ExpertsFMWhat racing reveals about working with AI - Episode 22
OpenAI Podcast · July 16, 2026
5 Selected Lessons:
heuristic · 3 framework · 1 reasoning_pattern · 1
Education-first AI adoption to unlock better prompting
Source
OpenAI Podcast — What racing reveals about working with AI - Episode 22 (July 16, 2026) · at 11:44 ▸
When this applies
when rolling out AI tools to domain experts who don't yet understand the technology
The takeaway
Before deploying an AI tool to experts, teach them the mechanism (it's input-driven, not magical) so they take ownership of crafting better inputs.
Why it works
Demystifying how models work (showing they are machines, not magic) lets experts reframe interactions as engineering problems with controllable inputs, improving the quality of inputs (prompts) and therefore outputs.
One level up
Adoption follows understanding; demystification precedes effective use.
Where it stops holding
This stops working when users don't have time or motivation to learn the mechanism, or when the tool's behavior is genuinely opaque even to its builders.
Detail
procedure:
- Show that models are machines under the hood, not magic
- Teach the input-output relationship: right input yields right output
- Put onus on the user to figure out the right input (prompt engineering)
- Do prep work ahead of race day so engineers feel prepared for strategic calls
benchmarks: {}
rationale: Once engineers understand the model's mechanism, they can conceptualize how to work with it better, improving outcomes without needing the AI to have been 'smarter' per se.
In their words
One of the most impactful things that we've delivered to the team is, frankly, is education. It's education on a little bit about how our models work. I think it helps to pull back the veil and show that these are just machines under the hood and it's not magic and these guys are engineers and they understand that.
Human-centric activities retain value even as AI/autonomy improves
Source
OpenAI Podcast — What racing reveals about working with AI - Episode 22 (July 16, 2026) · at 31:36 ▸
When this applies
when speculating whether automation will replace a human-performed activity whose appeal rests on watching skilled humans perform
The takeaway
Before assuming automation will displace a human activity, check whether the activity's value derives from the outcome or from the fact that a human produced it.
Why it works
The appeal of many human activities (sports, racing, space exploration) comes specifically from watching a human achieve something difficult, so automating the task removes the source of the appeal even if the automated version achieves the same objective outcome.
One level up
Automation substitutes for tasks, not necessarily for the meaning humans derive from performing or watching those tasks.
Where it stops holding
This stops working when the audience/market genuinely only cares about the outcome (e.g., fastest lap time) and not about who/what produced it.
Detail
approach: Analogize to other domains where automation exists alongside human performance (robots vs basketball players, probes vs astronauts) to argue the human version retains its distinct appeal.
signals:
- Activity is valued partly for human skill/effort on display
- Automated version already exists but hasn't displaced human version
- Audience explicitly says they'd be 'upset' if replaced
rationale: Existing autonomous racing experiments (e.g., Abu Dhabi open-wheel) are 'still hitting walls', suggesting the human version's dominance is not merely due to technological immaturity but to differing appeal.
In their words
I actually think the application of AI we're seeing in motorsports is extremely human centric because these are some of the most goal oriented people I've ever met and I've ever worked with.
Convert legacy spreadsheet formats to AI-native formats for better model performance
Source
OpenAI Podcast — What racing reveals about working with AI - Episode 22 (July 16, 2026) · at 21:43 ▸
When this applies
when an organization's critical data lives in a legacy format (spreadsheets, PDFs, ad hoc files) that AI models handle poorly
The takeaway
Invest in converting legacy data formats to model-friendly formats early, because the conversion cost is repaid by the volume of better downstream outputs.
Why it works
Models work with text/structured formats much more reliably than with spreadsheets' visual formatting, formulas, and highlighting, so converting data upstream yields disproportionately better downstream AI results.
One level up
Infrastructure investment that seems like 'overhead' (format conversion) is often the highest-leverage lever for AI performance.
Where it stops holding
This stops working when the cost/complexity of conversion exceeds the value of the downstream analysis, or when model capabilities improve fast enough that near-term conversion effort becomes obsolete.
Detail
rule: Show teams that investing to convert spreadsheet data into AI-native format is worth it because of the downstream results it enables; re-test every three months to see if models can handle data 'upstream' of where they currently need it converted.
threshold: Re-evaluate model capability against raw data format roughly every three months as state of the art evolves.
rationale: Models evolve, so the gap between raw data format and model-usable format shrinks over time; teams should track that evolving gap rather than assuming it's fixed.
In their words
I think it's really important that we test every three months, hey, what is state of the art? Can our models work with something that is a little bit upstream of this process so you don't have to do so much legwork to get it into a format that AI can understand.
Data wars: data advantage as the current competitive frontier
Source
OpenAI Podcast — What racing reveals about working with AI - Episode 22 (July 16, 2026) · at 21:43 ▸
When this applies
in any industry where the main competitive battleground has shifted to who can capture and act on data fastest
The takeaway
When an industry's core differentiators (hardware, physical performance) converge across competitors, the remaining edge shifts to information capture and processing.
Why it works
As telemetry, notes, and strategic thoughts proliferate, the ability to capture, organize, and act on every piece of this data becomes the differentiator, since raw physical/engineering capability gaps are narrowing.
One level up
Competitive frontiers migrate from physical/structural advantages to informational ones as physical gaps close.
Where it stops holding
This stops working when physical/engineering differences between competitors remain large enough that data advantages can't close the gap.
Detail
rule: Every piece of telemetry, handwritten note, or strategic thought must now be captured and leveraged, because teams are in what Chase calls 'the data wars.'
threshold: Applies now, as he describes a present ongoing escalation ('I believe we are in what's called the data wars right now').
rationale: Data streams are only going to get tougher to manage as the sport evolves, making capture and simplification the critical differentiator.
In their words
I believe we are in what's called the data wars right now in racing, maybe on the IndyCar side and the NASCAR side.
Lead with shared enthusiasm to win first customers
Source
OpenAI Podcast — What racing reveals about working with AI - Episode 22 (July 16, 2026) · at 26:39 ▸
When this applies
when selling an unproven product into a market where trust and buy-in must be built from scratch
The takeaway
Target early customers who already share interest or curiosity in your technology, since instilling interest from scratch is the harder, slower work.
Why it works
A prospect already curious about the underlying technology requires less persuasion effort, because their pre-existing interest substitutes for the proof/credibility a cold prospect would demand.
One level up
Pre-existing belief and curiosity in a prospect is a scarce resource worth identifying and seeking out, not manufacturing from zero.
Where it stops holding
This stops working when the target customer base has no existing curiosity or is actively hostile to the technology, requiring a different go-to-market approach.
Detail
rule: Find a first customer who is already interested in the technology themselves; that pre-existing interest does most of the persuasive work.
threshold: A single genuine conversation before a formal engagement was sufficient in this case.
rationale: Interest 'is something that you have to create and instill and people want to see' — finding it pre-existing short-circuits that work.
In their words
He was just interested in artificial intelligence himself. So that was most of the work, would say. It's just the interest was already there because that's something that you have to create and instill and people want to see.