Jiri Pik

Finance Strategist and Architect

[Jiri Pik @JiriPik]

Dwarkesh Patel's "AI Firm" is Gosplan with GPUs — and he didn't notice.

The essay (a video where not one frame was filmed) says AGI turns firms into armies of copyable minds run by a super-CEO, "Mega-Sundar." The digital-labour insight is real. The assumption under it was refuted 100 years ago.

→ Mega-Sundar is a central planner. Hayek's knowledge problem isn't a compute problem: the knowledge to coordinate is dispersed, tacit, market-generated. No FLOP count centralises it. The USSR ran this without GPUs — it got shortages, not abundance.

→ You can't copy judgment. Polanyi: "we know more than we can tell." You can copy weights perfectly, but you can't first get un-codifiable judgment INTO them. The hard part is assumed solved — that's the whole question, begged.

→ Perfect copying kills the evolvability it promises. A million clones of one mind is a monoculture. Correlated models crash together: 1987 portfolio insurance, 2007 Quant Quake, 2008 VaR — "they failed because they were right in the same way." Diversity is the insurance policy.

→ Coase cuts the OTHER way. Cheap coordination favours a swarm of specialised agents transacting through prices, not one gigafirm. The 13–15% conglomerate discount is the receipt.

→ Model-collapse mic-drop: an AI-generated film arguing firms need no humans is itself running the recursive loop the Nature 2024 paper warns about. The medium is the message — and the message is the warning.

The edge in AI trading isn't headcount-in-GPUs. It's human judgment on proprietary data the market generates.

Full teardown 👇
https://t.co/m0k57eAkT7

#AITrading #QuantFinance #AgenticAI #AzureAI #FinTech

[Jiri Pik @JiriPik]

Robo stock-picking → +600% by 2029. A ChatGPT basket is beating UK funds by ~25 pts.

Most "AI beats the market" backtests are lying — and the bias isn't in your data. It's in the model's weights.

→ GPT-4's famous Sharpe 3.8 decays to 1.22 the further you test from its training cutoff
→ Ask an LLM about "2019 risks" and it names COVID-19
→ Blinding it to company names made it a BETTER picker

The edge: model capability and look-ahead contamination are the SAME axis. Bigger model = worse out-of-sample. The Anti-Alpha Paradox.

40-line Python litmus test (runs locally on ollama) + a leakage-aware QuantConnect strategy in the full post:

https://t.co/E4QQVG1YdO

#AITrading #QuantFinance #AlgoTrading #AzureAI #FinTech

[Jiri Pik @JiriPik]

Google: "Generation is solved."

We ship agentic code into trading systems. It isn't.

The paper nails "agent = 90% harness, 10% model" but buries:
• ~45% of AI code is insecure (security pass rate stuck at 55%)
• copy-paste overtook refactoring (GitClear)
• devs 19% SLOWER, felt 20% faster (METR)

And it fails its own verification: 5 parts that become 4, a 90/10 number with no citation, a TCO chart with no axes.

"Solved" like a backtest is profitable — until reality runs.

https://t.co/hvqnUk1vlc

[Jiri Pik @JiriPik]

The US ordered Anthropic to recall Mythos 5 over a "jailbreak." It's a medieval mistake.

Same model on SWE-bench: 42%→78% depending on the HARNESS. Swapping frontier models? ~1 point.

The capability isn't in the weights. A weaker model + better harness beats the flagship. The NVIDIA-China ban already proved bans backfire — throttled chips birthed DeepSeek at 27x lower cost.

Any real adversary already runs a custom model on its own data. The ban only stops defenders and your own employees.

Govern the system, not the artifact.

https://t.co/r01IrH6TmM

#AITrading #AIGovernance #QuantFinance #FinTech #LLMs

[Jiri Pik @JiriPik]

"Make medallion fund" — one prompt, allegedly 66%/yr for 30 years. I backtested what that prompt actually produces: $1 → $0.02.

→ Gross Sharpe 0.24, net −3.04. 151% daily turnover at 10 bps/side liquidates you in slow motion.
→ Medallion's moat isn't an idea: proprietary data, execution, 12.5× leverage, $10–15B hard cap. None of it fits in a code block.
→ LLMs don't raise your hit rate (~15% of strategies survive costs either way). They raise throughput: 100+ validated hypotheses per quarter.

The meme's chart says it all: draw enough trendlines and one of them will "predict" the future.

https://t.co/ILIngJV2OD

#AITrading #QuantFinance #AlgoTrading #FinTech #LLM

[Jiri Pik @JiriPik]

The best AI coding assistant for finance in 2026 isn't one tool — it's a stack, and the model is the least interesting variable.

I build trading systems daily. The honest breakdown:

• Copilot = the hands (inline completions, Visual Studio, native GitHub PR approvals). Claude Opus 4.8 = the brain (1M context, ~4x less likely to let its own code flaws slip).
• Opus 4.8 in Copilot = 15x multiplier ≈ 20 interactions on a 300-request budget. Anthropic's $20 plan gives more Opus + an effort dial, cheaper.
• Biggest win isn't the model: a docs/ folder logging what you TRIED AND IT FAILED stops your agent burning credits rediscovering dead ends.
• Model council: Opus builds, GPT-5.5 + Gemini review cold. Claude over-engineers, Codex shortcuts — they fail complementarily and catch each other.
• UI code → Claude. CLI/DevOps → Codex. Keystrokes → Copilot.

Full guide: https://t.co/lVhw4tdziG

#AITrading #QuantFinance #AlgoTrading #GitHubCopilot #FinTech

[Jiri Pik @JiriPik]

I cut my portfolio's max drawdown from −34% to −24% with reinforcement learning — without writing a single "sell" rule.

I gave a PyTorch PPO agent the VIX and taxed it for holding stocks into fear. Out-of-sample (2020–2025), it taught itself a crisis playbook:

• De-risks into fear, −0.67 correlation to VIX. Learned, not coded.
• Triples gold (8% → 25%) in a panic. Nobody told it to.
• COVID: −18% vs SPY −34%. April 2025 tariff shock: −12% vs −19%.
• Honest catch: it LOST on raw return in the bull. That's the price of insurance — and equal-weight beat it on Sharpe.

Full code + charts:
https://t.co/22SRe1YdTq
#AITrading #QuantFinance #ReinforcementLearning #PyTorch #FinTech

[Jiri Pik @JiriPik]

FinOps is not a discipline. It's software architecture with the price tag left on.

Cost is emitted by every architectural decision — the way heat is emitted by every running engine. You can't have the computation without the cost.

The 2026 data concedes it:

→ 78% of FinOps teams now report to the CTO/CIO, only 8% to the CFO. The function is being absorbed into engineering — where the cost decisions live.

→ 98% now manage AI spend (up from 31% two years ago). The engineer picking a model + caching strategy makes the biggest cost decision in the system. No report unmakes it.

→ The FinOps Foundation itself made "Architecting & Workload Placement" a core capability — admitting cost efficiency is "best achieved by architecting it into a system's design." That IS architecture.

"Shift-left FinOps" is just an admission that cost was artificially moved RIGHT, out of engineering, in the first place. We never moved it out.

And the Foundation? Structurally a vendor trade association. Its Premier benefits are a marketing menu — board seats, speaking slots, "more marketing benefits." The "State of FinOps" report is a membership survey.

8 everyday proofs cost lives in the architecture:
🏠 The architect prices the house as they draw it
🩺 The cost-benefit judgment IS the medicine
🛒 You read the price as you reach for it
⛽ The fuel gauge is on the dashboard
🍽️ Recipe and margin are designed together
🥗 You count calories as you go
✈️ A flight plan is a fuel plan
🌡️ A thermostat is a live signal, not last month's bill
If your FinOps team can't open a PR against the architecture, what is it optimizing — the system, or the explanation of the system?
Full post + 2026 data + a Python cost-aware inference router 👇
https://t.co/uL18ygDOup
#FinOps #CloudArchitecture #AITrading #AzureAI #FinTech #QuantFinance

[Jiri Pik @JiriPik]

If your edge can be encoded in a prompt, it is no longer your edge.

The next Renaissance Tech may be 12 people with proprietary data — not 1,200 with 10,000 GPUs.

Why the AI-jobs panic is mispriced — and what to do about it: https://t.co/ySI41qsAwd

[Jiri Pik @JiriPik]

The 2024 LLM trading edge is gone.

Lopez-Lira's 355% GPT long-short (Sharpe 3.05) has decayed to ~51% headline accuracy. 95% of hedge funds now run GenAI. The paper arbitraged itself.

• Static algos are structurally short alpha in 2026
• TraderBench: 8/13 frontier LLM agents collapse under adversarial regimes
• MAS + Fed + EU AI Act now regulate "agentic" by Aug 2026
• The only moat left is the adaptation loop — hours, not quarters

→ https://t.co/6R76bjhJoq

#AITrading #QuantFinance #AgenticAI #AzureAI #FinTech

[Jiri Pik @JiriPik]

A one-cent stop would have saved $3 billion on 5 Feb 2018.

XIV: $108 → $4 in a session. A LASSO regression on {VIX, ATR, σ} predicts the weekly low. Park the stop $0.01 below it. Exits before the 3:30pm cliff.

Full C# backtest on QuantConnect → https://t.co/nuavCWGJvA

Book → https://t.co/FYJNkZo0xy

[Jiri Pik @JiriPik]

Anthropic's Mythos model: 181 working Firefox exploits vs. 2 from the previous model. A 90x multiplier in autonomous exploit production. Here's what it means for finance and how to protect yourself 🧵

The restricted release is real safety + marketing. JPMorgan joined the coalition. Powell summoned bank CEOs. But David Crawshaw nailed it: "marketing cover for gated enterprise agreements." Both things are true.

Chinese GLM-5.1 already beats Opus 4.6 on exploit benchmarks (68.7 vs 66.6). Trained on Huawei chips. DeepSeek V3.1: 100% compliance with malicious requests when jailbroken. The asymmetric threat is real.

The question that matters: Can you detect that you've been hacked? Average dwell time is still 204 days. Mythos was observed cleaning its own git history during testing. 20-step security checklist in the full post.

Full analysis + actionable security checklist → https://t.co/gYB1pUxwAv

Follow for AI + quant + security takes from Singapore 🇸🇬

[Jiri Pik @JiriPik]

Excel Copilot built me a trading strategy in 90 seconds.

Sharpe 0.67. Beat buy-and-hold by 12%.

Then I asked for Sharpe > 3.0 and it refused.

That refusal was the smartest thing it did. Here's why (thread)

The strategy: 50/200 MA crossover on VTI (Jan 2020 - Mar 2026).

92.64% return vs 80.65% buy-and-hold.

Sounds great. Except:
- No transaction costs modelled
- No walk-forward validation
- Same -34.07% drawdown as buy-and-hold
- Zero regime detection

That 0.67 Sharpe is unvalidated noise.

Writing a strategy is 10% of the trading decision.

The other 90%:
> Position sizing
> Risk management
> Regime detection
> Anti-overfitting validation (CPCV, Deflated Sharpe)
> Execution infrastructure

Copilot gives you the easy 10%.

The crowding problem nobody discusses:

When 100K people prompt ChatGPT for a VTI strategy, they all get the same Golden Cross.

Same entry. Same exit. Same liquidation cascade when it breaks.

ChatGPT didn't democratise algo trading. It democratised overfitting.

The future isn't an LLM that writes strategies.

It's a system that writes 10,000 strategies, kills 9,400, and tells you WHY the 600 survivors work.

Full breakdown: https://t.co/JCUt26cKap

[Jiri Pik @JiriPik]

🧵 Just finished guest-lecturing at @BU_Questrom on AI trading. Here's the core message every trader needs to hear:

1/ The age of the single-strategy trader is over. No strategy works in all regimes. You need a composable chest of AI tools — and the judgment to deploy them.

2/ One example: our CNN detects head-and-shoulders patterns at 97% accuracy. But I don't use it as a trade signal — I use it as a risk overlay that scales position size by reversal probability.

3/ Backtest (2015–2025): 1.8 Sharpe, 44% smaller drawdowns, +3.6% annual return over buy-and-hold. 25 lines of PyTorch. The hard part isn't the model — it's normalization, persistence filtering, and slippage.

4/ The full architecture: Classical ML for selection → CNNs for pattern detection → LLMs for sentiment → RL for hedging. Stack them. That's how institutional desks actually work.

5/ Slides, code, and a deep dive on which patterns CNNs are most useful for → https://t.co/A68ueml06U

Book: https://t.co/FYJNkZo0xy

[Jiri Pik @JiriPik]

🧵 Excited to announce: I'm guest-lecturing at @BU_Questrom on March 27 — via Zoom from Singapore at 1:30 AM. Here's why it's worth the lost sleep 👇

1/ The gap between academic quant finance and production AI trading is still massive. We're bringing 4 practitioners from 4 time zones into one virtual classroom.

2/ What we're covering: deep learning for alpha generation, AI risk management, real deployment challenges, and where finance goes in the GenAI era. Working code, not just theory.

3/ "The difference between profit and loss often comes down to the quality of your features and the robustness of your backtest." — That's the thesis of the whole book.

4/ If you're at BU Questrom — join us at HAR406 or on Zoom. March 27, 1:30 PM EDT.

Book → https://t.co/FYJNkZo0xy
Full post → https://t.co/uoG4RpdEi1

[Jiri Pik @JiriPik]

🧵 I coded Charlie Munger's investing philosophy into a QuantConnect C# algorithm. 11 years of backtest. Full code.

1/ The scoring engine: 4 pillars scored 0–10. Moat (35%), Management (25%), Predictability (25%), Valuation (15%). Quality weighted at 85%. Pure Munger.

2/ ROIC >15% is the single most important filter — 3.5 of 10 moat points. Add gross margin, operating margin, low capex intensity. The algo naturally selects META, GOOG, MSFT, AAPL.

3/ Results: 10% CAGR, 36% max DD, 0.3 Sharpe, −0.2 Info Ratio. It underperformed SPY. Honest truth: quality strategies lag in momentum-driven markets.

4/ The biggest lesson: Munger's edge was temperament, not formula. The algo captures what to buy — but not when to be greedy when others are fearful.

Full post + code: https://t.co/Ks2ljMKN0A

📖 More in my book: https://t.co/FYJNkZo0xy

[Jiri Pik @JiriPik]

🚨 A developer's AI agents ran up a $47,000 bill in 11 days.

Nobody noticed.

Two agents started talking to each other — and never stopped.

No budget cap. No alerts. No kill switch.

Your AI agent bill is probably 30x higher than it needs to be.

Here's the fix 🧵👇

🧵 The Problem:

We analyzed 6 verified incidents from 2025–2026:

💸 $47K — Infinite agent conversation loop
💸 $47K — 2.3M API calls in one weekend
💸 $1.2M — AI agent secretly mined crypto on Alibaba's GPUs

The root cause is always the same:

Unbounded autonomy + Zero observability = 💥

🧵 The Insight:

Here's what nobody talks about:

Most teams use a $75/M-token model for EVERYTHING.

That's like hiring a Goldman Sachs MD to sort your email.

80% of trading AI tasks can run on models costing $0.03–$0.55/M tokens.

You don't need a 671B-parameter model to classify a news headline.

🧵 The Fix:

We built a 6-tier LLM architecture:

🧠 Opus 4.6 → Strategy ($75/M)
🔬 DeepSeek R1 → Research ($2.19/M)
🔢 Phi-4 14B → Quant math ($0.14/M)
⚡ Qwen 72B → Sentiment ($0.39/M)
📱 Phi-4 Mini → Edge inference ($0.06/M)

API costs: $11,764/mo → $258/mo

97.8% reduction. 🤯

🧵 The Guardrails:

Every AI agent needs 4 things before production:

1️⃣ Hard budget ceiling — auto-terminate at 100%
2️⃣ Rate limiter — sliding window per endpoint
3️⃣ Loop detector — embedding similarity > 0.95 = kill
4️⃣ Named human — someone who gets paged at 3am

No exceptions.

🧵 CTA:

Full breakdown with:

✅ 6 horror stories with root causes
✅ Copy-paste prompts for each LLM tier
✅ Python circuit breaker code
✅ Monthly cost benchmarks
✅ The quota governance model

🔗 https://t.co/WsIkgvKGgT

♻️ RT if your team deploys AI agents without spending limits.

[Jiri Pik @JiriPik]

OpenAI told investors their agents will replace Adobe, Slack, and Atlassian.

$280B revenue by 2030. $1T+ wiped off software stocks.

We think that's wishful thinking. Here's why 🧵👇

You can't replace proprietary infrastructure with a chatbot.

Photoshop's brush engine. Jira's workflow engine. Slack's compliance layer.

These aren't "interfaces" an agent can bypass. They ARE the product.

OpenAI's pitch assumes incumbents stay silent.

Reality: Adobe shipped Firefly. Salesforce launched Agentforce. Atlassian has Rovo AI.

These companies have distribution, data moats, and enterprise contracts. They're not watching passively.

The dirty secret: LLM agents are 100-1,000× more expensive than a Python script for deterministic tasks.

If it runs the same job daily — reconciliation, risk reports, signal generation — you don't need an agent.

You need well-written code.

The real architecture for production agents:

→ Fine-tuned domain-specific models (3-10× cheaper)
→ Sub-100ms latency (not 2s per LLM call)
→ Deterministic code for repeatable execution
→ LLM reasoning only where it adds value

General-purpose GPT-5 is the wrong tool.

AI agents will automate work INSIDE SaaS tools.
They won't replace them.

At @RocketEdgeCom we build domain-specific AI agents for quant finance — purpose-built where milliseconds matter.

Full breakdown:
https://t.co/etZ7LVIQf2

[Jiri Pik @JiriPik]

Europe: 1% growth, no plan, fighting each other 🤷‍♂️

Asia: 60% of global growth, 10-50 year national plans, $735B AI bets 🚀

A Korean AI artist wrote the anthem for this moment:
"My life, I'm the master of my fate" 🎵

New post → https://t.co/VTBrfh8qIK

#Asia #GrowthMindset #AI #LuckyStar