The End Is Nigh. Buy Here?

Tesla

Across the shows: 11 speakers in other episodes, 12 bullish and 3 bearish calls, past 30 days

• Tobias Carlisle · Acquirers Funds / The Acquirers PodcastBearish

  Tesla's self-driving is still about ten years away despite Musk's earlier timelines.

  Carlisle notes Musk has said for five to ten years that autonomy was near. He concludes that, from the outside, Tesla is still roughly ten years away from self-driving. He contrasts this with Waymo, which he asks about and treats as further along.

  “Musk saying it's five years ago, ten years ago. Clearly, it's still ten ish years away for Tesla.”at 42:19

Waymo

Across the shows: 3 speakers in other episodes, 2 bullish and 1 bearish calls, past 30 days

• Jake Taylor · Farnam Street Investments / The Acquirers PodcastBullish

  Waymo is doing well, with strong accidents-per-mile results in a tightly mapped city.

  Taylor has ridden in a Waymo. He credits its focus on a single fully mapped city with reducing edge cases. He says statistically it seems to be doing well on accidents per mile driven. He hedges that he is unsure whether it is better than a human driver yet.

  “it seems like statistically, they're doing well as far as, like, accidents per mile driven, that kind of stuff.”at 42:55

AI: overall outlook

• Jake Taylor · Farnam Street Investments / The Acquirers Podcast

  AI self-improvement will be limited to domains like coding and math that give fast feedback.

  Taylor responds to the idea that OpenAI and Anthropic are using machines to build better machines, which could mean a singularity. He argues that self-improvement needs a feedback loop so the system can tell when it is getting better or worse. Coding and some obscure math problems provide that loop, so the system will keep pushing there by brute force. Much of the world does not give that feedback easily, and he is not sure you can get purchase on problem sets outside those domains. He treats this as a constraint on how far self-improvement can go.

  “you kinda need the feedback loop so that the machine knows when it's getting better or getting worse ... there's so much of the world that doesn't really give you the feedback loop very easily.”at 2:26

• Tobias Carlisle · Acquirers Funds / The Acquirers Podcast

  AI is already solving math problems and decoding the genome, which will be a big boost to medicine and aging.

  Carlisle cites present-day evidence that AI is solving math problems and decoding the human genome. He argues this will be incredible for medicine and, perhaps, for aging. He ties the math breakthroughs to models from Anthropic or OpenAI, and notes mathematicians thought the solutions were 'alien' connections a human would not have made. He hedges on the aging benefit with 'maybe'.

  “We're already seeing, like, it's solving math problems. It's decoding the human genome. So that's gonna be incredible for medicine and for aging maybe.”at 36:12

US stock market outlook

• Tobias Carlisle · Acquirers Funds / The Acquirers Podcast

  Valuation is not a useful timing tool, though expensive markets imply lower future returns and more volatility.

  Carlisle says his backtesting across CAPE, Tobin's Q and Buffett's market-cap-to-GNP ratio all point to the same answer: the market is expensive relative to its trend. He argues that an expensive market implies lower returns and some volatility ahead. He is equally clear that valuation does not work for timing. He also notes CAPE has calculation problems when earnings have run far ahead, but that the rough answer holds. His reading is that the market is expensive, but valuation will not tell you when to act.

  “Valuation is not a useful timing tool at all. Valuation doesn't work as a timing tool... Is not a timing tool.”at 11:33

• Tobias Carlisle · Acquirers Funds / The Acquirers Podcast

  Another market crash is highly likely, regardless of whether valuations or AI are the cause.

  Carlisle cites history showing a crash roughly every seven years, which makes another crash highly likely. He says the crash does not need to be explained by overvaluation or AI, since markets simply tend to crash. The crash risk is therefore a structural feature of markets rather than a call about current AI valuations.

  “There's a crash about once every seven years. We're highly likely to get another crash. It's got nothing to do with the market being overvalued or AI.”at 53:06

Enterprise AI adoption & AI software

• Tobias Carlisle · Acquirers Funds / The Acquirers Podcast

  AI-driven backtesting data-mines results that fail out of sample, so findings rarely persist.

  Carlisle gave Claude his backtest data and software. It ran 916,000 backtests across 1,217 parameter sets in one afternoon, where he would normally run dozens. He describes this as a data mine. He follows his own rule of optimising on the first half of the data and testing on the second, and finds it very hard to find anything that persists from one set to the next. What optimises in the first half usually fails in the second. He argues that the AI's speed of search makes overfitting easier, not the discovery of real edges.

  “And it's very hard to find anything that persists from one set to the next.”at 4:05

US growth & recession risk

• Tobias Carlisle · Acquirers Funds / The Acquirers Podcast

  The 10-3 yield curve inversion recession signal has stopped working, and un-inversion has not helped.

  Carlisle says the 10-3 inversion had a strong track record, with a near-perfect hit rate through 2022 to 2024. It then failed to call the downturn that followed. He says the curve un-inverted about a year before the conversation, with a sideways market since then. He concludes that un-inversion has not behaved like a recession signal, and that stretching the definition to keep it alive is getting hard.

  “But the uninversion doesn't seem to work.”at 10:55

US Treasuries & yields

• Tobias Carlisle · Acquirers Funds / The Acquirers Podcast

  The 60/40 portfolio's bond hedge has worked only about 20 years and failed in 2022.

  Carlisle explains the 60/40 logic. Bonds rally when equities fall, so rebalancing buys more stock at the bottom. That only works when stocks and bonds are negatively correlated. He says that held for a short window, about 20 years, while most of the time they have been positively correlated. In 2022 bonds sold off alongside stocks, so the hedge failed and the portfolio did not recover as expected.

  “6040 has only worked for twenty years. There's a short period of time where bonds and equities were uncorrelated. Most of the time, they've been pretty correlated.”at 15:30

Market sentiment, positioning & volatility

• Tobias Carlisle · Acquirers Funds / The Acquirers Podcast

  Moving-average trend rules whipsaw, have no good parameter, and are an expensive diversifier at best.

  Carlisle tested moving averages, including the 50-day, 200-day and crossovers. He found no choice of average that solves the problem. A rule can ride a market to its peak and then exit at the very top. Markets often fall back to their moving average before rallying, which causes whipsaws and overtrading that get worse at higher checking frequency. He calls moving averages trend strategies that work sometimes and cost money. They might earn a place as one uncorrelated stream, but they do not solve the problem. He also says CTA track records are flattered by survivorship bias, since failed funds drop out of the data.

  “And so moving averages are trend strategies. ... It works sometimes and doesn't work other times, and it's expensive.”at 20:11

Healthcare system & insurers

• Jake Taylor · Farnam Street Investments / The Acquirers Podcast

  Technology like personalized medicine, not government reform, is the realistic route to cutting US health costs.

  Taylor says the US spends about 20% of GDP on health care for outcomes that are reasonable but not good per dollar, and other countries get similar results for about half the spend. He doubts government reform because of entrenched interests and powerful lobbies. He sees new technology, such as personalized medicine that is better and cheaper to administer and becomes the new standard of care, as the only real hope.

  “But from the technology side, if we can just, like, chip away at it and be like, well, this is the new standard of care because it's so much better and it's cheaper to administer and personalized medicine can probably help a lot there”at 38:08

Frontier AI labs & model race

• Tobias Carlisle · Acquirers Funds / The Acquirers Podcast

  OpenAI and Anthropic are already building AI that builds smarter AI, so we're in the singularity now.

  Carlisle argues the singularity is defined by machines that make smarter machines. He says OpenAI and Anthropic are doing this now or will do so this year, so 'we're living in the singularity right now.' He frames this as a present-tense fact, with the timing set by 'this year'.

  “OpenAI and Anthropic are making the machines that make the machines smarter. They're either doing that now or that's... They're gonna be doing that this year.”at 41:18

Robotics & autonomous vehicles

• Tobias Carlisle · Acquirers Funds / The Acquirers Podcast

  Humanoid robots are getting close, though they are not in homes yet and cost is unclear.

  Carlisle points to robot Olympics and a robot fighting league as evidence humanoid robots are advancing quickly. He notes they are not yet in the home and that the cost of one is unknown, but he judges them 'close'. This comes from observed demonstrations rather than hard data.

  “It's getting close... they're not in the home yet, and who knows how much one of these things actually cost, but they're close.”at 43:56

AI infrastructure & data centers

• Tobias Carlisle · Acquirers Funds / The Acquirers Podcast

  AI bottlenecks like power, compute and data center space are solvable and just take time.

  Carlisle lists the possible constraints on AI as power, computing power and data center space. He argues each is a solvable problem that is not especially hard to fix, though it takes some time. He treats these bottlenecks as temporary rather than binding.

  “you gotta think that all of those problems are solvable problems that they're not that difficult to solve. They just take a little bit of time.”at 46:14

Software & SaaS

• Tobias Carlisle · Acquirers Funds / The Acquirers Podcast

  AI can rebuild existing software from its binary, so consumers will stop paying for it.

  Carlisle describes AI decompiling a binary into source code and rebuilding a working version of Microsoft Word from what it can see. If AI can rebuild any software for free, he asks what people will pay for. He concludes 'all software is cracked' and that the value goes to consumers as surplus rather than to software vendors.

  “Or you can just put Word, Microsoft Word, and it'll give you a functioning version of Word built backwards from what it can see. But I've already got Word.”at 47:51

AI & jobs

• Tobias Carlisle · Acquirers Funds / The Acquirers Podcast

  Robots plus superintelligent AI could leave humans with little work to do.

  Carlisle reasons that a robot could do all physical labor while a superintelligent machine does the thinking work, with no mistakes and around-the-clock availability. He asks what humans would then do, and his only answer is 'make more humans'. He treats the outcome as a speculative question rather than a confident forecast.

  “If you've got a robot that can do all the physical labor and you've got this ultra intelligent machine that it's smarter than you are... What's the human day? What are we doing?”at 48:23

Nuclear energy

• Tobias Carlisle · Acquirers Funds / The Acquirers Podcast

  Nuclear power never became too cheap to meter because nuclear is too expensive to build.

  Carlisle recalls a 1950s prediction that nuclear would make energy too cheap to meter, and says it has not happened. He offers his own explanation that nuclear is too expensive to build, though he says he has not checked this. The view is a tentative explanation rather than a researched conclusion.

  “because we haven't got nuclear. It's too expensive to build nuclear.”at 49:31