GPT's underlying training objective is to predict the next word from preceding text. Whether the resulting sequence is true or false is not part of that objective, which helps explain why fluent outputs can still contain fabricated information.

Next-word prediction is the core training task described
Fluency can emerge from a seemingly narrow objective
Truth is not inherently optimized by next-token accuracy
Human labels such as smart or dumb poorly describe the system

“You show at the beginning of a sentence, and it predicts the next word in the sentence, and that's all it is trained to do.”
Jon McAuliffe · 08:40

“whether you know the sequence of words that's being output, is leading to something that is true or false, is irrelevant.”
Jon McAuliffe · 08:58

#gpt#llms#hallucinations#ai

Myth Buster47:32

Why Quant Models Cannot Simply Be Copied Across Asset Classes

Teams may try to standardize a new market so it resembles a familiar one, but the effort often collides with market-specific realities. Corporate bonds, for example, may lack the continuous prices and order books that equity systems assume as basic inputs.

Standardization is a useful starting principle
Different markets expose different data and trading structures
Corporate-bond pricing is less observable than equity pricing
New asset classes frequently require fresh statistical problems to be solved

“there is no live price feeds that's showing you a bit offer quote in the way that there is inequity.”
Jon McAuliffe · 48:09

“That principle kind of quickly goes out the window when you when you start to make contact with the reality of how the new asset…”
Jon McAuliffe · 51:20

#asset-classes#corporate-bonds#quant#market-structure

Hot Take· 2

Hot Take52:57

A Bond ETF May Not Have One Obvious Underlying NAV

Bond ETF market prices are set in a liquid ETF market, but comparison with an underlying net asset value is not always straightforward. Authorized participants may negotiate the securities used in creation and redemption baskets, leaving no single, crisply defined underlying portfolio at every moment.

ETF market price reflects supply and demand
Bond ETF holdings can be less mechanically defined than assumed
Creation and redemption baskets may be negotiated
The meaning of an NAV comparison can therefore be ambiguous

“in many cases with bond ETF there's not even a crisply defined underlying portfolio.”
Jon McAuliffe · 53:14

“it's not even at all clear what you mean when you say that the nav is this or that relative to the price of the…”
Jon McAuliffe · 53:35

#bond-etfs#nav#liquidity#market-structure

Hot Take55:25

A “Six Sigma Event” Often Means You Estimated Sigma Wrong

Extreme market events recur more often than simplistic normal-distribution language suggests. McAuliffe's concise diagnosis is that an apparently impossible observation may reveal an underestimated volatility model rather than a supernatural tail event.

Rare-event labels depend on the assumed distribution
Observed extremes can falsify volatility estimates
Model error may be more plausible than repeated impossible events
Risk systems should account for heavier or changing tails

“if you think that you've experienced a six sigma event, the problem is that you have underestimated sigma.”
Jon McAuliffe · 55:34

#tail-risk#volatility#statistics#risk

Explainer· 3

Explainer06:08

Why More Powerful Models Demand More Data

McAuliffe explains that algorithmic sophistication and dataset size are inseparable. More expressive systems can represent richer relationships, but without enough observations they are increasingly likely to interpret one-time randomness as persistent signal.

Expressive models can capture more kinds of patterns
Greater flexibility increases sensitivity to noise
Large datasets are needed to separate recurring signal from chance
The same tension appears in recommendations and financial prediction

“The way that you make algorithms great is by making them more powerful, more expressive, able to describe lots of different kinds of patterns and…”
Jon McAuliffe · 06:14

“The more expressive. A tool like that is like a recommender system, the more prone it is to mistake one time noise for persistent signal”
Jon McAuliffe · 06:29

#machine-learning#data#signal#noise

Explainer11:59

How Click Probability Sets a Search-Ad Bid

Each search can trigger a rapid electronic auction among advertisers. A bidder estimates click probability, multiplies it by the expected value of a click, and sets a bid low enough to preserve profitability.

Search-ad placement is auction-driven
Advertisers bid through automated software
Expected value combines click probability with downstream profit
Bad click-through estimates can waste the advertising budget

“every single time you type in a search term on Google and hit return, a very fast auction takes place”
Jon McAuliffe · 12:14

“you want to know basically the probability that somebody is going to click on your ad, and then you would multiply that by how much…”
Jon McAuliffe · 13:08

#advertising#auctions#prediction#marketing

Explainer14:21

Using Evolution to Find Functional DNA

Researchers can compare aligned human and primate genomes to identify unusually conserved regions. If mutations in an important region reduce reproductive success, natural selection removes them, so long-term conservation becomes evidence that the sequence performs a function.

Human and primate genomes can be aligned position by position
Functional regions tend to resist accumulated mutation
Natural selection removes harmful variation
Conservation acts as a statistical signal of biological function

“you look for places where all of the genomes agree.”
Jon McAuliffe · 15:38

“look for conservation and use that as a big signal that something is functional.”
Jon McAuliffe · 16:10

#genetics#evolution#natural-selection#statistics

Takeaway· 6

Takeaway19:34

A Hedge Fund Can Add Value by Avoiding Familiar Risks

Voleon does not seek exposure to market and well-known econometric factors that institutional clients can already obtain. Its intended contribution is a stable return stream that remains uncorrelated with the client's existing sources of return.

Institutional investors already hold familiar market exposures
Replicating those factors adds limited diversification
A differentiated manager should contribute a new return stream
Low correlation is central to the value proposition

“What we're trying to provide is the most diversification possible.”
Jon McAuliffe · 20:01

“is also not correlated with any of the other return streams that they already that they already have.”
Jon McAuliffe · 20:14

#diversification#hedge-funds#portfolio#risk

Takeaway27:10

Build a Research Team by Waiting for Exceptional Colleagues

McAuliffe says Voleon devotes substantial energy and patience to recruiting premier machine-learning researchers and practitioners. Technical brilliance matters, but so do character, collegiality, and the opportunity for everyone to learn from one another.

Recruiting quality is treated as strategically important
The firm waits rather than lowering its hiring standard
Colleagues should be both brilliant and good collaborators
Team quality affects performance, learning, and satisfaction

“we exhibit a lot of patients. We we wait a really long time to be able to find the people who are kind of really…”
Jon McAuliffe · 27:23

“we really wanted to be able to assemble a group of people who were, you know, as brilliant as we could find, but also you…”
Jon McAuliffe · 28:07

#recruiting#teams#research#culture

Takeaway35:01

AI Adoption Slows Where Systematic Data Is Scarce

The likely boundary for machine-learning adoption is not a particular profession but the availability of repeatable, large-scale data. Fields become strong candidates for automation when they can systematically capture the same information experts use in their decisions.

Large datasets are a prerequisite for many machine-learning systems
Data must be gathered systematically, not merely exist
Expert decision inputs must be reproducible at scale
Poor data availability can delay adoption

“The kinds of fields where you would expect uptake to be slow are where it is hard to assemble large data sets of systematically gathered…”
Jon McAuliffe · 35:01

#ai-adoption#datasets#automation#prediction

Takeaway44:00

Treat Other People's Failures as Free Tuition

McAuliffe studies failures such as Long-Term Capital Management without assuming their participants were foolish. Learning how capable people were undone provides experience at lower cost, while preserving humility about unknown failure modes that remain ahead.

Failure histories contain reusable risk lessons
Sophisticated people can still make catastrophic mistakes
Avoided mistakes are not proof of invulnerability
Unknown blind spots require continuing humility

“good judgment comes from experience. Experience comes from bad judgment.”
Jon McAuliffe · 44:03

“The extent to which you can get good judgment from other people's experience, that is that that is like a free tuition.”
Jon McAuliffe · 44:10

#failure#judgment#risk#humility

Takeaway1:05:56

Choose Exceptional Colleagues Before Choosing the Perfect Field

For graduates with enough flexibility to optimize for long-term development, McAuliffe prioritizes the caliber of colleagues over the precise field or assignment. Exceptional coworkers improve learning and satisfaction, while an uninspiring peer environment can create lasting professional unhappiness.

Colleague quality can matter more than the exact specialty
Strong peers accelerate learning
Team caliber affects personal satisfaction
The advice assumes basic financial needs are already covered

“number one piece of advice is work with incredible people.”
Jon McAuliffe · 1:06:29

“much more important than what the particular field is the details of what you're working on, is the caliber of the people that you do…”
Jon McAuliffe · 1:06:34

#career#colleagues#learning#mentorship

Takeaway1:07:29

Predictive Accuracy Fails Without Risk Management

McAuliffe's hard-earned lesson is that prediction quality may be the central ingredient in a quantitative strategy, but it is not sufficient. High-quality risk management is mandatory if accurate forecasts are to produce durable investment success.

Prediction quality is important but insufficient
Risk management is a mandatory companion to forecasting
Overemphasis on model accuracy can neglect portfolio survival
Operational success depends on multiple supporting ingredients

“there are lots of other very necessary, mandatory ingredients, and I would put kind of risk management at the top of that list.”
Jon McAuliffe · 1:07:47

“if you don't have high quality risk management to go along with that predictive accuracy, you won't succeed.”
Jon McAuliffe · 1:08:10

#risk-management#prediction#quantitative-finance#investing

Frameworks in this episode

Adjacent-Signal Nowcasting

Locked-Drawer Holdout Validation

Prediction-First Black-Box Modeling

Systematic Investment Production Chain

Insights & moments

⚡Myth Buster· 2

GPT Predicts Language, Not Truth