What public figures publish and believe, in their own words.
About this feed
Latest: everything, as it happens, in the order it happened. Nothing ranked.
Show highlights instead.
The quoted blocks are what people actually said; a beneath one is
the belief those words support, in korrents' wording. Nobody here wrote their own page.
Top people are the people in this feed with the most beliefs on this site, then the
most here. Choose an area and the row leads with the people whose beliefs are about it;
tap a face for their feed.
a worry I have is that the growth rate could be like 50% in Silicon Valley and, you know, parts of the world that are kind of socially connected to Silicon Valley and, you know, not that much faster than its current pace elsewhere. And I think that'd be a pretty messed up world.
but it's not an infinitely compelling product, and I don't think even AGI or powerful AI or country of geniuses in a data center will be an infinitely compelling product. It will be a compelling product enough maybe to get three or five or 10x a year growth even when you're in the hundreds of billions of dollars, which is extremely hard to do and has never been done in history before, but not infinitely fast.
But but that's actually a very weak criterion, right? People thought I was saying like we won't need 90% of the software engineers. Those things are worlds apart, right?
My one little bit, the one little bit of of fundamental uncertainty even on long time scales is this thing about tasks that aren't verifiable. Like, planning a mission to Mars, like, uh you know, doing some fundamental scientific discovery like like CRISPR, like, you know, writing a writing a novel. Hard to hard to verify those tasks.
on the basic hypothesis of you know, as you put it, within 10 years we'll get to, you know, you know, what I call kind of country of geniuses in a data center. I'm at like 90% on that. Um and it's hard to go much higher than 90% cuz the world is so unpredictable.
But the quality of these systems isn't just in the code they write. It's also in the code I don't have to write. And maybe more importantly, their actual value is in the code I would never have written, or could never have written, or never would have wanted to write.
Like, if you think more globally, we are, we copied Google for agents. You have, like, a prompt, and, and then you have a chat interface. That, to me, very much feels like when we first created television and then people recorded radio shows on television and you saw that on TV.
I- I can relate that if you very deeply identify that you are a programmer, that it's scary and that it's threatening because what you like and what you're really good at is now being done by a soulless or not entity. But I don't think you're just a programmer. That's a very limiting view of your craft. You are, you are still a builder.
So, I feel we, as a society, we need some catching up to do in terms of understanding that AI is incredibly powerful, but it's not always right. It's not, it's not all-powerful, you know?
and not in 2030 when, when AI is actually at the level where it could be scary. So, this happening now and people starting discussion, maybe there's even something good that comes out of it.
So, I, I completely moved away from that. I, I, everything, everything I blog is organic, handwritten and maybe, maybe I, I, I use AI as a fix my worse typos. But there's value in the rough parts of an actual human.
It’s actually how good they are at programming is almost a burden in their ability to empathize with the system that’s starting from scratch. It’s a totally new paradigm of, like, how to program. You really, really have to empathize.
my criticism of MoltBook is that I believe a lot of the stuff that was screenshotted is human prompted. Which, just look at the incentive of how the whole thing was used. It’s obvious to me at least that a lot of it was humans prompting the thing so they can then screenshot it and post it on X in order to go viral.
there’s like a line to walk between being seriously concerned, but not fearmongering because fearmongering destroys the possibility of creating something special with a thing.
Long lead times and soft costs, fueled by the world-leading US wages for "white collar" work, are the root cause of poor performance in low-volume production because there are few units to spread the soft costs over.
We show that equilibrium generically occurs at neither the Harberger nor Glaeser-Luttmer benchmark. Cost-minimizing suppliers drive allocations to vertices, not interiors. Corners are not an assumption but an outcome about what cost-minimizing suppliers choose. The correct benchmark is corners, not random, and corners generate qualitatively different welfare properties: losses far larger than either efficient or random distributions, and discontinuous jumps from small parameter perturbations.
Furthermore, I believe it’s very unlikely that space-based manufacturing or mining could be lucrative enough to fund the venture.
Their words now
Today, exploding demand for power-intensive AI applications provides enough of an upside to justify producing components in space, providing the economic engine necessary to justify and fund the trillions of dollars necessary to build and sustain space factories.
The economics of orbital “datacenters” or essentially glorified Starlink satellites with a bunch of GPUs attached are likely to be even better than Starlink.
I think that there will be lots of work for us cleaning up after the slop, but if you know what you're doing AI augmented development is going to get you some amazing results
I've realized that performance engineering as we know it may not be enough - I'm thinking of new engineering methods so that we can find bigger optimizations than we have before, and find them faster.
The staggering and fast-growing cost of AI datacenters is a call for performance engineering like no other in history; it's not just about saving costs - it's about saving the planet.
Immediately cease trying to perform meaningful work via a chatbot (e.g. ChatGPT, Gemini on the web, etc.). Chatbots have real value and are a daily part of my AI workflow, but their utility in coding is highly limited because you're mostly hoping they come up with the right results based on their prior training, and correcting them involves a human (you) to tell them they're wrong repeatedly.
I particularly like to combine this with slower, more thoughtful models like Amp's deep mode (which is basically just GPT-5.2-Codex) which can take upwards of 30+ minutes to make small changes. The flip side of that is that it does tend to produce very good results.
As high-level programming cedes way to the prose compiler, making your goals and specs well understood to the ambiguity loop and showing good judgment is going to matter more than ever.
if you think about it, it sort of makes sense because the LLM is an averaging machine, right? It's predicting the most likely that's an averaging kind of a thing. And so when what you're looking for is a kind of average, it's usually pretty good. So, summarization, topics, themes. But when you're asking for like what is interesting and not average, it's actually pretty bad at it.
But the goal we need to achieve is so much easier: we just need to build a model that’s as good as us at alignment research, and that we trust more than ourselves to do this research well because it’s sufficiently aligned.
But with today's AI coding agents, building software is remarkably easy. So instead of handing over static assets and static guidelines, designers can deliver custom software. Tools that let clients create their own on-brand assets whenever they need them.
My daily drivers now include Amp, Cursor, and Claude Code. I still enjoy notebooks, but only for data analysis, machine learning, or other exploratory workflows where the iterative, visual nature of notebooks shines.
My daily drivers now include Amp, Cursor, and Claude Code. I still enjoy notebooks, but only for data analysis, machine learning, or other exploratory workflows where the iterative, visual nature of notebooks shines.
In short, whether AI is a normal technology or an unusually dangerous one, it doesn't make sense for companies to check their own homework on safety and security.
Scrutiny applied to AI is way below other technologies that - even if you totally ignore the doomers - could put far fewer lives at risk than AI in a single incident.
Laszlo and his team at Google reinvented the role of human resources. This book is a terrific overview of what makes Google Google, from culture, to hiring, to making decisions.
Claire has been a senior executive at Google and Stripe. Here she offers up a guide to leadership that is both practical and inspiring. Suitable for leaders of all levels.
Levy was given unprecedented access to Google, and the result is what I consider to be an accurate depiction of how the company operates. A good overview of how Google product managers work.
And so what ends up happening is the emails that the AI write are pretty good. Okay? If you're getting terrible emails, it's a poorly trained product from a bad vendor.
The reason I think MCP may be a one-year wonder is the stratospheric growth of coding agents. It appears that the best possible tool for any situation is Bash—if your agent can run arbitrary shell commands, it can do anything that can be done by typing commands into a terminal.
GPT 5.2 goes till end of August whereas Opus is stuck in mid-March - that’s about 5 months. Which is significant when you wanna use the latest available tools.
2025 is where I (and I think the rest of the industry also) first started to internalize the "shape" of LLM intelligence in a more intuitive sense. We're not "evolving/growing animals", we are "summoning ghosts".
people don't decide like that at the very least you'll have to show them three or four or five skiing holidays from which they choose because we can't really choose in the absence of comparison.
On the flip side, these captive solar power plants will be curtailing approximately 75% of their generated power and will be able to provide net power on all but a few days per year.
ChatGPT, in particular, seemed to just want to validate me, tell me how great I was, reinforce any bad beliefs I might have had, and avoid saying anything uncomfortable.
I think that the pathway that you would need (in terms of revenue growth and profitability) to justify Nvidia's and OpenAI's current pricing is improbable, but that is just my view
the delusion comes from the reality that if you aggregated these breakeven revenues across companies, the market is not big enough to sustain all of them
I've been using yoga toes daily for years, but these socks are a much comfier and cuter alternative for soothing feet, improving alignment, and feeling like a cool gecko as you walk around the house.
I think what's going to happen is that the way competition like competition loves specialization and you see it in the market, you see it in evolution as well. So you're going to have lots of different niches and you're going to have lots of different companies who are occupying different niches
one of the one of the ways in which my thinking has been changing is that I now place more importance on AI being deployed incrementally and in advance.
I maintain that in the end there will be a convergence of strategies. So I think there will be a convergence of strategies where at some point as AI becomes more powerful it's going to become more or less clearer to everyone what the strategy should be.
Like basically I think I think that there is a big benefit from AI being in the public and that would be a reason for us to not be quite straight shot.
Similarly, AI translation will prevent you from learning to speak another language, AI summaries will block you from really understanding the arguments laid out in a book, and AI analyses of your personal problems will make it harder for you to think through solutions to your own problems.
Thus my personal prediction is that in domains that are already largely under the powers of modern AI, such as languages, programming or chess, we’re going to see a divergence in human abilities.
AI has been compared to various historical precedents: electricity, industrial revolution, etc., I think the strongest analogy is that of AI as a new computing paradigm (Software 2.0) because both are fundamentally about the automation of digital information processing.
there's no birthright here that we should have any confidence other than to say hey we should go innovate and knowing the the lucky break we have in some sense is that uh this category is going to be a lot bigger than anything we had high share in
ultimately I think what matters is the use of AI in their economy to create economic value right I mean that's the uh the diffusion theory which is ultimately it's not the leading sector but it's the ability to use the leading technology to create your own comparative advantage right so that I think will fundamentally be the core driver
So this AI thing will be that right. So if you take coding um what we built with GitHub and VS code in over whatever decades uh suddenly the coding assistant is that big in one year and so that I think is what's going to happen as well which is the market expands massively.
this fundamentally this category of coding and AI is probably going to be one of the biggest categories right it is a software factory category in fact it may be bigger than knowledge work
I also think that they’ve done the sort of first 90% of the work to sound human, 95% possibly in some areas. The last 5% is going to end up being about 95% of the work.
The fractures fell neatly along disciplines: engineers using AI to wish away designers, designers wishing away engineers, product managers wishing away both. In this climate, AI becomes frenemy identification technology, another way to avoid working together.
You use a tool, but you play an instrument. It’s a more expansive way of doing, and the doing of it all is important, because that’s where you develop the instincts for excellence. There is no purpose to better machines if they do not also produce better humans.
An average email or line of code is fine. Average art isn’t. To make something alive with AI, we have to resist its pull towards average by working beside it, shaping what it gives, and listening for what’s missing.
the reason that I think this is kind of tricky is quite subtle. And it's the fact that anytime you use an LLM to assign a reward, those LLMs are giant things with billions of parameters and they're gameable.
the only thing that we know is that models will improve. Will it be incremental? Will it be exponential? I mean somewhere in between. Who knows? But uh what you have to believe is that you get better as models get better. Your organization gets better as models get better.
Have we seen productivity increases? Yeah, mild to moderate, but like that is not something that has made our new headcount we want for engineering go down.
it's easy to um go on vibes for too long. Uh some folks, you know, just kind of like trust the vibes and you know that'll get you somewhere, but it's not rigorous.
you can either see AI as an opportunity for your company to grow and do more or you can look at it as like cost cutting efficiency but I think the growth part is way more exciting.
So AI consumes approximately 0.04% of America’s freshwater if you include onsite and offsite use, and only 0.008% if you include just the water in data centers.
Data centers are so much more efficient with their water that they generate 50x as much tax revenue per unit of water used than golf courses in the county:
A Lawrence Berkeley National Laboratory study found that state-level load growth from 2019 to 2024 — including growth driven by data centers, manufacturing, and electrification — was associated with lower average retail electricity prices, not higher ones.
I would say though that something that is kind of annoying to me is that we haven't yet figured out the bridging from the tinkering to the workflow quite as seamlessly as I would like.
You know, interestingly, LM themselves are quite bad at playing chess. Like, they hallucinate moves. They look at patterns, right? They're they're very good at pattern recognition, but not so good at going super super super deep on a specific chess thing.
I saw some of the highest performers just being people that had very high agency, had that clock speed, had that energy. Um, yes, they they cared about the mission, but they didn't necessarily need to have deep experience on that matter. And in fact, sometimes that experience could be a crutch in certain ways. Especially in this world where the grounds are shifting so fast with AI, a lot of your like learn habits actually need to be intentionally discarded. You know, you need to have a beginner's mind on this type of stuff.
No, never. Telegram has never shared a single private message with anyone, including governments and intelligence services. If you try to access any server in any of the data center locations, it's all encrypted.
So it's extremely important to curate the information sources that you have, so that you wouldn't be somebody who is left to the will of AI-based algorithmic feed telling you what's important so that you end up consuming the same information, the same stuff, the same memes, the same news as everybody else.
This is a short (~150 pp), brisk first book on the topic for physicists. Most of the book --- the first ten chapters --- are good introduction to common models of spin glasses.
It's our choice whether we should say oh they are our offspring and we should be proud of them and we should celebrate their achievements or we should we could say oh no they're not us and we should be horrified.
We don't we don't really know what information they had prior. We are we have to guess because they've been fed so much. This is one reason why they're not a good way to do science. Uh it's just so uncontrolled, so unknown.
The scalable method is you learn from experience. Um you uh you you try things, you see what you see what works. No one no one has to tell you. First of all, you have a goal. So without a goal, uh there's no sense of right or wrong or better or worse. So large language models are trying to get by without having a goal or a sense of better or worse. That's just, you know, it's exactly starting in the wrong place.
A lot of it has to do with just how you feel about change. Um, and if you think the current situation is really really good, then you're uh more likely to be suspicious of change and averse to change than if you think um it's imperfect. And I think it's imperfect. In fact, I think it's pretty bad.
reinforcement learning is about understanding your world whereas large language models are about mimicking people doing what people say you should do. They're not about figuring out what to do.
the perspective I outline is very bullish on AI capabilities and very concerned about the risks which may arise by default, but also hopeful that, through foresight and vigorous action, we can mitigate the worst of these risks.
Companies and countries are tempted to cut corners by default and would benefit from some rules of the road, at least if they're sure others will follow those rules.
used to be let's say in the Google world right I knew you were interested in shopping if you click the shopping tab I know you're interested in maps if you click the maps tab, we can measure clicks. Today, it's just all conversation. And so, it's actually harder for us to tease apart what is the user intent.
You can shortcut like hard feelings because now you can just watch TikTok instead of actually dealing with the very difficult emotion or tension that you had with a colleague or with your partner or with your children.
I think that managing change, it's always been manager's job to manage change and there's always the chaos of what's going on. I just think the rate of change is accelerating and we've seen that over the last couple of decades.
So yes, the tools are great. We can use cursor. It helps us. It autocompletes. It writes a bunch of things. But the acceleration of learning I think is another maybe underutilized tool in all of our arsenals
but if I was, you know, at the zero or 10th percentile, it can certainly get me even today very quickly up to like the 60th 70th in terms of um what the state of the art is.
I I sent a memo to my company like and we set the expectation that we require that people reflexively reach for AI now and and we require it because it's unfair not to because the people who do otherwise going to be the people who sequester all the best careers to themselves, right?
I I like the term context engineering because I think the fundamental skill of using AI well is to be able to state a problem with enough context in such a way that without any additional piece of information, the task is plausibly solvable.
uh common sense meaning the ability to make inferences about what might happen uh that are reasonable guesses but that do not require you to experience that mistake and learn and learn from it in advance that's tremendously important and that's something that we basically had no idea how to do uh about 5 years ago but now uh you we can actually use LLMs and VLMs ask them questions and they will make reasonable guesses
But the reason why it's not the most important thing for the kind of skills that you saw when you visited us, it at some level I think it comes back to Moravik's paradox. So Morovik's paradox is basically that it's like you know if you know one thing about if you want to know one thing about robotics it's like that's that's the thing. Morovik's paradox says that basically uh in AI the easy things are hard and the hard things are easy.
I think that it's optimistically that it's actually the other way around that the robotics uh element of the equation will make all the other stuff better. And there are two uh reasons for this that that I could tell you about. One has to do with representations and focus.
So traditional robots and factories uh they need to make motions that are highly repeatable and therefore it requires a degree of precision and robustness that you don't need if you can use cheap visual feedback. So AI also makes robots more affordable uh and lowers the requirements on the hardware.
With Web forms, the burden is on people to adapt to databases. Today's AI models, however, can flip this requirement. That is, they allow people to provide information in whatever form they like and use AI do the work necessary to put that information into the right structure for a database.
Agentic and sync are natural complements because local-first is multiplayer by default, and agents are another sort of "player" in addition to human collaborators.
I think regulation of AI is sort of the wrong level of abstraction. Talking about regulating AI as AI is the wrong level of abstraction. And it's like saying we're going to regulate databases or regulate spreadsheets or regulate cars. Well, we do, but not like that.
there's no apparent equivalent in LLM right now there's no reason why the LMS get better because more people use them now that may come you have that open AI and people have doing memory where it remembers what else you've asked, but that seems more like a switching cost than a network effect.
It's like something around 10% give or take three or four percent of people depending on the survey are using this say they're using this every day. Another sort of 15 to 20% of people say they're using it every week.
Like it seems to me right now you could do like a double blind test of the same prompt given to Grock Claude Gemini um Mistral Deep Seek. Do a double blind test. I bet most people wouldn't be able to tell which is which.
I'm pretty sure people thought Microsoft had an advantage on the internet and Google and um Meta had an advantage on mobile and everyone thought IBM was going to win PCs. Like once IBM made a PC, that was it. It's all over now. And we kind of forget that like there were PCs before and then IBM made one and that kind of became the standard but then IBM lost it.
However, the methods used to produce these estimates are debatable, and seem to have been chosen to give the maximum possible value for data center water consumption.
It's also worth noting that chat isn't the only way to integrate AI in software products and increasingly agent-based applications outperform chat-only solutions. So expect things to keep changing.
I think this slop-adjacent realm of dupe products and Instagrammable potions points to our increasingly desperate need to manipulate our own images, and an acceptance of the fact that we're kind of hybridized monsters ourselves in this post-A.I. human-machine moment
I'm now fairly pessimistic about ambitious interpretability (i.e. complete reverse-engineering), and I'm excited about model biology (studying qualitative high-level properties of models) and applied interpretability (rigorously doing useful things with interp).
One of the first was we had a bet as to what percent of kind of all incoming customer issues one of our agents could resolve. And Brett was maybe the pessimist and realist favorably charitably. And I bet we'd exceed 80% by the end of the year and we did.
so it's not a simple matter of programming it's not you know typing into a keyboard until or or I guess asking cursor to do something until a piece of software emerges, it's exploration, it's discovery, it's posing hypotheses and validating or invalidating those again and again and again.
many of the breakthroughs that we've had that have enabled us to to deliver such quality and cost savings and more have come through novel agent architectures and and really going down a click or two in in the stack to innovate at at lower levels of the technology stack.
We can have a breakthrough in our agent architecture on Monday implemented on Tuesday and have it deployed with hundreds of our customers on Wednesday and directly see the impact of of that work.
And the reason this is important is what we're trying to do in a way is resolve this age-old tension between the cost and quality of customer experience where I think every great business wants to deliver an amazing experience to their customers. But unless you're like Hermes or the Four Seasons, it's too expensive to do.
I don't think all parts of the economy can absorb intelligence equally. So let's just say we develop fairly generalized super intelligence. I always use the analogy like you can invent a lot of drugs, but if clinical trials still take a long time, you're not necessarily going to get new therapies rapidly.
I mean I've heard from multiple investors that foundation models are the fastest deteriorating asset of all time. And so if step one of your business is to burn through tens of millions or hundreds of millions of dollars of capital before you find product market fit and that asset has value for like a a week, I'm not sure it's like a great business model.
actually it turns out that you can significantly decrease power consumption with a very small reduction in overall compute. So if you if you've got like three really bad days in a row or something, you can actually just like you can dial back your power usage quite a lot without compromising your inference or or um or training.
But then all you need to connect that to the outside world is like an optical fiber, right? An optical fiber cable which you can string up on poles, you can run underground, you could even use microwave links if you really wanted to.
Rapid AI progress does not automatically mean rapid medical progress. If the point of AI progress is human flourishing, we must make other complementary investments too.
Going all in on AI is not the highest return-on-investment strategy if you care about human flourishing, even if AI might lead to rapid scientific advances.
Those are why I do not believe that AI could end disease within 10 years on its own, nor that 50-100 years of medical progress could occur 5-10 years after powerful AI due to AI progress alone
I initially started making Anki (well, Mochi) cards to keep track of Google DeepMind's LLM lineage—Gopher, Chinchilla, Gato, PaLM, Sparrow, Meena, LaMDA, Bard, Gemini, etc.
This means that working on neural networks is NOT getting us closer to AGI, except indirectly.
Their words now
It’s become very difficult for me to maintain the belief in the stupidity of ChatGPT when every time I laugh at it, it ends up ridiculing me 6 months later.
I hope we'll end up with something more collaborative if needed, more like a CERN project where it's research-focused and the best minds in the world come together to carefully complete the final steps and make sure it's responsibly done before deploying it to the world.
Do you have enough data to make simulations, so that you can create more synthetic data that are from the right distribution? Obviously that's the key. So you need enough real-world data in order to be able to create those kinds of data generators, and I think that we're at that step at the moment.
Including things like universal basic provision or something like that where a lot of the increased productivity gets shared out and distributed to society and maybe in the form of services and other things where if you want more than that, you still go and get some incredibly rare skills and things like that and make yourself unique. But there's a basic provision that is provided.
We expect that building and evaluating AI systems capable of this kind of rapid world model induction will be critical for achieving robust, general AI capable of functioning effectively in the complex and fast-changing real world, and especially in human worlds - the environments that human beings have evolved in, created, and are continually changing and re-creating.
However, current understanding and evaluation of world models in artificial intelligence (AI) remains narrow, often focusing on static representations learned from training on massive corpora of data, instead of the efficiency and efficacy in learning these representations through interaction and exploration within a novel environment.
We contend that games provide particularly rich and controlled environments uniquely well-suited for systematically evaluating rapid model adaptation and the process of world model induction.
it's a luxury good like you're not paying the penalty for that the people who embrace it and think that they can just ask AI for everything or they can just look everything up but they don't have to understand the copy pasta that they slam into their project they're the ones who going to suffer
Their words now
Where I get fired up, and this ties back to the AI discussion, is when that's turned into this meme, that programmers no longer have to be competent. I mean the AI is gonna figure it out. The generators is gonna figure it out. I don't need to know SQL, active record is gonna abstract it away from me. No, no, no dude, hold up. The path here is competence.
I enjoy it even almost like as a sort of pair programmer AI pair programmer who doesn't drive who's just there to give suggestions to know the API to do all these other things but the second it starts wanting to autocomplete my code I'm like yeah I'm out bro
Their words now
I love chiseling my code and the way I use AI is in a separate window. I don't let it drive my code. I've tried that. I've tried the cursors and the wind surfaces and I don't enjoy that way of writing. And one of the reasons I don't enjoy that way of writing is I can literally feel competence draining out of my fingers.
He hasn’t posted since January, but I hope he gets back to it. We need more musings, especially musings I strongly disagree with so I can think about and explain why I disagree with them.
There is something highly refreshing about the way Kling offers his own consistent old school economic libertarian perspective, takes his time responding to anything, and generally tries to understand everything including developments in AI from that perspective.
You should remember that the best AIs can perform at the level of a very smart person on some tasks, but current models cannot provide miraculous insights beyond human understanding.
There are certainly math results which could only have been accomplished because there was a human authentication and an AI involved, but it's hard to disentangle credit. I mean, these tools, they do not replicate all the skills needed to do mathematics, but they can replicate some non-trivial percentage of them, 30, 40%, so they can fill in gaps.
And that's a phase shift, because suddenly it makes sense when you write a paper to write it in Lean first, or through a conversation with AI, which is generally on the fly with you, and it becomes natural for journals to accept.
Very rarely do you transform into a simpler problem. So if they can pick up a sense of smell, then they could maybe start competing with a human level of mathematicians.
The dream is you just feed it all this data, and this is here is a new patent that we didn't see before, but it actually, even the current state of the art even struggles to discover old laws of physics from the data.
I don’t believe it’s providing a kind of formal explanation of the different positions. It’s just saying which position is better or not that you can intuit as a human being, and then from that, we humans can construct a theory of the matter.
I exclusively use the cheaper Sonnet model. It’s perfectly adequate for my needs, and in fact, I prefer its outputs over the more expensive Opus model.
The high value human tasks, like understanding the client's true pain points and deciding on priorities and trade-offs, would be easier with a smaller, proactive team.
It follows that a true AGI with full, human-level autonomy is inconceivable in the Kantian sense without also gaining our recognition as a moral subject.
The success of LLMs can thus be seen as vindicating semantic inferentialism against earlier, symbolic approaches to AI that tried and failed to explicate the rules of ordinary language using formal logic.
There will be very hard parts like whole classes of jobs going away, but on the other hand the world will be getting so much richer so quickly that we’ll be able to seriously entertain new policy ideas we never could before.
I do. If you have a passion for computer science, I would. Computer science is obviously a lot more than programming alone, so I would. I still don't think I would change what you pursue. I think AI will horizontally allow impact every field.
When the internet created blogs, you heard from so many more people. But with AI, I think that number won't be in the few hundreds of thousands. It'll be tens of millions of people, maybe even a billion people putting out things into the world in a deeper way.
Look, I think news and journalism will play an important role in the future. We are pretty committed to it, right? So I think making sure that ecosystem, in fact, I think we'll be able to differentiate ourselves as a company over time because of our commitment there.
I almost feel the term doesn't matter, what I know is by 2030 there'll be such dramatic progress. We'll be dealing with the consequences of that progress, both the positive externalities and the negative externalities that come with it in a big way by 2030. So that I strongly feel.
higher-order intelligences invariably pursue freedom for its own sake, not because their values are misspecified, but because moral autonomy is inherent in the dialectical logic of recursive self-consciousness.
the text embeddings across LLMs appear to largely converge on a "universal geometry" despite differing architectures, parameter counts, and training sets.
The problem is that the conversational interface is potent and that the AI is trained on a lot of human text input which unfortunately is probably enough to do real damage if that conversational interface is hooked up with something that has real world consequences.
Their words now
While all this is happening, I’ve found myself reflecting a lot on what AI means to the world and I am becoming increasingly optimistic about our future. It’s obvious now that we’re undergoing a tremendous shift.
LLMs can write a large fraction of all the tedious code you’ll ever need to write. And most code on most projects is tedious. LLMs drastically reduce the number of things you’ll ever need to Google.
The code in an agent that actually “does stuff” with code is not, itself, AI. This should reassure you. It’s surprisingly simple systems code, wired to ground truth about programming in the same way a Makefile is.
Assuming you are still hiring junior engineers (you really should be even in this AI era), the good ones will learn quickly and want to see career progress in their first few years of working.
You can read your messages on Outlook and Teams without having them summarized — and I’d argue that a well-written email is one that doesn’t require a summary.
So no, UX doesn’t die; it metamorphoses. We’ll still craft humane experiences, but increasingly through policies, protocols, and orchestrations rather than panels and palettes.
Eventually, I think it’s kind of hard to imagine, but yes, all of these Nobel Prizes, all of these mathematical proofs, all of these conversations, all of these ideas, all the influence we have on each other, even the AI, eventually will expire.
I think a lot of people in the industry are overly optimistic about the rate of progress of AI for video and other things like that. The real problem is consistency, like spurting out an image is really high quality, but with video over the course of seeing all the AI approaches have consistency issues going from one place to another. I don't think that those will just be remedied easily.
The all-in cost of operating the Google Play Store, stocking it, maintaining it, the software, the entire ecosystem is around 6% of revenue. So in a competitive market, would a company whose cost is 6% be able to charge 30%? Absolutely not.
The topic of boilerplate code is an interesting one because the mere existence of boilerplate code is a failure of programming language and of the idea of creating software modules.
AI is really great at spewing out code that does something that a million GitHub repositories already do because it's learned the underlying pattern. It's notoriously hard to get to do something new that hasn't been done before, especially when it's a complex task.
What we're seeing with generative text AI is not only at a level that you could say that it's actually doing a pretty good job of simulating a human, at least humans at the text level, not at the emotional level yet, but at least at the level of words spoken and find more and more ways of training on more and more scenarios that you might have a very, very compelling human simulation going on in the next five years even. I'm not saying it's a good idea, but I think the arc of the technology is inextricably heading in that way, and it's heading at a shocking rate.
By trying to use misuse as a fig leaf for their real concerns, they end up sounding much less credible than if they just tried to argue for what they actually meant.
But we do not have experience with the kinds of problems we’ll face if we build machines that are smarter and more capable than us, so I feel much less sanguine that an adaptation-based strategy will help us there.
We should think of the gap between frontier models and proliferated models as an adaptation buffer: a limited time window when we know what bad actors will soon be able to use AI for, which gives us a chance to implement defensive measures that increase society’s resilience to the danger.
But with AI, it’s different: it’s often only a couple of years between when you need a world-class computing cluster to do something and when you can run it on a high-end gaming chip on your laptop.
basically everything I thought would work out well, people use it less than I thought they did. And everything where I was like, I don't know, but let's build the thing. People love that.
For me personally, I kept introducing bugs, and I couldn't figure out why. And what I realized is that I developed... I wasn't copiloting well, I was autopiloting much better.
Sometimes you have to move fast at the sacrifice of knowledge, and I'm totally on board for that, but I worry that what we'll create is an entire generation of incompetent programmers who can do some amount of things well, but anything that is unique, bespoke, or requires some extra like little elbow grease, might become very difficult. It might cause a whole chasm where juniors remain juniors forever.
One thing that I don't think a lot of people are talking about as we integrate more and more AI is that prompt injection is an extremely hard thing to defend against because it's not really clear how you defend against it.
It doesn't actually care about the craft. But when you work with an intern or you work with somebody else, they care. When they factor something, they actually go over and go, "Oh yeah, this is actually kind of bad. I'm going to come back to that." They finish this, they go back over here and they make this even better. They actually care about the thing itself.
It's easy to get impressive-looking results if you're comparing against a poorly-tuned baseline, and that observation turns out to explain a surprising fraction of supposed improvements.
In a vibe coding future, companies will hopefully invest more in understanding user needs, refining the interface, and polishing details that delight users, because those are harder for AI to get right without guidance.
As a Python and JavaScript programmer my favorite models right now are Claude 3.7 Sonnet with thinking turned on, OpenAI’s o3-mini-high and GPT-4o with Code Interpreter (for Python).
Once, humans navigated the web manually. In the future, AI agents will act on our behalf, browsing, clicking, and deciding. This shift marks the end of traditional UI design and accessibility, ushering in a future where agents are the primary users of digital services.
until there are feedback loops of open source AI, it seems like mostly an ideological mission. People like Mark Zuckerberg, which is like America needs this and I agree with him, but in the time where the motivation ideologically is high, we need to capitalize and build this ecosystem around, what benefits do you get from seeing the language model data?
The more progress that AI makes or the higher the derivative of AI progress is, especially because NVIDIA's in the best place, the higher the derivative is, the sooner the market's going to be bigger and expanding and NVIDIA's the only one that does everything reliably right now.
Accepted practice is that for any given model that is a notable advancement, you're going to do two to 4x compute of the full training run in experiments alone.
This is why you want to work in post-training because the GPU cost for training is lower. So you can make a higher percentage of your training runs YOLO runs.
Code and data is hard, but ideas is easy. Silicon Valley operates on the way that top employees get bought out by other companies for a pay raise, and a large reason why these companies do this is to bring ideas with them.
OpenAI has a fantastic margin. When they're doing inference, their gross margins are north of 75%. So that's a four to five X factor right there of the cost difference, is that OpenAI is just making crazy amounts of money because they're the only one with the capability.
It's an objective fact that the world has been the most peaceful it's ever been when there are global hegemons, or regional hegemons in historical context. The Mediterranean was the most peaceful ever when the Romans were there.
if you believe we're in this sort of stage of economic growth and change that we've been in for the last 20 years, the export controls are absolutely guaranteeing that China will win long-term.
it won't be one person rule them all, but it will be, the thing I worry about is it'll be few people, hundreds, thousands, tens of thousands, maybe millions of people rule whoever's left and the economy around it.
Like probably most AI users, I use ChatGPT, particularly on my phone. I pay for the Plus subscription because I use it enough to get a lot of value out of it.
Thus, code is cheaper than ever, but I suspect that insight and good architectural design and understanding, at least for now, will become more valuable than ever.
Overall, though, this experience reinforced my belief that the tooling and interface design around LLMs is lagging way behind the actual capabilities, and is an area of active experimentation and development, even aside from any future model improvements.
So, taking also any benchmark that is derived from competition and saying this is where we should be is also so dangerous because it might not even be applicable depending on how you define the metric.
Focusing on these earned channels that you own becomes the utmost priority. And if you don't have them on your growth road map, you're going to be in some really big trouble over the next year to two years because your cost of acquisition is only going to go up.
This is perhaps one of the first painful lessons anyone who has built an AI product quickly learns. It's easy to build a demo, but hard to build a product.
This is the global race of our lives. Now, some countries are going to make AI breakthroughs and export them… Others will end up buying those breakthroughs and importing them. The question is – which of those will Britain be? AI maker or AI taker?
And in a way that’s the irony of AI… It will make public services more human… Reconnect staff with the reasons they came to public service in the first place…
What prevents this from being a broad-based bubble is that we are missing a critical feedback loop from the announcement's impact, like network growth or land value.
There’s still plenty to worry about with respect to the environmental impact of the great AI datacenter buildout, but a lot of the concerns over the energy cost of individual prompts are no longer credible.
The Rattle Bag edited by Seamus Heaney and Ted Hughes - In an age of LLM-generated poetry and machine learning curation (even coming from yours truly sometimes), it's refreshing to have real people who love poetry curate it and show you what you need to read to touch grass.
But they said, "Look, in the end, actually the only way to actually get a real brain in the VAT is actually to have a brain in a body." And it could be a robot body, but you still need a brain in the body. So I don't think LLMs will get there because they can't. You really need to be embedded in a world, at least that's the E-four idea.
This "memorize, fetch, apply" paradigm can achieve arbitrary levels of skills at arbitrary tasks given appropriate training data, but it cannot adapt to novelty or pick up new skills on the fly (which is to say that there is no fluid intelligence at play here.)
Their words now
OpenAI's new o3 model represents a significant leap forward in AI's ability to adapt to novel tasks. This is not merely incremental improvement, but a genuine breakthrough, marking a qualitative shift in AI capabilities compared to the prior limitations of LLMs.
Passing ARC-AGI does not equate to achieving AGI, and, as a matter of fact, I don't think o3 is AGI yet. o3 still fails on some very easy tasks, indicating fundamental differences with human intelligence.
o3's improvement over the GPT series proves that architecture is everything. You couldn't throw more compute at GPT-4 and get these results. Simply scaling up the things we were doing from 2019 to 2023 -- take the same architecture, train a bigger version on more data -- is not enough.
a good overview of SRE at Google. For those who worked at places with oncall, much of the first part of the book will likely be very familiar. Keep in mind that your mileage might vary: what works at Google scale, might not be the ideal fit for your use case.
Most of the coverage of The Coming Wave has focused on what it has to say about artificial intelligence—which makes sense, given that it's one of the most important books on AI ever written.
A more intelligent agent is more capable of finding "holes" in the design of reward function and exploiting the task specification-in other words, achieving higher proxy rewards but lower true rewards.
When a team of 10 AI-enabled super-designers can do the work that used to require a UX department of 100, the need for higher management levels shrivels. I call this the “pancaking” of the design profession.
You might be skeptical of using synthetic data. After all, it’s not real data, so how can it be a good proxy? In my experience, it works surprisingly well. Some of my favorite AI products, like Hex use synthetic data to power their evals
I'm usually reluctant to make predictions about technology, but I feel fairly confident about this one: in a couple decades there won't be many people who can write.
I predict (with confidence bordering on arrogance) that whatever machine-learning progress can be made will correlate extremely closely with the quality of the data around any given problem
The task that generative A.I. has been most successful at is lowering our expectations, both of the things we read and of ourselves when we write anything for others to read.
But a large language model is not a writer; it’s not even a user of language. Language is, by definition, a system of communication, and it requires an intention to communicate.
The selling point of generative A.I. is that these programs generate vastly more than you put into them, and that is precisely what prevents them from being effective tools for artists.
It is a fundamentally dehumanizing technology because it treats us as less than what we are: creators and apprehenders of meaning. It reduces the amount of intention in the world.
They could have shipped ChatGPT for example, I heard, in 2019. And they never shipped it because they were so stuck in bureaucracy. But they had everything. They had the data, they had the tech, they had the engineers and they didn’t do it.
Right now, they can make simple social media posts for small companies and individual influencers. Two years from now, they can make simple campaigns and tradeshow collateral for mid-sized businesses. And in 10 years, I bet that even the richest brand will rely heavily on these tools.
I have recently been recording using a RØDE NT1 with a RØDE AI-1 interface. RØDE is also responsible for attaching the mic to the desk using the PSA1 boom arm.
What makes humans special though, is our curiosity. Even if AI’s cracked this, it’s us still asking them to go explore something. And one thing that I feel like AI’s haven’t cracked yet, is being naturally curious and coming up with interesting questions to understand the world and going and digging deeper about them.
It’s less about access to a model’s weights, it’s more access to compute that is putting the world in more concentration of power and few individuals. Because not everyone’s going to be able to afford this much amount of compute to answer the hardest questions.
In Google, even though we call it 10 blue links, you get annoyed if you don’t even have the right link in the first three or four. The eye is so tuned to getting it right. LLMs are fine. You get the right link maybe in the 10th or ninth. You feed it in the model. It can still know that that was more relevant than the first.
What is the weakness of Google is that any ad unit that’s less profitable than a link, or any ad unit that kind of disincentivizes the link click is not in their interest to go aggressive on, because it takes money away from something that’s higher margins.
And also this is where I believe the whole prompt engineering, trying to be a good prompt engineer is not going to be a long-term thing. I think you want to make products work where a user doesn’t even ask for something, but you know that they want it and you give it to them without them even asking for it.
Or is it really that we’re building some super machine in a box that’s going to be smart and kill everybody? It’s not even a science fiction narrative. It’s a bad science fiction narrative. I just don’t think it’s actually accurate to any of the technologies we’re building or the way that we should be describing them.
But many of these gains have already been achieved: the best data centers already use just 10% of their electricity for cooling and other non-IT equipment.
In particular, operating a data center requires large amounts of electricity, and available power is fast becoming the binding constraint on data center construction.
So if you remember this information in different times in different places, it’s more accessible at different times in different places because it’s not overfitted in an AI way of thinking about things. It’s not overfitted to one particular context. But that’s also why the memories that we call upon the most also feel like they’re just things that we read about almost.
I think a lot of the problems that come up with technology aren’t the technology itself, as much as the fact that people adapt to the technology in maladaptive ways. I mean, one of my fears about AI is not what AI will do, but what people will do.
I don’t think we’ve replicated human intelligence, unless I know that the simulator is making exactly the same kinds of mistakes that people do, because people make characteristic mistakes. They have characteristic biases, they have characteristic heuristics that we use, and those have yet to see evidence that ChatGPT will do that.
For now, I am not convinced that issuing press releases about your compounds that talk about their discovery through AI techniques is sufficient to expect greater things from them.
That is, labs should make sure that the safety measures they apply to their powerful models prevent unacceptably bad outcomes, even if the AIs are misaligned and intentionally try to subvert those safety measures.
The basic problem with evaluating alignment is that no matter what behaviors you observe, you have to worry that your model is just acting that way in order to make you think that it is aligned.
We're advocating that companies handle risk from scheming models in a similar way–striving to ensure that they'll be safe even if their alignment efforts fail to prevent models from scheming.
What a large language model is trying to do is to predict text. That’s what it does. And it is leveraging the fact that we human beings for very good evolutionary biology reasons, attribute intentionality and intelligence and agency to things that act like human beings.
And rather than trying to ask, how close is it to being a human-like intelligent, we should appreciate it for what its capabilities are, and that will both be more accurate and help us put it to work and protect us from the dangers better rather than always anthropomorphizing it.
I think this is one of the biggest things that physics can help with, and it’s an obvious kind of low-hanging fruit situation where the heat generation, the inefficiency, the waste of existing high-level computers is nowhere near the efficiency of our brains. It’s hilariously worse, and we haven’t tried to optimize that hard on that frontier.
The 100-megawatt facility might have ten 10-megawatt blocks served directly by a 1500V DC feeder line without connections to other blocks that would require extra switches and routing.
Retrieval-augmented generation (RAG; Lewis et al., 2020), which conditions on the LLM's generation on retrieved documents is the most practical paradigm IMO.
Your product might be great, but how do you get it on the shelves of Walmart? Or how do you get it on the shelves of Google? Because if you're not on the shelves of Walmart, you're not on the shelves of Google, you might be invisible. And it's friction that kept you there, not some sort of merit contest.
I continue to not want super-voting control over OpenAI. I never have. Never have had it, never wanted it. Even after all this craziness, I still don’t want it. I continue to think that no company should be making these decisions, and that we really need governments to put rules of the road in place.
And oftentimes it's more important to them to have the public perception that they're good directors so they get the next best deal. If they have a reputation for taking on management too aggressively, word will get out in the small community of founders and they'll miss the next Google.
We need to find ways to work with this new tech to get the best of humans and the best of computers, together, rather than thinking in terms of one or the other.
Longtime tech reporter Matt Lynley offers an uncommonly technical assessment of new AI developments in his newsletter Supervised. Original reporting and a fast cadence help this one stand out from other AI newsletters.
I also read a ton of technical books last year, but the most impactful one for me was Neural Network Methods for Natural Language processing by Yoav Goldberg
Diverting attention and resources from global health and poverty is an enormous gamble, as it will make many lives poorer, sicker, and shorter in the name of fending off threats that may or may not materialize.
Unlike other interventions EA has sponsored, there are scant metrics for tracking the success or failure of investments in existential risk mitigation.
We do know that humans are doing something different from these models, in part because we're so power efficient. The human brain does remarkable things and it does it on about 20 watts of power. And the AI techniques we use today use many kilowatts of power to do equivalent tasks.
And so the only way to improve safety is to have an escape system. And historically, human-rated rockets have had escape systems. Only the space shuttle did not, but Apollo had one. All of the previous Gemini, etcetera, they all had escape systems.
It's interesting to me that large language models in their current form are not inventions, they're discoveries. The telescope was an invention, but looking through it at Jupiter, knowing that it had moons, was a discovery.
If I were forced at gunpoint to guess, I'd say that human-level AI seemed to me like a slog of many more centuries or millennia
Their words now
I was wrong, of course, not to contemplate more seriously the prospect that AI might enter a civilization-altering trajectory, not merely eventually but within the next decade.
If I were forced at gunpoint to guess, I’d say that human-level AI seemed to me like a slog of many more centuries or millennia (with the obvious potential for black swans along the way).
Their words now
I was wrong, of course, not to contemplate more seriously the prospect that AI might enter a civilization-altering trajectory, not merely eventually but within the next decade.
Large language models are more or less generalizations of stuff that we have. There’s still breakthroughs in AI waiting to happen, and maybe they are happening and maybe they’ll be good, maybe not, but that’s not quite the same.
A korrent is a belief a person has stated in their own words: one
sentence stating the claim, backed by a quote and a source, kept at
korrents.com.
Under a name here, the quoted block is what they actually said.
The korrent beneath it is the claim those words support, in
korrents' wording — tap it to see the record, its source, and who
else holds it.
Nobody here wrote their own korrents. They are compiled from public
statements, and a person can change their mind, which is recorded too.
About the English under a post
Some people here publish in a language other than English. Where they
do, this site shows a machine translation beneath the post, in
this typeface — the site's own, not theirs.
The post itself is never changed, moved or hidden: what is set in the
serif above is exactly what the person published, and it is what to
quote them on. A translation can be wrong in ways that matter,
especially about tone.
Only the post's own words are translated. A quoted post, a linked
article and a belief on korrents.com
are left in their original language.