ILTB — Gabe Stengel / Rogo: investing superintelligence, harness, last mile

Opening

Buy-side AI is past demo theater: o1 Pro made reliable financial search usable; Opus 4.5-class models can do junior banker/analyst work with the right context — and the scarce work for the next 2–5 years is reinventing the firm (harness, compliance, last-mile UX), not waiting for a smarter base model. Patrick O'Shaughnessy interviews Gabe Stengel (~64m), co-founder/CEO of Rogo (AI platform for finance). Ground covered: Rogo product eras tied to frontier models; deal-side vs public-equities readiness; 10k-agent PM vision; data/model/harness stack; MNPI/auditability as wedge; vertical $5B+ revenue thesis vs frontier-lab incentives; durable human skills (proprietary inputs, judgment). Whisper asr from Megaphone audio — no YouTube yet at run time; Colossus episode page as source. Figures marked (asr).

Key takes

The 2–5 year problem is firm reinvention, not model IQ. Stengel: Jane Street took ~15 years to build a dominant market-making franchise; today's best investors will spend the next two to five years learning how to "fly AI into the investment life cycle." Frame: what would Goldman / Millennium / Citadel do if a "country full of geniuses" showed up in the data center — the lag is workflow integration. [asr]

Rogo capability eras track model eras: o1 Pro = reliable search/metric tool; Opus 4.5 (late 2025 / early 2026) = junior investment professional / junior banker work given instructions + context. Early GPT-3 era demos were "magical" and "nothing worked." First-mover disadvantage: ship before models are ready and users brand you garbage; if you built for the end-state (compliance, regulatory workflow, last-mile hookup), arrival of capable models flips the product to "transforming the way I work" / "saving hundreds of hours a month" (customer feedback as stated). [asr]

Current users skew transaction/deal workflow — sell-side and buy-side preparing materials to execute deals (data rooms, company models, customer PowerPoints, DDQ answers on concentration; opposing agents tearing through data mapped to firm investment philosophy). Public-equities "soup to nuts" diligence / IC memo / objection-handling is the next last-mile "chef" problem — deal makers further along than a junior PE HF analyst end-state. [asr]

Vision stretch: every PM with ~10,000 agents "fraternizing," reading notes, debating for 24 hours, returning one idea. Nearer claim: models are already "smarter than anyone I know" — value is plumbing them to thesis, working style, and firm context. [asr]

Harness > model for applied finance; architecture moved from ~60-call Rube Goldberg toward simpler high-quality tools as models got smarter. Spend goes into: compliance/regulatory so Delaware discovery of AI research doesn't intermingle MNPI; evals routing by performance / token cost / latency; integrations into behind-the-scenes firm systems (not only chat/email). Claude usage expansion credited more to harness/presentation of long-running capability than raw model delta. [asr]

Vertical wedge: capital-markets workflows are messy enough to support $5–10B businesses going deep; frontier labs won't pick that up on a $100B→$1T path ("stopping to pick up a penny"). MNPI ingestion, audit flags into bank/IM systems, and compliant live data rooms (not static Dropbox) are examples of painful last-mile that creates defensibility. [asr]

Durable human edge if models keep eating junior production: proprietary field inputs — time in the field, expert networks, relationship graphs that feed the model data no one else has. Move 37 analogy: if AI finds non-consensus public-equity moves, calculation matters less than unique inputs + judgment. Jury out on replacing that judgment. [asr]

Internal dogfood: every Rogo conversation recorded into a company brain ("Shrek" dashboard); monthly AI-tool usage stack-rank by division as leadership forcing function. Board pressure (Pat Grady / Sequoia cited) toward aggressive hiring/commercial/product goals; series A path via David Tisch intros when no blue-chips yet on the cap table. [asr]

Key math

Jane Street ~15 years to dominant MM franchise (asr — analogy) — reinvention timescale foil. [asr]

Investor AI integration window ~2–5 years (asr) — Stengel's planning horizon. [asr]

o1 Pro then Opus 4.5 as capability steps (asr) — product era markers. [asr]

Customer feedback: "hundreds of hours a month" saved (asr — unverified) — usage claim. [asr]

PM vision: ~10,000 agents / 24h debate → one idea (asr) — stretch product image. [asr]

Early Rogo: ~60 model calls in a Rube Goldberg (asr) — architecture starting point. [asr]

Vertical revenue ambition: deep FS workflows can support ~$5B+ (even $5–10B) businesses (asr) — company thesis, not run-rate. [asr]

Frontier-lab path framed $100B→$1T (asr — Anthropic foil) — why labs skip vertical pennies. [asr]

Quotes

"The world's best investors today are going to spend the next two to five years figuring out how to integrate AI into what they do." — Gabe Stengel [asr]

"Figuring out how to fly AI into the investment life cycle is the biggest challenge over the next five years for every great investor." — Gabe Stengel [asr]

"With Opus 4.5… the models just became capable of basically anything a junior investment professional or junior banker was doing as long as you gave it the right instructions and context." — Gabe Stengel [asr]

"Imagine if every PM at a hedge fund had 10,000 agents that were just kind of fraternizing… and then at the end of 24 hours of debate just gave you one idea." — Gabe Stengel [asr]

"The way that you harness these models is so so important." — Gabe Stengel [asr]

"The core skill set is folks who can go out and gather data and inputs into their model that no one else will have." — Gabe Stengel [asr]

Variant perception

Priced in — Applied AI for finance is hot; junior analyst tasks automate first; compliance/MNPI is hard; harness matters as much as weights; Sequoia-backed vertical AI narrative.

What's new — Concrete o1 Pro → Opus 4.5 capability map onto banker/analyst work; explicit deal-workflow penetration vs public-equities last-mile lag; $5–10B vertical FS systems-of-record ambition vs frontier-lab incentive foil; Delaware-discoverability as product constraint; internal usage stack-rank / always-on company brain as operating model.

Bear case — Customer hour-saved claims are vendor NPS; Opus-class "anything a junior does" overstates production reliability on live books; public-equities last mile may stay stuck (IC taste, liability); $5B revenue path is aspiration; model jumps can still wash product work ("chewing glass").

Discount — CEO of Rogo on ILTB — maximum book-talking and fundraising narrative. Sponsors (Ramp/WorkOS/Vanta/Ridgeline) frame adjacent IM software. Whisper asr from podcast audio; no YouTube captions to cross-check proper nouns (Opus versions, investor names).

Positioning

Enterprise agent stall — WEAKENS (soft) for finance workflows, NEUTRAL for F500 production-$. Deal-side agents already in data rooms / DDQ / philosophy-mapping per vendor; no Fortune 500 named with headcount or dollar production deployment — stall thesis unbroken on the formal breaks_if, but finance vertical looks further than generic enterprise pilots.

AI capex durability — NEUTRAL / soft STRENGTHENS. Token routing, eval cost, and "country of geniuses in the data center" assume continued model/serving scale — does not add hyperscaler capex evidence.

Inference margin inversion — NEUTRAL. Mentions routing by token cost/latency but no lab GM or price-cut arithmetic.

HBM supply binds — NEUTRAL. No semiconductor supply content.