Alexandr Wang - A Primer on AI - [Invest Like the Best, EP. 272]
What this covers
Alexandr Wang, founder and CEO of Scale AI, talks with Patrick O'Shaughnessy in what is framed as a primer. It begins with geopolitics: the shift from two decades of counterterrorism toward great power competition, and China's faster integration of AI into specific uses, including facial recognition built by government contractors. Wang rejects the nuclear-weapon analogy. He treats AGI as a broad, software-like capability that gives roughly 10X better decision-making, not a discrete weapon. The conversation then moves to mechanics: collecting raw data, annotating it, training, deploying, and repeating the loop because models are brittle.
The middle of the episode covers the range of what AI can and cannot do. Wang says what is easy for algorithms is largely orthogonal to what is easy for humans, which he thinks fuels needless fear. He cites DeepMind's AlphaCode beating most competitive programmers, systems passing standard engineering interview questions, and GPT-3's reliance on two decades of Reddit text. He expects software work to converge on product management. Talent, compute, and data are named as the three requirements for scaling. Contested claims include talent scaling further in AI than in software and the Scale Versus Democratization Curves, where larger models favor incumbents while falling costs spread access, with inequality as the risk.
The closing stretch turns to company-building. Wang describes borrowing the Amazon Playbook, running many products in parallel on strong primitives, against the usual advice to focus on a few things. He also offers a personal heuristic for judging whether someone cares: look at the small details at the edges of their work.
Wang argues that data, not code, is becoming the main source of differentiation in AI, and that AI's economic value comes from scaling repetitive human tasks, with data availability determining which tasks get automated first.
- Code has become a commodity across AI tasks, so proprietary data sets and data engines now drive competitive advantage
- Tasks with abundant digital data are automated first, while physical tasks with scarce data remain hard
- AI's value comes from taking human tasks from one to N, with a potential impact 10X to 100X larger than SaaS
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
A common misconception is that tasks easy for humans will be easy for AI; in fact, what algorithms find easy is largely orthogonal to what humans find easy, which causes unnecessary fear about AI.
“the intuitive belief is that the things that are easy for humans to do are going to be the things that are easy for AI and machine learning to do, which is absolutely not the case”
Building an AI model involves collecting large amounts of raw data, annotating it into structured form, training the model, and deploying it, and because models are brittle, this must be a continuous loop of recollecting, reannotating, and retraining rather than a one-time build.
“The very important piece is, this is not a sort of like one way process this is actually a loop”
Chinese government contractors such as SenseTime and Face++ built world-class facial recognition computer vision for state national objectives, and China has integrated AI into specific use cases faster than the US.
“Facial recognition technology, I think is one of the primary examples where this company's like SenseTime, Face++, et cetera, that are primarily contractors of the Chinese government”
Talent has a greater scaling effect in AI than in software, not because AI requires brighter minds, but because AI systems have a much higher ceiling on impact and adaptability, so a small team of experts can generate outsized economic value.
“the cap on the level of impact or the cap on the amount of utility to be delivered is much, much higher”
The highest-performing algorithms for image recognition, speech recognition, text summarization, and question answering all use essentially the same code base, so code has become a commodity and differentiation comes from data.
“they actually all use the exact same code base. That's been this very meaningful shift”
Two competing curves shape AI's effect on society: benefits that favor scale, as bigger models perform better, versus falling fixed costs and democratization of the technology, and how these intersect will determine outcomes.
“there's these two competing curves in AI... One is the benefits to scale and the other is lowering of fixed costs and the democratization of the technology”
AI can perform very difficult human tasks like competitive programming and proving math theorems because abundant visual data exists and outcomes can be verified or simulated, even though these tasks are premium human skills.
“there are incredible pools of visual data as well as abilities to verify or simulate the outcomes here, which allow these algorithms to reform actually incredibly”
Broad scaling of AI depends on three critical components: talent, compute and computational power (vulnerable to supply chain disruptions), and data and data scale.
“There's three critical component to broad scaling of artificial intelligence. One is certainly talent... The second is compute and computational power... Then the last one is data and data scale.”
Contrary to the conventional business truism of focusing on a few things, Amazon's approach of parallel execution across many products built on strong primitives is more powerful, and Scale follows that approach.
“One of the great lessons from Amazon as a business is that parallel execution is incredibly powerful.”
The total addressable market for AI and machine learning is astronomical, at minimum 10X and probably 100X the total business value that SaaS and software have generated historically, because AI can automate repetitive enterprise work and exceed trained humans.
“I think that is at minimum 10X, probably 100X the total business value that has been generated by SaaS systems or software historically”
AI systems can now pass the classic programming interview questions Silicon Valley uses to hire engineers, so a Turing-style test of programming competence has been met.
“if your touring test is a set of programming interviews to get a job at Google, you probably have a machine learning system now that can pass that touring test”
The alchemy of software, namely large-scale data collection, simple workflow tooling over data such as CRM, and near-infinite scalability, has produced major enterprise value but is capped by software saturation.
“these primitives or these forms of alchemy have some cap, that's for the saturation of software”
Technologies that favor scale generally increase inequality, and because AI benefits from bigger models and more data, society must watch this dynamic closely and counter it through democratization.
“Generally technologies that favor scale do result in greater levels of inequality. I think that's something that we really need to watch out for.”
DeepMind's AlphaCode performs better than the median competitive programmer and beats about 99.9 percent of humans at competitive programming tasks, a capability Wang said he would have doubted five years earlier.
“they can solve these competitive programming problems better than the median competitor... Beating the competitors is probably beating 99.9% of humans”
The past 20 years' paradigm of counterterrorism is shifting to great power competition among a small set of technically capable states, making technological excellence critical across economic, ideological, and military competition.
“this shift from the paradigm of the past 20 years being counter-terrorism, this shift from counter-terrorism to great power competition”
The real secret behind GPT-3 and large language models is that they leverage roughly two decades of Reddit data, which represents a vast pool of human-typed digital text.
“the real secret behind it is that it leverages two decades of Reddit data”
Annotation converts large pools of unstructured data, such as video of pedestrians or long text, into structured labels that algorithms can learn from, and the work is moving from manual labor toward automation that pre-labels data and flags human errors.
“Then it goes through this process of annotation, which is the conversion of this large pools of unstructured data to structured data”
As the share of software that is AI-infused rises from 0.01 percent toward 50 percent, the main source of competitive differentiation shifts from proprietary code to a company's existing data assets and its engine for continuously producing differentiated data.
“as you go from 99.99 to 90:10 or 80:20, or even 50:50 over time, the vector of differentiation totally shifts to data”
Algorithms excel at tasks where large pools of digital data exist or can be easily collected, while tasks lacking digital data and expensive to capture, like physical chores, will be the last to be automated.
“generally that more or less boils down to data availability, which is that where there are large pools of digital data an algorithm can learn from”
The lower-bound impact of AI is scaling repetitive human tasks from one to N, which is economically valuable and frees humans, who generally dislike repetitive work, to be more creative and leveraged.
“take repetitive human tasks instead of going from zero to one, go from one to N human work”
In traditional software the lifeblood is code that dictates system behavior, whereas in AI and machine learning the lifeblood is data, which is the ingredient that determines what the system learns and does.
“We often say this thing, which is data is the new code.”
AI is a multifaceted, software-like general technology rather than a discrete weapon, so AGI is not the equivalent end state to a nuclear weapon. Its strategic effect is better understood as enabling roughly 10X better decision-making and action.
“I don't quite fully agree with this analogy that an AGI is the equivalent end state to a nuclear weapon”
Software work is converging toward product management, where the differentiating skill becomes understanding what to build rather than building it, and tools like GitHub Copilot already accelerate programming.
“programming and product management, in some sense, we're going to keep converging where differential tasks or the differential skill is in understanding what to build”
You can identify whether someone truly cares by noticing their attention to small details at the edges of their work, such as the minor flourishes in a performance, and by asking what details took most of their time.
“you notice it along the edges, so to speak. You notice it in maybe the depth of thought, the depth of obsession, the attention to detail”
The greatest macro shift in conflict is the move from physical battlegrounds to primarily digital ones, where perhaps 90% of conflicts get resolved or contested in a purely digital context long before any physical conflict arises.
“we're moving much more to a paradigm where hopefully 90% of conflicts are actually resolved or battled in a purely digital context long before an actual conflict arises”
AI could give a country roughly 10X better strategic decision-making and action, analogous to a chess player facing a superior algorithm, and this advantage would tend to dominate extended conflicts.
“AI and AGI, the best way to think about is it enables maybe 10X better strategic decision making”
Business leaders should first identify the data that powers their business and the highest-value problems AI could solve, such as recommendations and expensive repetitive processes, and should partner with firms holding scarce AI talent, since the human capital shortage will last for many decades.
“the big questions are really around understanding the data that powers your business”
A good data-gathering engine rests on three tenets: quantity is a quality of its own, data must be curated to find the rare valuable 'needles' in the haystack, and the strategy must continually expand data diversity rather than collecting the same data.
“first quantity is a quality of its own... You'll want strategy for data diversity”
The Amazon approach of taking complex problems that are half operational and half technological, and adding substantial technology and automation, yields orders-of-magnitude better performance than operations alone, and this is the playbook for AI data work.
“They've taken both of these very large complex half operational half technological problems, and they've added significant amounts of technology and automation”
When Scale visited machine learning teams across industries, data quality and data volume were consistently among the top three problems, and the gaps were both technological and cultural, since ML engineers are often unmotivated by data work.
“always in the top three data quality and data volumes, one of them, and you'd dig into it”
Artificial intelligence has a compounding, order-of-magnitude effect on virtually any other technology, which is why it matters so much for national competitiveness.
“Artificial intelligence is just one of those technologies that happens to have compounding order of magnitude effect on any of the other technologies”
The unclear relationship between US technologists in the private sector and the government over AI for defense and intelligence is a significant cultural problem that must be solved for the US to be well positioned in AI, with Google's Project Maven as the clearest example.
“The Google cloud project Maven conflict was probably the most visible, most clear example of this”
Every firm will need its own version of a TikTok-style recommendation algorithm for every economic transaction, identifying the next thing a customer wants to transact with, which will make trade significantly more efficient.
“Imagine TikTok recommendation algorithm, but for like every economic interaction or every economic transaction in your life”
Data annotation has historically been done in a low-tech, manually intensive way because most technologists avoid operational messiness, choosing computer science for its more constrained problems.
“most people who go into computer science... go into computer science because it's a significantly more constrained problem”
Scale builds core primitives such as data annotation, data management, and model monitoring, then combines them into customer products, mirroring how AWS combined storage and compute into services like RDS.
“they started with the primitives of storage and compute, and then they sort of combined those primitives in many interesting ways”
In machine learning, the imaginative and creative work lies in choosing the label or outcome being targeted, not in the features used to predict it, so questions become more valuable than answers.
“it's the label, it's the outcome that you're targeting that requires all the imagination and creativity, not the features”
Most firms could not obtain high-enough-quality data sets to build great algorithms, making data annotation a massive bottleneck for overall machine learning progress, and Scale's initial mission was to fix it.
“what we noticed is that it was this massive bottleneck for the overall progress of machine learning is that most firms could simply not get high enough quality data sets”
Because of the pace of AI acceleration, what the US and other powers do over the next 5 to 10 years is probably about 10X as important as what was done in the prior 5 years, and will set the stage for the next phase of great power competition.
“I think what we do over the next five years or do over the next 10 years is probably 10X as important as what was done over the past five years”
The long-term effect of superior AI-driven recommendations will be a general reduction in marketing and sales expenses, because algorithms will better predict what users want next and reduce the need for active sales.
“The long-term impacts of that you could think of as like a general reduction in marketing expenses or sales and marketing expenses”
OpenAI, with roughly 250 total staff, is a small number of people to generate the broad, multi-decade scaled impact the company is expected to deliver.
“their total head count is 250 people as a corporate”
Business ROI from high-quality AI systems is expected to be about 10X the value generated by deploying a CRM or ERP system, driven by a proliferation of high-impact enterprise use cases.
“the business ROIs generated from high quality AI systems are going to be 10X more than the business value generated by let's say, deploying a CRM or deploying an ERP system”
Deterrence has been an incredibly positive force for world peace and a stable world order over the past 40 to 50 years, as exemplified by the atomic bomb's deterrent effect.
“generally speaking, deterrence has been a incredibly positive thing for world peace and the stable world order over the course of the past 40, 50 years”
Major world events such as the Russia-Ukraine conflict do not unfold in predictable logical ways and function as tail risks for human well-being.
“they do not happen in predictable ways. There's not a predictable logical way in which events will play out over time”
Modern deep learning was developed primarily in the private sector at Google, DeepMind, OpenAI and Facebook, with contemporary research in China as the other major hub.
“modern, deep learning was initially developed primarily in the private sector through a lot of incredible research that was done at Google and DeepMind and OpenAI and Facebook”
Scale's approach was to offer ML engineers a developer API that delivers as much high-quality data as needed, hiding the operational dirty work behind the scenes, following the AWS playbook.
“We built great developer APIs and a great developer platform that enabled the machine learning engineers... to basically hid an API, get as much high quality data as they needed”
Technology is morally neutral and can be used for good or bad, and the concern is what leverage it gives people who should not have it, with broad outcomes so far having been positive.
“I'm of the view always that technology is sort of like a morally neutral, it can be used for good or bad”
Cybersecurity is a very challenging attack plane because it is almost impossible to defend against, and attacks are numerous and can generate many attack vectors.
“cyber security which you mentioned, which is a very challenging general attack plane because it's generally speaking, almost impossible to defend against”
Highly inspired people show a recurring pattern across their lives, moving from one source of inspiration to the next with full force and will, which can be detected by hearing how they talk about their past.
“the people who are more inspired you'll notice this like theme in their life. They find things that inspire them and they find the next thing”
Scale values itself primarily on the customer outcomes it generates, not only on growing its data primitive, because AI has been overhyped with few firms focusing on actual customer value.
“that being a primary metric fund which we value ourselves”
If the Turing test is a set of programming interviews for jobs at Google, AI systems can now pass it, and this reverses the role of programmers who take pride in telling machines what to do.
“if your touring test is a set of programming interviews to get a job at Google, you probably have a machine learning system now that can pass that touring test”
Attracting high-skilled immigration is one of the big, easy-to-change inputs for American AI leadership, with consequences lasting decades because people want to build lives and families in America.
“this is one of the ones that has again, multi decade consequences”
AI and machine learning are a cross-sectional technology that can intersect with intelligence gathering, actual warfare, and cybersecurity, making them very significant across all of these fronts.
“I think AI and machine learning is a very, very significant one. Its ability to intersect with many of these attack planes”
Tesla's autonomous vehicle strategy treats all its cars as collection vehicles that encounter rare situations, producing a diverse data set that benefits long-term machine learning differentiation, though Tesla faces other challenges.
“all of the Teslas in the world are in some sense collection vehicles, they're all encountering very rare situations”
A violin teacher who ended a lesson for lack of practice taught Wang that if you do something, you must do it with full passion and force, or it is not worth doing, a belief that shapes how Scale hires.
“If you're not practicing at all, we should stop wasting your parents' money and just stop on lessons.”
The 'dirty jobs' of AI data work involve converting data formats that machines cannot read, such as transaction records and audio intent, into formats that machine systems can use, and these tasks are the hidden foundation of AI.
“all these dirty jobs, they fundamentally start with some data format that machines can't read today and making them effectively legible”
Home robots that fold laundry and dishes will take a very long time to arrive, while AI systems that are world-class copywriters already exist today.
“it's going to be a very, very long time before we have home robots that can do things like fold your laundry and your dishes”
A teacher's visible love of art and craft is infectious, and the lesson that passion spreads is a learning Wang carries from his violin teacher.
“I think one learning is just that passion is incredibly infectious thing”
The intern analogy for AI tasks is useful: work that is complicated enough that a person is on it now but simple enough to hand to an intern is likely automatable.
“anything that an intern could do for you, you might be able to scale up through one of these models”
The actions taken all day in Word and Excel are among the most likely to be automated, and much of that work is uninspiring.
“if you spend all day in a Word Processor, all day in Excel, the actions that you take in those products are some of the most likely to be automated”
Despite widespread distrust and claims that AI is snake oil, AI is an incredibly powerful technology, and the key gap is firms that focus on actual customer value creation.
“There's a lot of distrust of AI systems. There's a lot of belief AI isn't good or AI is so bad or AI is just this snake oil technology. That is absolutely not true.”
Success for Scale in ten years means the majority of companies in the world can use AI effectively for high-impact applications, with many of those powered by Scale products, whether directly or through SaaS partners.
“in 10 years are the majority of companies in the world able to effectively use artificial intelligence in some very high impact application”
Given the 1960s-era low jobless claims and the observation that automation has not yet hurt employment much, the claim that knowledge work is first to be automated is plausible though its effects are slow.
“I think that as of this morning, jobless claims are at their like all time low or since the 1960s or something”