Bill Gates Reveals Superhuman AI Prediction

Sam Altman reported to Gates that OpenAI is seeing productivity improvements of up to 300% among their developers, and other sectors are reporting 25-50% productivity increases.

“I think I think Sam Alman said on your on your podcast unconfused me which I enjoy uh that there's seeing a productivity Improvement of up to 300% I think among their developers uh and in other sectors I think we've seen reports of you know 25 50% increases in productivity”

The barriers to entry in AI product development are uniquely low; there is currently unprecedented capital flooding into AI as a new category, with companies raising $6 billion in a single funding round and hundreds of millions across many others—more capital and talent pursuing a new technological category than in the internet or early auto industry booms.

“it's important to distinguish two parts of economic activity one is the economic activity building AI products and both Bas level AI products and then vertical AI products and we can say for sure that the barriers to entry are uniquely low uh in that we're in this Mania period where you know somebody literally raised $6 billion dollar in cash you know for a company and many others raised hundreds of millions and you know so the idea that there's you know there's never been as much Capital going into a new category you could even say a new Mania category I mean this makes the internet or the early Auto industry Mania look you know quite small in terms of the percentage of IQ and and the valuations”

The question of whether AI empowers the little guy or the big guy is complex: big companies dominate AI model development, but everyone has free access to GPT-4 Omni, creating an equalizing element.

“we're seeing that there just a few big big companies seem to be the dominant players in the development of the technology um but on the other hand it does seem that everyone has access to gp4 Omni um at now for free uh so there there's also an equalizing element”

Open-sourcing sophisticated AI models creates risk that malevolent actors will have access to advanced AI, and defensive applications may be better served by restricting access to frontier models.

“you know like the people who say oh it's fine that it's open source you know they're willing to say well okay if it gets you know too good maybe we'll stop open sourcing it but you know will they know what that is uh and would they really say okay maybe the next one um you know so you you pretty quickly go to let's not let people with Mal intent benefit you know from having a better AI than uh you know the sort of Defense good intent side of you know cyber defense or uh War defense or biot Terror defense”

Nvidia, a chip design company (not a chip manufacturer), added $1 trillion in market value in six months, demonstrating the magnitude of value creation in AI infrastructure.

“I mean you know there was no company before the turn of the century that had ever been worth a trillion dollars here we have one ship company who doesn't make chips it's a chip design company uh that in six months adds a trillion dollars of value”

Translation has historically been a holy grail capability that would generate tens of billions in revenue; now AI companies are providing free, arbitrary audio and text translation as an afterthought feature, demonstrating the magnitude of the productivity transformation.

“you know let me try and let me try and I think like you know I think it's a the comment when people say that not withstanding what you just said bill they they're creative and their naysaying capabilities um because I think that you're your response is accurate um for sure it's like when the car was developed you know you it's the second order effect when the car was developed it could get you from point A to point B um and you might even be able to predict uh the development of roads and highways Etc but you might not be able to predict Los Angeles or suburbs um and uh all you know drive-in movie theaters and in the case of I think when in more modern um stance the worldwide web came OG and there were lots of brochure wear and you know there was uh travel agent Expedia came along and that was all sort of like run-of-the-mill first order effect but people point at Uber as a second order effect on the technology on that that sort of was like oh you couldn't have predicted that now maybe you could maybe you couldn't but that's what I think that's what Adam's question I think is going for when you look at AI I mean per in many ways search is the game of search has already changed which is ubiquitous um um uh consumer activity uh and certainly chat gbt was a Monumental the fastest growing technology uh adopted uh technology in our ever so I don't I'm not minimizing or giving credence than the Nays but it's really about the second order effects now we yeah we you know chat gpt3 was not that interesting I mean it was interesting enough that a few people that open the eye felt the scaling effect would cross a threshold and I I didn't predict that and very few people did”

The personal agent will be a superior AI assistant embedded in form factors like earbuds and glasses, operating at a much higher semantic level than today's software; it will understand context, anticipate what you need, and serve as executive assistant, therapist, friend, girlfriend, and expert—all driven by deep AI.

“the personal agent you know that I've been writing about for decades that's you know Superior to a human insistant in that it's uh tracking and reading all the things that you wanted to read um and you know just there to help you and understands the context enough that you know silly things like you don't trust software today to even order your email messages it's in a you know stupid dumb time ordered form because the contextual understanding of okay what am I about to do next what's the nature of the task that these messages relate to you don't trust software to combine all of the you know new information in including uh new Communications you know you go to your mail and that's time ordered you go to your text and that's time ordered you go to your social network and that time ordered I mean computers are operating at a a almost trivial level of semantics in terms of understanding what's your intent when you sit down with the machine or helping you with your activities and now that they can essentially read like a white collar worker uh that interface will be entirely agent driven you know agent executive assistant agent mental therapy agent friend agent girlfriend agent expert all driven by uh deep Ai”

The GPT-4 demo Bill Gates saw in September was more impactful than the Xerox Park graphical user interface demo he witnessed in 1980, because unlocking a new type of intelligence that can read and write is more fundamentally transformative than graphical interfaces, which are now taken for granted.

“I'd say yes I mean I'd seen graphical interface prior to the xero park stuff and um that you know was an embodiment that helped motivate a lot of what apple and Microsoft did uh with personal Computing in the you know the decade after that uh you know people kind of take it for granted today that that we have those interfaces but you know compared to unlocking a a a new type of intelligence that can read and write uh Graphics interface is is clearly less impactful uh which is saying a lot”

Concrete AI applications already exist and are widely useful: meeting summarization, translation, programmer productivity enhancement, support call optimization, and sales call improvement; the idea that no breakthrough applications exist is false—these use cases are already delivering concrete value across white-collar work.

“I mean I agree they they don't think you know no yeah summarizing meetings or you know doing translation or making product programmers more productive I mean that's it it's mindblowing you know this is White Collar capability with a footnote that if if in many open Ed scenarios it's not as reliable as humans are”

Gates does not have a solution for the possible future problem of determining human purpose in a world where machines can solve problems better than humans; he views it as a very important problem that people should contemplate, but doubts anyone of his age and immersion in scarcity can imagine a post-scarcity world.

“I don't think somebody who spent you know 68 years in a world of shortage uh and you know okay we use Market mechanisms to you know deal with the shortage we use various incentive structures to innovate to create more Supply capacity I doubt that either at that absolute age or having been immersed in in such an utterly different environment that the ability to imagine you know this post shortage uh type world will come from anyone near my age um so right you know I view it as a very important problem uh that people people should contemplate but no I I that's not one uh that um I I have the solution”

AI could help bridge political polarization by providing tools to identify misinformation, highlight bias, and help people understand the views and concerns that have pushed people on the other side to vote differently.

“certainly if somebody wants to understand okay uh where did this come from this article or this video you know can you what is the Providence you know is that provably a reliable source uh or is this information accurate or you know in general in my news feed you know what am I seeing that you know somebody who's voting for The Other Side what have they seen uh and you know try to explain to me what has pushed them in that direction uh you'd hope that you know sort of the again going back to the Paradigm of uh White CER capability being uh almost free that you know well-intended people who want to bridge those misunderstandings would have the tools of AI to highlight misinformation for them or highlight uh bias for them”

No slowing mechanism exists that is both plausible and feasible for AI development, given individual, company, and government-level incentive structures that push toward advancement.

“you know as Mustafa writes in his book the incentive structures don't really have some mechanism that's all that plausible of of how that would happen given the individual and uh company and even government level thing”

Chat GPT-3 was not particularly interesting, and the transformative moment came with GPT-4; the scaling effect crossed a threshold less than two years ago with the general availability of GPT-4, so AI disruption is very recent.

“we you know chat gpt3 was not that interesting I mean it was interesting enough that a few people that open the eye felt the scaling effect would cross a threshold and I I didn't predict that and very few people did and we only crossed that threshold less than two years ago uh a year and a half in terms of General availability”

AI technology is advancing at superhuman levels in capability but not yet in reliability; adding metacognition—the ability to step back and reason about how to think about a problem—will solve the erratic nature of current AI genius and is the key frontier beyond mere scaling.

“you know so this technology is in terms of its capability is and it will reach um even modulo reliability will reach superhuman levels we're not there today if you put in the um the reliability constraint a lot of the the new work is adding a level of metacognition that done properly will solve uh the sort of erratic nature of the the genius that uh is easily available uh today in the white colar realm”

AI will be the most important thing happening and the dominant change agent, shaping humanity in a very dramatic way, alongside synthetic biology and robotics (also controlled by AIs).

“and even more so it's absolutely the most important thing going on it'll it'll shape Humanity in a very dramatic way it's at the same time that we have you know synthetic biology and Robotics you know being controlled by the AIS you know so we have to keep in mind those other things but the dominant uh change agent will be AI”

Gates' main excitement about AI is driven by the severe shortage of white-collar workers in the Gates Foundation's work on global health in subsaharan Africa and developing countries, and the lack of teachers who can engage deeply with students in their native language; AI can address these shortages at modest server costs via mobile infrastructure.

“all these shortages there's no organization that faces white collar shortage as much as the Gates Foundation where we look at Health in uh subsaharan Africa or other development countries or you know lack of teachers who can engage you in a deep way you know ideally in your native language um and so the idea that modulo whatever the the server cost is that by using the mobile phone infrastructure that you know continues to drive pretty significant penetration even in in very poor countries the idea that medical advice and personal tutors can be delivered uh where you know because it's meeting you in your language and your semantics there isn't like some big training thing that's taking place there you just pick up your phone and listen uh you know to to what it's saying so you know it's very exciting”

Governments will likely focus on concrete near-term AI issues—copyright rules, deepfake abuse, unreliability in health diagnosis or hiring decisions—rather than addressing the deeper, slower-moving challenge of managing productivity-driven labor transitions.

“now governments will will take the things that are most concrete like what are the copyright rules or what are the abuses of deep fakes or you know in some applications does the unreliability say health diagnosis or um um hiring decisions you know mean that you ought to move more slowly or or create some uh liability for those things they'll tend to focus in on those short-term issues which you know that's fine but you know the biggest issue has to do with the adjustments to productivity that overall you know should be a phenomenal opportunity if political capacity and the speed which which it was coming uh were were paired very well”

When AI tools are applied to other industries (not AI itself), there's hope that AI will level the playing field or not between small and large players, delivering the same service quality at lower cost to customers.

“once you leave the AI tools domain which as big as it is is a Modest part of the global economy how that gets applied to okay I'm a small Hospital Chain versus a big hospital chain you know now when I have these tools to that you know level the plane field or not you would hope that it would uh and that you can offer it for the same price are less a far better level of service all of these things are in the further furtherance of getting the value down to the customer”

People are already disclosing large amounts of personal information into digital systems through emails, online meetings, and phone calls, so enabling an AI agent to access audio from one's life would offer immense value in summarizing meetings and follow-ups, with partitions available for different types of information depending on user preference.

“computers today see every email message that I write um and certainly digital channels are seen you know all my online meetings and and phone calls so you're already disclosing into digital systems uh a lot about yourself and so yes the value added of the agent um in terms of summarize that meeting or help me with those follow-ups um you know be phenomenal and the agent will have different modes in terms of which of your information it's able to operate with so there will be partitions that um you have but for your essentially executive assistant agent you know you won't exclude much at all from that that partition”

Blue-collar job substitution from AI will be delayed relative to white-collar disruption, which is fortunate because it means the more educated workforce experiences AI disruption first and may have more capacity to adapt.

“you can almost say it's good they that the blue collar job substitution stuff is more delayed than the the white collar stuff so that you know it's not just anyone sector and actually it's the more educated sector that seen these changes first”

AI adoption will not face significant impedance barriers because the software meets users where they are with natural language rather than requiring users to learn new interfaces (like menus or formulas); uptake will be driven by intuitive voice and language interfaces that work in users' native context.

“I want to make an image okay what do I have to learn I have to learn English uh this is the software meeting us not us meeting the software you know so it's not like there's some new menu you know file edit window help and oh you got to learn that you have to type the formula into the cell this is you saying hm I wish I could do data an analysis to see which of these you know products is responsible for a Slowdown and it understands exactly what you're saying so the idea that there's a impedance of adoption uh it's not the normal thing”

Company processes and organizational habits are barriers to AI adoption, but these are surmountable through observation-based learning; salespeople can learn to use advanced AI tools through a week of watching expert users without formal training or manuals.

“company processes um that are very used to doing things the old way uh will have to adjust but if you look at you know Tes support Tes sales uh data analytics you know give somebody a week of watching a an advanced user and you know say no Manual of any kind just you know learn by example of how the stuff has been used the uptake assuming there's no limit in terms of the you know server capacity that connects these things up which I don't expect certainly in rich countries there'll be a gigantic limitation there”

A CEO whose top salespeople reported their most time-consuming task was drafting follow-up emails created an AI instance that transcribes all sales calls and auto-generates follow-up emails using best practices, allowing the best half of the sales team to work twice as efficiently while laying off the other half.

“I know one of those crazy white collar workers who who's a CEO of a company that's growing very quickly who ask his top salespeople what takes you the most time during this day and they said writing drafting follow-up emails following sales calls and he created an instance of GPT to you know pulled in all their all their best practices best Communications automatically transcribes every phone call and automatically generates the follow-up email uh and he he's he's laying off half of his sales team so that the best half of his sales team uh can do can work because they can now work twice as efficiently”

If it were hypothetically possible to stop AI development exactly where it is now, it would take 10 years for companies and individuals to fully apply the technology that currently exists.

“even if it were possible hypothetically to stop AI development exactly where it is right now it would probably take 10 years of forum 3 and other folks helping companies and individuals figure out how to apply the technology that currently exists right I mean the the we already have an AI That's powerful enough I think to have profound implications uh it it feels it feels like we could easily benefit from 5 to 10 years just thinking of all the useful applications of the the level of AI we have today”

AI will be an accelerant for accomplishing in 5-10 years what has taken 20 years in healthcare and education; the last 20 years of progress were largely low-hanging fruit from getting vaccines cheaper and ensuring distribution, but now facing tougher issues like malnutrition and HIV vaccine development where AI will help with upstream discovery and advice, delivery, diagnosis, and scientific discovery.

“we you know was pretty miraculous in that we cut child to death in half from 10 million a year to 5 million a year that was largely by using getting tools like certain vaccines uh to be cheaper and making sure they were getting to all the world's children and so that was kind of loow hanging fruit and now we have you know tougher issues but with the AIS um the the Upstream Discovery part of okay why do kids get malnourished or and why has it been so hard to make an HIV vaccine yes we can be you know way more uh optimistic about those those huge breakthroughs”

You never want productivity to go backwards; even with distribution challenges, running the clock backwards to lower productivity is not desirable.

“you you know you'd never want to run the clock backwards and say you know thank God we were less productive 20 years ago”