "What will be left for us to work on?" https://lnkd.in/eNwSE9Dv I made three arguments. First, the AI as Normal Technology framework is a correct and useful as a way to think about AI’s impacts, unless and until there is some future discontinuity such as through recursive self-improvement. Second, even though we should take recursive self-improvement seriously, there is no milestone that companies might achieve in the lab that will suddenly put us all out of work. Third and finally, jobs of the future will be radically different, and a lot of adaptation will be needed. I shared my thinking about what this might look like and ended with a vision of human/AI “co-superintelligence”. Relatedly, Sayash Kapoor and I have collected the essays that we think are particularly helpful for understanding the AI as Normal Technology framework, organized into Foundational essays, Applications of the framework, and Technical research and explainers https://lnkd.in/ePNfYbrx
-
Arvind Narayanan shared thisI've been going on for a while about how automation and collaboration agents need to be designed differently, but that's just the start. Here are 11 different configurations of firm/worker/agent that I think will require different agent designs. The industry's current focus is on automation / delegation, overindexing on metrics like time horizon while ignoring others like collaboration skill, and treating the human as the bottleneck. So there's a huge unmet market need. This is also a great opportunity for research, provided we treat harness design not as an afterthought but as a set of open problems that will require expertise not just in machine learning but also human-computer interaction and organizational behavior.
-
Arvind Narayanan shared thisWhat impact do data center bans or moratoria have on AI progress? Let's do some napkin math. A number such as "1 GW of capacity blocked" can be unintuitive. But we can take advantage of the fact that the efficiency of AI inference at a given capability level has been consistently improving over time and this trend will likely continue. So blocking a certain amount of capacity roughly translates to delaying AI progress by a certain amount of time — the amount of time it will take to catch up to a given capability level through improved efficiency rather than increased compute. As the table shows, if a typical U.S. state enacts a 1-year moratorium, it slows AI efficiency progress by 5-10 hours. That's assuming the blocked capacity doesn't get rebuilt elsewhere, which is extremely unrealistic. Data centers sites are highly substitutable. Assuming 90% "leakage", even a big state like NY banning all new data center construction only slows AI by less than a day. While I may have gotten some numbers slightly wrong (I used AI for the analysis and did some spot checks), the rough order-of-magnitude is enough to make my larger point: data center bans are not an effective way to slow AI. Of course, the local environmental effects are a big reason for these bans. That's a highly contested topic that I won't comment on here. But to the extent that support for these bans comes from people's generalized anxieties about AI and its societal and economic impacts, slowing new construction is an exceptionally ineffective way to channel those anxieties.
-
Arvind Narayanan posted thisI’m often critical of the Bitter Lesson in AI because it tends to get overzealously applied, so I want to take a minute to acknowledge its core truth and usefulness using a recent example. Only about a year ago, we were in awe of “Deep Research” tools. But this kind of task-specific harness quickly got swallowed by general-purpose agents. And frankly, these agents are much better “deep researchers” than last year’s tools because you can prompt them how you want (in my case, I can get them to do my searching for me without trying to do my thinking for me).
-
Arvind Narayanan posted thisA question I've gotten for decades from non-technical folks who want to work in tech (say, tech policy) is how technical they should be. I've never had a satisfying answer, falling back on "it depends" — until now. I think most people in this position should be technical enough to use agents to get work done. You can do so even without any technical knowledge, but a little bit of investment in understanding how the tech works will go a long way toward getting more out of them and, more importantly, avoiding the ever-present pitfalls (verification, overreliance, skill erosion). It is also a great way to get an intuition for where the frontier of AI capabilities lies. Acquiring this just-enough-to-be-useful level of technical knowledge has gotten easier than ever because agents themselves are good teaching tools. Of course, you still have to put in the work to learn, and maintain a critical mindset. But the "where do I even start?" problem has gone away, you don't need a human tutor, and don't have to worry about hitting some technical snag (setting up a compiler or whatever) that's going to block your progress. Usually, people asking how technical they should be want to know if they should learn to code. My answer used to be "probably", but not anymore. The alpha of learning to code has gone down a lot. It used to be a good way to understand what computers can and can't do, but AI has shifted that boundary. And non-technical folks who learn to code typically aren't planning to write production software but rather do simple things like web scraping — tasks can be delegated to agents now. Finally, there is enduring value in learning basic computer science concepts, and AI is unlikely to undermine that.
-
Arvind Narayanan shared thisA few people have recently claimed my posts are actually AI. Apparently I go to great lengths to bypass AI detectors, but these people have seen right through it, as they are connoisseurs of fine human writing. I find this hilarious and don't plan to change my writing style, which has had some AI characteristics long before AI writing existed. As I've said before, it's not AI's style itself that makes it nauseating to read, but the disconnect between the punchy turns of phrase and the shallow substance behind those words. I'm confident enough in the substance of my writing that if my style is AI-like, I take it entirely as a compliment. It's sad what the prevalence of slop has done to us, making us constantly second-guess everything we read. I find it best to stop worrying about whether something is AI and instead read it a second time to see if it reveals more depth than the first time or starts to feel hollow.
Arvind Narayanan replied to a comment
2w
Agreed — I think the rollout was rushed because the normative questions are unresolved and transparency has been inadequate. In other words, the backlash is justified, but the criticism of Anthropic has focused on the wrong thing. The distortion issue is much more legible than the real concerns I have, but ultimately technically incorrect.
Arvind Narayanan replied to a comment
2mo
I use vertical tabs in Chrome, which live in a sidebar, and I can't throw my mouse to the edge of the screen to select a tab unless the window sticks out :)
Arvind Narayanan replied to a comment
6mo
Rajeev Kozhikkattuthodi Great question. Two quick responses:
* We don't actually subsume safety into reliability — we set aside adversarial notions of safety that we see as more related to the kind of system-level, interaction-based safety failures that Leveson talks about. Our narrow focus is on what one might call operational safety, which (from what we can tell so far) appears to be caused by the same stochasticity that causes the other kinds of reliability failures. This is in contrast to Leveson's point that improving reliability won't necessarily improve safety.
* At any rate, this is a measurement project; we are not making an explicit claim about how to improve reliability/safety. So even if our hypothesis is wrong, and operational safety failures in agents have a different root cause, there is no harm in measuring them together. After all, anyone interested can disaggregate the different metrics that go into our index.
(N.B. we actually haven't aggregated safety in the current version but plan to do so in a future version.)
-
Arvind Narayanan liked thisArvind Narayanan liked thisLong time listener, first time caller #HardFork 🎧 excited to see what Casey and Kevin are up to next, but it was SO interesting to hear from Arvind Narayanan that I have to share some great #quotes I totally agree with and gave me a lot to think about 🧠 talking about #AI: “the broader point is that these tools are not only powerful, they're very, very flexible in how we can use them.” broader point on #agency: “And I think each of us should not just accept the way that the developer has created the tool, but configure them, personalize them, in ways that meet our appropriate comfort level for what should be delegated to AI and what should be in the domain of people. That's not an answer to your broader question about AI as a kind of powerful global force and how we can slow that down. But I think at an individual level, I do want to emphasize that we have a lot of agency.” https://lnkd.in/exv3x8Nq
-
Arvind Narayanan liked thisI've been flowing Arvind Narayanan since I read his paper "AI as Normal Technology" some time ago. He brings a rational and insightful view of how #AI is and will intersect with our work and life. I can confirm his commentary about how AI helps the experienced professional more than the junior: I've spent a lot of time building a local "second brain" so my AI #agents can act as assistant, reviewer, librarian, and recommender across a corpus of work information. I wouldn't have been able to do this without my existing knowledge of data retrieval, integrations with other software, the need to build and maintain indexes, how AI stores and uses external (non-training) data, etc. One-shotting a project like this simply isn't possible if you want it to work well. You still need the background experience in 2026. So how do employers bridge the gap between new employees who may only have academic experience, and expecting them to effectively use AI to accelerate their work? My thoughts stem from how I've approached new things, even now, mid-career: I tend to be deeply uncomfortable with my ability to lead a topic unless I dedicate at least some time "hands on" with the subject matter. This applies double for #technology. If it's new to me, I'll work through a minor implementation until I can appreciate what's really happening. For employers, I'd recommend an initial on-boarding experience that includes developmental or rotational assignments, and temper the expectations around AI use in favor of building those base skills in your environment. It'll be more expensive, but much more effective. I've personally benefited from programs like this in my career, and would recommend any employer that offers that opportunity.
-
Arvind Narayanan liked thisA great reflection from Princeton’s Arvind Narayanan. His point that there is no learning without cognitive load is important. This is the litmus test for learning. Let’s use technology to increase our cognitive loads and work that brain of ours into shape!
-
Arvind Narayanan liked thisvery sensible advice on AI use for students (and others) from the one and only Arvind Narayanan :)
-
Arvind Narayanan liked thisThis pretty much captures where I am at on AI and legal education. The only thing I'll add is that just as our students are experimenting with how AI can make them *better* not just *faster*, law professors and legal professionals tasked with training law students and junior lawyers must face this challenge with clear eyes. We need to be finding ways to introduce junior lawyers to the tools while being even more explicit about what core competencies remain and doubling down on prioritizing and incentivizing that they learn those skills. Plus we must experiment too. Its a scary time but maybe an exciting one as well? I certainly think so.
-
Arvind Narayanan liked this"Similarly, the panic about what to major in is also misplaced. There will be enduring demand for computer science, philosophy, and just about everything else. (In fact, AI companies hiring philosophers has been a big recent trend.)" #philosophy #liberalarts #highereducation
-
Arvind Narayanan liked thisExcellent advice! The Good News, The Bad News ,The Hope about AI and its impact on jobs!
Explore more posts
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content