In this issue: - Technologies as Cleverness Substitutes—Manufactured products are a way to provide mass access to the cleverness of whoever invented them. Every general-purpose tool reduces the need for smart topic-specific solutions, and also changes the shape of collective R&D investment.
- Sin Taxes—Sometimes when a tax works as intended, it's a problem.
- Optionality—Sometimes it pays to be subscale.
- The Dow—Nvidia joins the benchmark.
- The GPU Financing Ecosystem—Credit is always a bet on a narrow slice of outcomes, so it's harder when the distribution of outcomes is fat-tailed.
- AI Fragmentation—Nvidia still has a strategic imperative to ensure that there are many winners in AI.
Technologies as Cleverness Substitutes
The second industrial revolution was probably the peak of humanity's capacity for clever self-contained solutions to physical problems. Rising abundance and more precise dies, molds, and lathes meant that it was physically possible to mass-produce these sorts of devices, and cheaper long-distance transportation by rail and ship meant that markets were larger, so the payoff from this cleverness was higher. And the result was a profusion of devices with embedded cleverness. The can opener, for example, went through a long evolution: in the beginning, people improvised (and cans have gotten thinner since then—canned food was a good idea for people who were very hungry and had tools available); after that canned food used to come with a can-specific twist key opener; claw-based can openers started showing up in the mid-19th century; and the earliest design recognizable as a modern can opener showed up in 1870 and evolved substantially to reach its modern incarnation.
The Kodak camera is an even more extreme example. Film photography is an absurd cross-disciplinary effort that requires good lenses, control over the shutter, chemistry (knowing how silver halides react to light), other chemicals for developing and fixing the film, building a lightproof camera body, and then chemistry again for the base that needs to be both durable and flexible. It's a mixture of technologies that mostly make sense for their specific purpose, and that manages to be an effective data storage and transmission mechanism while only interacting with physical materials rather than abstract digital bits.
And there are plenty of other bits of embodied cleverness from this time period. Double-action revolvers use the mechanical force applied to a trigger to cock the gun, release the hammer, and spin the chamber to make the next bullet available. Alarm clocks integrate the "keep track of time" function and the "play a sound at a particular time" function into the same device. Even humble pencil sharpeners manage to combine faster speed of sharpening, a sharper point, and far less waste than the alternative, which involves sharpening a pencil with a knife.
Such a device wouldn't be invented today, not because the collective investment in R&D is any lower, or because of regulatory barriers, or any of the vaguely political-economic issues like that, but because we have so many labor-saving shortcuts. Ever-smaller batteries and ever-more-efficient electric motors mean that you don't need a clever way to build something mechanical. You don't need your physical devices to embody some kind of logic (like if current_time >= alarm_time then: play_alarm_sound() ), we can generally convert the state of the device into a stream of bits, feed those to a very cheap commodity chip, and write the rules that chip uses to make its outputs. It also doesn't need some built-in way to share its status, the way a tea kettle uses steam to create sound. LEDs are cheap for simple status indicators, and LCDs are also useful for more complex interfaces.
So for modern use cases that require some kind of cleverness embedded in a physical device, it's usually more straightforward to build with off-the-shelf parts than to find some perfect confluence between the limitations of materials and the possibilities of mechanism construction. If we somehow forgot how to make can openers, but still had canned food, we'd probably invent something a lot less elegant than the two-handle, two-gear model. But it wouldn't take decades to figure out, either; we'd reinvent electric can openers promptly and probably never bother with the purely hand-operated variety.
One way to model this is that general-purpose technologies don't just allow new use cases. They take share, and they scale, and that means that they get good enough and cheap enough to be the default solution to a wide variety of problems, even problems that have already been solved in other ways. (Rails predate the steam engine by a few centuries; they're a good way to direct humans or animals hauling things in and out of mines.) Since these general-purpose technologies have so many use cases, it makes sense to invest more effort in incrementally improving them than in pursuing objectively larger improvements that will be used for orders of magnitude fewer products.
So yet another way to look at these cleverness substitutes is that they collectively embed much more cleverness than earlier contraptions. Far more time has been invested in making the GPU, LIDAR, and image signal processor chip on a modern iPhone better than was invested in making the original Kodak camera. And why wouldn't they be? All of these tools have broader applications, and they thus have a bigger end market. The maximum quality of those thought-hours is higher, too; there have to be historical inventions that were either delayed or didn't happen at all because whoever was designing them lost their train of thought while looking up the right entry in a table of logarithms or while cranking through a particularly laborious manual calculation. The people researching modern tools have access to information that simply wasn't available before, and have much easier access to the store of existing knowledge.
Modern increases in compute fit into this model in two contradictory ways. First, we've added intelligence as one more ever-cheaper generic input that can be added to a growing variety of products. The microwave today might be the last one that you interact with by entering a time instead of letting it just scan the food itself, figure out the optimal time, and ask you in its uncanny-valley voice whether you're still upset that it burnt the popcorn a few weeks ago. So, speakers and screens giving access to not-quite-deterministic outputs will continue to eat number pads and single-function buttons.
But in the other direction, it's more conceivable to convert the messy unstructured real-world constraints of use cases, materials, and labor costs into a straightforward optimization problem. Given enough GPU-hours, you'll get all the serendipity you need. (And, of course, those same GPUs also train the ad targeting algorithms ensuring that this niche product gets shown to exactly the audience that's excited to buy it.)
The general trend of technology and economic growth is that they unlock various ways to benefit from the intelligence of others at a low marginal cost. Every new unit manufactured amortizes the inventor's effort over more end customers. And these products embody some kind of cognition that reduces the amount of knowledge any one person needs. Thomas Sowell points out early in Knowledge and Decisions that, while people living in advanced economies know a lot in the aggregate, the number of things where knowledge is a matter of life and death is much higher among hunter-gatherers, and their knowledge is much more specific to immediate circumstances. But the returns on knowledge compound at a higher rate, and for longer, in those same advanced economies. This trend goes full circle when intelligence itself is one of those products of a complex economy devoting immense resources to research and development in order to find one more universal input.
Elsewhere
Sin Taxes
The Economist notes a paradox in the otherwise-appealing concept of a Pigouvian tax, i.e. a tax that's designed to discourage some undesirable activity. If it works to reduce demand that leads to fiscal constraints and unpopular new taxes. There's a very slow death spiral, where as the number of people who indulge in a vice goes down, they lose lobbying power even as their revenue contribution drops, and that process feeds on itself. So part of the political dynamic is that an addiction to popular taxes that will have to be replaced by unpopular ones later on is a sort of vice on the part of people who support vice taxes. The taxes still work fine, of course, as long as we choose the right vice and tax at a level that doesn't lead to too much smuggling. But when they work, they bring up new problems for the people who supported them.
Optionality
Perplexity AI has a tool for providing AI-generated answers to questions related to the US election. This is within the technical capabilities of many labs, but high-profile hallucinations are a bigger business risk to OpenAI (synonymous with AI businesses) or Gemini (the Google brand is too valuable to risk). It's hard to guarantee that answers are as accurate as the source material without just tuning the model so it's regurgitating the exact text it was trained on. But that legal risk is apparently one that Perplexity is willing to undertake—especially because offering a tool like this and linking back to media outlets is a way for Perplexity to send them some traffic, as a demonstration that it's better for publishers to have their data in Perplexity and get traffic and licensing money than to miss out.
The Dow
The Dow Jones Industrial Average is an old index, and like many attempts then and more recently to build a model that reflects the state of the economy, it was designed to balance between realism and data/compute limits. And since it was designed in the late 19th century, those compute limits were severe indeed, and the Dow is weighted by share price rather than market cap and only tracks thirty stocks. Despite this, it does almost exactly what the S&P 500 does: offer exposure to a diversified slice of American industry in a way that's easy to implement. Since it's an index that isn't automatically constructed, it's a good read on what parts of the investment zeitgeist seem permanent enough that they should be represented there. So Nvidia and Sherwin-Williams are replacing Intel and Dow. The Dow is a benchmark, not an attempt to pick outperformers, so including Nvidia is just an admission that it's a big enough company that it's part of what any investor with equity exposure is trying to outperform.
The GPU Financing Ecosystem
$11bn is not a large number in the world of asset classes, but a strikingly big number when restricted to just loans against GPUs ($, FT). The GPU-backed lending business is fun, because it's not strictly backed by GPUs: "Like traditional asset-backed lending, in the event of a default, the lender would own the GPUs as well as the contracts — known as power-purchase agreements — with the companies that lease them." PPAs get less valuable when GPUs improve, so both sides of this bet are long near-term demand for compute but short perceived long-term demand for compute. These loans default if demand for training and inference plummets, but they also drop if demand is so high that Nvidia can afford to develop more blockbuster GPUs while bigger cloud companies ship better tensor processing units. But that's always the basic bet credit investors make: they look best when things do reasonably well, because if things go badly you wouldn't have wanted exposure anywhere in the capital structure, but the better the outcome the better it was to have owned equity all along.
AI Fragmentation
Nvidia is considering an investment in Elon Musk's xAI. The Diff has argued for a while now that Nvidia's habit of strategic investments in major customers (but less so in their biggest customers!) is actually a sound plan: it means that there are more labs launching more models that incrementally leapfrog one another, instead of one biggest-and-best lab that has the resources to keep expanding its lead and thus the negotiating power to reduce Nvidia's margins. The usual dynamic with cyclical providers of capital equipment is that they have great economics periodically, but very little control over the structure of their industry or the cycle they're tied to. Nvidia is unique among cyclicals in that it's able to trade better short-term economics for more long-term control.
Diff JobsCompanies in the Diff network are actively looking for talent. See a sampling of current open roles below: - A Google Ventures-backed startup founded by SpaceX engineers that’s building data infrastructure and tooling for hardware companies is looking for a staff product manager with 5+ years experience, ideally with AI and data intensive products. (LA, Hybrid)
- An AI startup building tools to help automate compliance for companies in highly regulated industries is looking for legal engineers with financial regulatory experience (SEC, FINRA marketing review, Reg Z, UDAAP. JD required; top law firm experience preferred. (NYC)
- A hyper-growth startup that’s turning customers’ sales and marketing data into revenue is looking for a product engineer with a track record of building, shipping, and owning customer delivery at high velocity. (NYC)
- Ex-Ramp founder and team are hiring a high energy, junior full-stack engineer to help build the automation layer for the US healthcare payor-provider eco-system. (NYC)
- A well-funded startup that’s building the universal electronic cash system by taking stablecoins from edge cases to the mainstream is looking for a senior full-stack engineer. Experience with information dense front-ends is a strong plus. (NYC, London, Singapore)
Even if you don't see an exact match for your skills and interests right now, we're happy to talk early so we can let you know if a good opportunity comes up. If you’re at a company that's looking for talent, we should talk! Diff Jobs works with companies across fintech, hard tech, consumer software, enterprise software, and other areas—any company where finding unusually effective people is a top priority.
|