Agustin Lebron — comprehensive briefing

0 EquiLibre Technologies — who you’re interviewing with

Company site: equilibre.ai. Official framing: Frontier AI Trading Research Lab — not a classic prop shop pitch deck. They build reinforcement-learning trading agents that “learn, adapt, and compete,” and they claim their models trade billions of dollars daily on hard markets.

Origin story they want you to know

1. DeepStack (2017) — first AI to beat professional humans in heads-up no-limit Texas Hold’em (imperfect information; search + value nets; continuous re-solving). Published in Science.
2. Founders also linked to Player of Games — strong play across perfect- and imperfect-information games.
3. Time at IBM Research then DeepMind Edmonton (visiting/project work with Alberta poker / RL community).
4. ~2022: leave DeepMind, found EquiLibre in Prague — “rather than playing poker, algorithms play algorithmic trading.”
5. Thesis: poker and markets share structure — uncertainty, hidden information, adversarial other agents, sequential decisions, simple scalar score (money / chip EV).

Founders (from equilibre.ai)

Site stress: mix of big-tech AI labs, top-tier trading, academia — “builders and dreamers.” No sales/marketing theater — “success is quality of models and algorithms.”

What they say they do (product / tech)

• Self-play emphasized in press as carry-over from DeepStack-era methods.
• Partnership: quant firm Tower Research Capital — algos trading large notional (press: billions daily volume on major equity indices).
• Performance claims (press, treat as marketing until you verify): “zero negative months since inception” after live trading (crypto 2025 → stocks). Do not parrot as gospel; you can say “public materials claim strong live track record with Tower.”
• Self-description: “a lab first, not a finance firm” (TechCrunch interview with Schmid) — research identity matters culturally.

Funding & scale (public, mid-2026)

• Founded late 2021 / public launch 2022 (sources vary slightly on wording).
• Pre-seed / seed path via European VCs (e.g. Credo, Blossom seed reported).
• Series A ~Jun/Jul 2026 at roughly €438M / ~$500M valuation (Creandum lead reported as large check; treat numbers as press).
• Use: scale AI trading agents / research lab growth.

Advisors (signal density)

• Rich Sutton — RL pioneer, Turing Award; Alberta / ex-DeepMind
• Michael Bowling — Alberta; DeepMind Edmonton site/lead history; poker research group / Science papers
• Csaba Szepesvári — RL theory, DeepMind / Alberta / Amii
• Michal Pěchouček — CTU Prague; ex-Avast CTO
• Murray Campbell — IBM; Deep Blue co-author

This board says: serious RL + game theory credibility, not a random fintech wrapper.

Prague / culture notes (from their site)

• HQ Prague — talent from regional unis + CEE tech migration; “deep work” lifestyle pitch.
• Hiring contact: hiring@equilibre.ai · general: schmid@equilibre.ai
• Open positions linked from site /hiring

How this maps to your story (use in interview)

Sources: equilibre.ai · TechCrunch / Recursive / CNBC coverage of founding & Series A · DeepStack (Science 2017) public literature · Tower Research partnership press. Verify live details before relying on valuation/volume numbers in conversation.

1 Who Lebron is

“I decided to look for jobs that were half engineering and half poker… that pretty much describes trading.” — Predicting Alpha podcast · career origin story

2 Vibe & how he thinks

3 Edge — his core concept

Edge is something that either you know or you can do that the marginal participant in that market either doesn’t or can’t. — Multiple interviews; formalized in the book

Forms of edge he lists

4 The 11 Laws of Trading

From The Laws of Trading: A Trader’s Guide to Better Decision-Making for Everyone. He taught new traders with these as mental shorthand; claims they’re timeless (Shakespeare knew versions of them). Aaron Brown’s foreword frames them as laws of the jungle — dense competition for scarce resources.

Note: chapter titles vary slightly across summaries; the substance above is consistent across his interviews + book reviews. Motivation / Adverse Selection / Risk / Liquidity / Edge are the ones he drills hardest on podcasts.

5 Jane Street & quant firms

“If I can buy baloney, bread, and cheese for $3, make a sandwich, sell it for $4 — that’s a business. That is what ETF market making is.” — Outlier podcast · explaining prop market making

Why quant profits rose (hypotheses he floats)

• World got richer → more capital / volume.
• Disintermediation of banks — pensions go straight to ETF MMs; risk-takers capture what banks used to take as middlemen.
• “Gamblization of everything” — retail boom, sports betting, prediction markets → more uninformed flow for MMs.

Industry structure

• Harder for small shops: tech fixed costs rising (esp. options MM consolidation).
• Automation lowers marginal cost of adding product → large firms reach further down the book (e.g. small-cap once “not worth it”).
• Still some room at the edges, but the waterline has risen.

6 Market making & adverse selection

Adverse selection outside markets (he loves this)

• Hiring: applicants who apply are selected against; best people are retained; final decision is the candidate’s (winner’s curse on employers).
• Crypto deals / investments: roleplay on Derivative podcast — always ask who knows what you don’t.

7 Strategy building (practical, from Outlier sessions)

He did a cold “poke holes in a strategy” session — useful as a template for how he evaluates ideas.

Crisis / regime shift

• Systematic: judge distributional shift; still may trade if absolute edge is huge even with weaker relative edge.
• Volatility + forced flow → good for unforced liquidity providers.
• Discretionary: process, trusted peers, “soft” quality of decisions.

8 ML lens — talk his language

He doesn’t center “deep learning hype.” He centers decision quality under competition. Translate your ML story into that frame.

Topics where ML naturally fits his worldview

• Estimating adverse selection / toxicity of flow.
• Regime detection / when to disable a strategy (human+model hybrid — he already believes this exists).
• Cost-aware optimization (Law 7).
• Uncertainty quantification (you can’t “know” edge — only bound it).
• Simulation of execution / market impact (capacity).

8b Example: NoCap as risk capital (your story)

Use this when he asks for a concrete ML example. Map your nocap-agent-architecture work into his laws — not as “I built a cool agent,” but as risk-managed research capital under competition.

“I treat cloud credits and GPU hours the way a desk treats risk capital: I only take the scientific risk I’m paid to take, and I systematically hedge the risks that just burn bankroll — confounded experiments, silent re-spend, and ideas that look good on a noisy proxy.” — Spoken framing for interview

What NoCap actually is (one breath)

A closed-loop experiment runtime for BottleCapAI NoCap-Test: instrumented GPT-2 trainer + two-stage FineWeb triage + Modal serverless DAG + pre-registered keep/kill gates. Goal: reach val loss ≤ 3.3821 on one GPU faster than a re-benchmarked stock baseline — with a ~$25 cloud budget and negative results as first-class evidence.

Risk translation table (Lebron → your system)

Where is the edge?

Credits as automation of risk management

If you also build a broader agent orchestration framework with credits, connect it like this:

• Credits ≈ risk capital / buying power for autonomous work units (experiments, tools, scrapes).
• Orchestrator ≈ desk risk system — what may run, in what size, with what timeout, what is skipped as already done, what triggers flatten (kill / human gate).
• NoCap is the worked example of that philosophy on a real hiring metric: multi-phase DAG, keep/kill policy, ~$25 book, idempotent execution, audit trail (logs, IDEA.md negatives, RESULTS.md).
• General product claim: you automate the boring half of research alpha — capital allocation, stop-loss on bad arms, and evidence packaging — so humans only set policy and interpret edge.

30-second spoken version

“For the NoCap challenge I built a serverless research runtime, not a chatbot. Cloud credits are the bankroll. The orchestrator is the risk system: smoke before size, parallel arms with pre-registered kill gates, idempotent phases so I don’t double-spend, and a hard rule that the only risk I’m paid to take is ‘does this idea beat my re-benchmarked baseline on one GPU.’ Architecture experiments that bought loss but destroyed wall-clock got killed or parked — Law 3 and Law 7. That’s the same muscle I’d bring to ML in a trading firm: treat experiment spend as risk capital, and treat methodology as part of the edge.”

If he pokes holes (be ready)

• “Isn’t the edge just data filtering?” — Possibly part of the PnL; the process that discovered, isolated, and cost-accounted it is what transfers to a firm.
• “Small budget, toy model.” — Agree on scale; the skill is capital-efficient learning under constraints (table selection + costs).
• “Where’s the model risk?” — Wrong baseline, leakage in val, composition risk, OOD from proxy→full; each has an explicit hedge in the design.

Source: software_i_built/modal/nocap-agent-architecture.html · rainbowpuffpuff/nocap

8c Baloney sandwich · AI agents as your edge

He explains Jane Street–style market making as the baloney sandwich business: buy components cheap, assemble, sell the finished product at a spread — a service with no deep directional view, but extreme pressure on costs, selection, and execution. Steal the analogy. Map it onto how you use AI agents.

“If I can buy baloney, bread, and cheese for $3, make the sandwich, and sell it for $4 — that’s a business. That is what ETF market making is.” — Outlier podcast · his framing of mechanical edge

Your sandwich (default mapping)

Multiple spoken variants (pick 1–2, don’t recite all)

How to land it without sounding gimmicky

• Name his analogy first (“your baloney sandwich point”) so it’s homage, not theft in the dark.
• One concrete example immediately after: NoCap, transcript atlas, research DAG — finished sandwich, not abstract.
• Admit the hard part is adverse selection on agent output — that shows you heard the real lesson.
• Don’t claim “agents replace traders.” Claim: agents compress research manufacturing under risk limits.

Why you’re not the marginal participant

Edge (his definition): something you know or can do that the marginal person in the same game doesn’t or can’t. In “people using AI at work,” the marginal participant is not “someone without GPT.” It’s someone who treats the model as magic autocomplete: vague prompt → accept first answer → pay full rework tax. Your gap vs that margin is concrete.

Risks you’re paid for vs risks you refuse

8d His phrase bank — embed this in your brain

Built from high-count content bigrams/trigrams in the corpus plus how he actually uses them. Goal: when he speaks, you already have the slot filled — and you can reuse the same language without parroting randomly.

Institutional world (he lives here)

Edge, risk, selection (core doctrine)

Retail vs pro / “wrong table”

Origin story triggers (if bio comes up)

Flash drill (2 minutes before the call)

8e Colonize adjacent · your arc + conquering hiring AS

He describes great trading firms as: start with something simple and self-contained, get good, then take the next most similar thing you’re not doing yet — “colonize adjacent markets.” Plug holes so even the weakest part is still decent. Your career is the same pattern on skills and proof surfaces.

Your colonization map (say it in order)

Why this is edge (not a random resume walk)

• Self-contained start: DS problems you could finish and measure.
• Transfer, not tourism: each step reuses eval discipline; only one major variable changes (modality → hardware → product → competitive ranking).
• Contestability: competitions = other agents in the market; win rate is a crude but real “am I better than marginal participants in that game?”
• Auditable inventory: site + product URLs reduce reliance on private claims.

Hiring is adverse selection — how you conquer it

His point: the hiring market is stacked against employers and noisy for candidates.

• Applicants on the market are selected against — best people are often retained; open market is not a random sample of talent.
• Final accept is the candidate’s — employers can win a bidding war and still get winner’s curse if signal was wrong.
• So both sides need process that reduces AS, not better vibes.

What employers fear (winner’s curse)

• Talks well, can’t ship under constraints
• Leaderboard tourism without reusable method
• AI cosplay — chat logs, no judgment on risk/eval
• Will leave / never was available (retained people problem inverted)

Your anti-AS kit (make selection work for you)

Pre-interview checklist (this section)

1. Name one competition with true participants + result (precise).
2. Name one failure you killed with evidence (AS-friendly honesty).
3. Open carolcalin.com + think2earn.com paths you’ll actually point to.
4. One sentence: why this role is the next adjacent tile, not a random jump.
5. One sentence: risk you’re paid for = time/reputation here; what you hedge with tools/agents.

8f Motivation · costs vs edge · why work at a firm

Two laws at once: Law 1 Motivation (know why before you act) and Law 7 Costs (if costs seem tiny vs edge, you’re wrong about at least one). Your honest read: you have some edge, but costs are high — so the rational move is not “go solo forever,” it’s to change the cost structure by joining a firm that supplies what you’re missing.

If costs seem negligible compared to your edge, you’re wrong about at least one of them. Attention and risk management are costs too. — Laws of Trading · costs / capacity family

Your situation in his language

Your motivation (Law 1) — use this, it’s clean

You want to be paid to take the right risks and buy the missing inputs:

• Paid to learn — learning is currently an unpaid cost; you want it on the firm’s books as investment in you
• Stability — reduces personal variance so you can size intellectual risk sanely
• Structure — process, reviews, defaults = institutional priors (stored posteriors)
• Innovation of the company — real problems with real constraints, not toy loops
• Expertise of founders / team — mentors compress what takes years alone (he stresses this constantly)
• Collective work hours — leverage the integrated product of many people’s time (the firm’s sandwich line), not only your solo kitchen

Map motivation → his hiring / firm story

8g Bayesian software decision process

Practical loop for product, ML, and incidents — aligns with his process-vs-outcome honesty and your keep/kill experiment culture.

• Pre-register what would move you (kill gates, SLOs) so you don’t invent LR after seeing data.
• Benchmarks price model capability before you bet a day (LR on “model can do task T”).
• Competitions update prior on “am I better than marginal participants in this game?”
• Firm structure is a warehouse of team posteriors — joining buys better priors faster than solo tuition.

9 Career advice & “don’t trade”

If you’re smart enough and hardworking enough to make a living in financial markets, there’s probably an easier way to make money and have a satisfying life. — Trading, Crypto & Adverse Selection podcast

Who should become a quant trader?

• Analytical + statistical comfort and emotional comfort with incomplete information and money on the line.
• Understands good decision ≠ good outcome (and the reverse).
• Willing to be mentored; he stresses mentors accelerate learning for retail and pros.
• Sam Bankman-Fried anecdote (as intern manager): noticeable for asking all the questions day one — intellectual aggression as a positive signal in that culture.

10 Hiring, talent, culture

• He led a CTO-room conversation on compensation: labor markets, retention, honesty over platitudes.
• Hiring is riddled with adverse selection (see §6).
• Software orgs are “sociology” — processes, culture, conventions; he generalizes decision frameworks beyond trading.
• At Jane Street style firms: plug weaknesses; breadth of competence matters because any hole gets exploited.
• For candidates: curiosity, clear thinking, ability to reason under uncertainty — more than buzzwords.

11 Questions you can ask him

Prefer questions that show you read his mental models — not generic “what’s a day in the life.”

12 How to present yourself (ML + trading role)

One-sentence positioning (customize)

“I build ML systems the way a careful trader builds strategies — clear motivation, explicit edge, cost-aware evaluation, and a plan for when the world goes out of distribution.”

Full pitch skeleton (2 minutes)

1. Motivation (Law 1): paid to learn · stability · structure · company innovation · founder/team expertise · collective hours.
2. Edge: non-marginal AI user — CLI/libs, model envelope, benchmarks; sandwich line; colonization DS→vision→hardware; competitions; public portfolio.
3. Costs (Law 7): solo costs (attention, RM, learning) are high vs that edge → join to improve cost structure, not because edge is zero.
4. Risk paid for: time/reputation here; hedge unpaid toil with agents/tools.
5. Hiring AS: make yourself easy to select — artifacts, contestability, fast falsifiability.
6. Close: adjacent tile, not leftover inventory; curious peer, honest uncertainty.

Deep links: motivation + costs · NoCap · sandwich / agents · colonization + hiring AS · Bayesian process · phrase bank · today’s synthesis.

12b Today’s synthesis — all threads in one place

Everything we built in this prep arc, compressed for final review.

Company deep-dive: §0 EquiLibre / DeepStack.

13 Landmines to avoid

• Retail hero stories — “I beat the market with RSI on YouTube.” He’ll think wrong table.
• Ignoring costs — commissions aren’t the only cost; spread, impact, attention, risk. Also: pretending solo costs are low.
• Certainty cosplay — he respects probabilistic honesty; overclaiming edge is a red flag.
• Emulate MM as retail — he explicitly calls that a fool’s game for individuals.
• Outcome worship — “it made money so it was right” violates his process/outcome distinction.
• Name-dropping SBF casually — he managed that intern cohort; keep factual/neutral if it comes up.
• Black-box flex without economic story or failure analysis.
• “I just need a paycheck” without Law 1 — pair stability with paid learning / structure / team density.
• “I have no edge” — too weak; better: edge real, costs high, firm improves the ratio.
• Begging for hire without falsifiability — increase AS for them; show how they’d know quickly.

14 Source index (your 10 transcripts)

Full text: equilibre/transcripts/ · Deepgram JSON: equilibre/deepgram/

90-second pre-interview checklist

1. EquiLibre: lab-first RL trading · DeepStack lineage · 3 founders · Prague · Tower partnership · continual multi-agent adaptation.
2. Motivation (Law 1): paid to learn · stability · structure · innovation · founders/team · collective hours.
3. Edge: vs marginal AI user (tools, envelope, benchmarks) + colonization proof + one competition result.
4. Costs (Law 7): edge real, solo costs high → firm improves edge/cost; not “I have nothing.”
5. Risk paid for: time/reputation here; agents hedge unpaid toil.
6. Sandwich or NoCap: one finished-artifact example + adverse selection on outputs.
7. Hiring AS: artifacts + contestability + “you’d know fast if wrong.”
8. Bayesian: prior → evidence → posterior → new prior (defaults/runbooks/team).
9. Question for them (continual learning / DeepStack transfer / keep-kill culture).
10. Energy: curious peer, honest uncertainty — not fan, not retail tips seeker.