ExpertsFMEric Vishria - A Decade of Lessons Investing in Software & Hardware - [Invest Like the Best, EP.486]
Invest Like the Best with Patrick O'Shaughnessy · August 11, 2026
5 Selected Lessons:
reasoning_pattern · 2 heuristic · 2 framework · 1
Bootstrap new-domain AI models with high-value curated data first, then post-train for breadth
Source
Invest Like the Best with Patrick O'Shaughnessy — Eric Vishria - A Decade of Lessons Investing in Software & Hardware - [Invest Like the Best, EP.486] (August 11, 2026) · at 39:20 ▸ · at 42:20 ▸ · at 44:29 ▸
When this applies
when training any AI system for a domain that lacks a ready-made large-scale dataset analogous to the internet text corpus
The takeaway
When no large general dataset exists for your domain, deliberately curate a smaller but high-value, tightly-matched dataset to bootstrap a strong base model, then rely on lightweight post-training to extend capability — don't wait for an internet-scale corpus to appear.
Why it works
LLMs succeeded by pretraining on a massive, implicitly curated-for-value corpus (internet text weighted toward higher-quality sources) and then using small amounts of post-training/RL to teach specific new behaviors; robotics companies can replicate this by first collecting a narrow but high-value, vertically-integrated dataset (e.g., purpose-built gloves matching the robot hand) to bootstrap a strong pretrained model, then adding small post-training examples to generalize to new tasks.
One level up
Capability acquisition order matters: build a strong, narrow base from your best available data before trying to generalize broadly.
Where it stops holding
This stops working when the target domain's tasks are too heterogeneous for any narrow high-value dataset to transfer, or when vertical integration (tailoring hardware/data collection together) is not feasible.
Detail
procedure:
- Recognize the domain lacks an internet-scale bootstrap dataset
- Vertically integrate hardware and data collection (e.g., custom gloves matching the robot's hand) to maximize data transferability and quality
- Use this curated high-value dataset for pretraining a base model
- Add small, targeted post-training/RL examples to generalize to new specific tasks
- Build rigorous evaluation baselines (e.g., measuring folded laundry quality) to validate emergent behavior
benchmarks: {"evolution observed":"from a 'janky cardboard glove' demo to a dozen robot arms reliably folding arbitrary laundry"}
rationale: The task itself (e.g., laundry) matters less than whether the pretrain-then-post-train flywheel is working, because a working flywheel lets capability multiply across many tasks.
In their words
I think one of the most interesting things that these AI robotics companies are doing are saying, okay, well, let's just go after the high-value data to start.
Oligopoly-not-winner-take-all market sizing
Source
Invest Like the Best with Patrick O'Shaughnessy — Eric Vishria - A Decade of Lessons Investing in Software & Hardware - [Invest Like the Best, EP.486] (August 11, 2026) · at 4:33 ▸ · at 9:02 ▸ · at 59:06 ▸
When this applies
when forecasting whether a fast-growing new technology market will be winner-take-all or will support multiple large players
The takeaway
Before assuming zero-sum, check whether the market's growth rate outpaces any single player's ability to scale and consume it — if so, many large winners can coexist.
Why it works
Markets perceived as 'this one company will eat everything' are usually underestimated in total size; as the market grows, multiple vendors can each capture enormous absolute value even while losing relative share, because the pie itself expands faster than any single player can consume it.
One level up
Correctly sizing how big a market can become is usually a bigger investing edge than correctly picking the single winner within it.
Where it stops holding
This stops working when the market is small relative to the first mover's scaling capacity, or when strong network effects/lock-in genuinely allow one player to consume the whole pie.
Detail
approach: Trace how a dominant-seeming platform's market evolved over a decade; check if 'threat to everything' narratives proved wrong because the market was far bigger than anyone modeled.
signals:
- loud 'X will eat everything' narrative
- huge expected market growth
- infrastructure buildout constraints (power, chips) limiting any single scaler's pace
rationale: In 2014 the AWS narrative wrongly predicted total dominance; by 2026 AWS/Azure/GCP/Cloudflare are each $100B+ businesses alongside thriving SaaS layers built atop AWS.
In their words
a bunch of zero-sum thinking and not realizing like how big. What if it all works? What if it all works? It all works.
Check all assumptions at the door when advising founders in an unstable substrate
Source
Invest Like the Best with Patrick O'Shaughnessy — Eric Vishria - A Decade of Lessons Investing in Software & Hardware - [Invest Like the Best, EP.486] (August 11, 2026) · at 22:41 ▸ · at 24:54 ▸
When this applies
when applying experienced judgment to a domain whose underlying technology/market substrate is changing faster than the experience was formed in
The takeaway
Before applying any past 'best practice,' explicitly re-derive it from first principles against the current substrate, rather than assuming experience transfers.
Why it works
Experienced leaders carry deeply encoded heuristics tuned to a prior, more stable substrate; when the substrate (technology capability, demand curve, market size) changes qualitatively, those heuristics misfire even though the leader's judgment and rigor are otherwise excellent.
One level up
The faster a domain's underlying substrate moves, the more aggressively past expertise must be re-validated rather than trusted by default.
Where it stops holding
This stops working when the substrate truly is stable and historical heuristics still hold — re-deriving everything from scratch would then waste time.
Detail
approach: For every inherited assumption (quota models, territory rollout, leadership profile), ask: does this still apply given an unstable technology substrate and a different demand curve?
signals:
- experienced leader from prior era struggling to hit targets
- demand being pulled, not pushed, by 'magic' product
- reps massively outperforming historical quota norms
rationale: Great leaders from the SaaS era repeatedly failed inside AI scaleups despite excellent pedigree, revealing an 'impedance mismatch' between old heuristics and the new substrate.
In their words
I've really started to question every assumption and every lesson that I learned before, which of it translates and which of it doesn't.
Build for the jagged edge, not the smooth human arc
Source
Invest Like the Best with Patrick O'Shaughnessy — Eric Vishria - A Decade of Lessons Investing in Software & Hardware - [Invest Like the Best, EP.486] (August 11, 2026) · at 11:37 ▸ · at 13:57 ▸
When this applies
when building on a rapidly improving underlying technology whose capability profile is uneven (great at some things, bad at others) rather than smoothly competent
The takeaway
When your core technology improves in discontinuous, uneven jumps, treat your product as disposable scaffolding ('sandcastles') to be rebuilt each cycle, not durable architecture ('castles').
Why it works
Models have a 'jagged edge' of capability unlike smooth human skill curves; product value comes from precisely mapping what the model does well/poorly onto the customer's actual problem, and that mapping must be rebuilt as the edge moves.
One level up
The speed of your underlying technology's improvement should dictate whether you build for permanence or for rapid obsolescence and rebuild.
Where it stops holding
This stops working when the underlying technology's capability curve stabilizes and stops shifting, making durable architecture valuable again.
Detail
rule: Understand the model's jagged capability edge and the customer problem, and continuously rebuild the bridge between them rather than building permanent architecture.
threshold: Expect meaningful new model capabilities roughly every four weeks, requiring product re-architecture at that cadence.
rationale: Companies that try to build 'perfect foundations' like artisans fail because the ground (model capability) keeps shifting under them.
In their words
We used to be building castles. Now we're building sandcastles that washed away.
Partner, don't just invest: the chemistry/mutual-selection filter
Source
Invest Like the Best with Patrick O'Shaughnessy — Eric Vishria - A Decade of Lessons Investing in Software & Hardware - [Invest Like the Best, EP.486] (August 11, 2026) · at 44:29 ▸ · at 49:51 ▸
When this applies
when deciding whether to enter any long-term, high-involvement relationship (investment, partnership, hire) that is financially sound but lacks deep interpersonal fit
The takeaway
Distinguish between opportunities that are merely investable (will probably make money) and opportunities worth the deep commitment of true partnership (mutual learning, chemistry, a desire to be reachable at any hour) — only pursue the deep commitment when both are present.
Why it works
Deep, multi-year collaborative value (mutual learning, feeding each other's thinking loops) only emerges from genuine mutual chemistry and mutual selection; a merely 'good enough' financial opportunity without that fit yields a transactional, lower-value relationship even if it 'works' financially.
One level up
The highest-value long-term relationships require selecting for fit, not just for expected value, because fit determines how much compounding collaborative value is created over time.
Where it stops holding
This stops working when the role genuinely only requires transactional capital, not an ongoing deep working relationship, making chemistry irrelevant.
Detail
rule: Pass on 'investment grade' opportunities lacking true chemistry/fit, and reserve commitment for relationships with mutual learning and mutual desire to partner.
threshold: 'Would I pick up the phone if this person called at 9pm on a Saturday?' as a concrete chemistry test.
rationale: An investor would take a probably-profitable deal regardless of fit; a 'partner' requires chemistry sufficient for sustained mutual teaching/compounding.
In their words
The other question, if this person calls me at 9 p.m. on a Saturday night, will I pick up the phone?