Scott Barber

I present you: the latest state of the art in parameter tuning. SEC's case against Jian - just published last week - was a wild read. For those who are living under a rock, this was the researcher whose bonus jumped to $23M within 5 years out of school and then went around gloating on Chinese social media. Notable details: 1. No one seemed to bat an eye about awarding a 9x YoY bonus jump from $2.5M to $23M for the newcomer, until it went on social media: > Wu’s scheme began to unravel in January 2023, when TSI discovered an anonymous social media post revealing that a TSI employee had received over $23 million in compensation for 2022, prompting an investigation into Wu’s Models. 2. The curse of modern day code review practices: > because engineers implemented ticket system changes without substantively reviewing the requests or ensuring management approval, modelers could also use the ticket system to circumvent the [...] process" When I first heard of this case, my original suspicion was a flaw in the PnL attribution model or a novel way to game it through market impact. It turned out to be much simpler - who knew you could 10x your bonus just by adding a few zeroes. 3. This case exposes a practical issue of perverse incentives with model-based attribution for alpha researchers. 4. Which brings me to an interesting thought: if you're running a high ADV strategy and about to leave for a competitor, your competition is yourself... (I know for a fact some of my signals still live on in top 5 ADV settings after several years.) So what's there preventing you from just applying *1e-1, *1e-2 or gently perturbing a few monetization parameters before heading out of the door?

145

12 Comments

This is what AI 'thinks' of my interview with Prof. Kevin Werbach 😅: "David Hardoon delivers a masterclass in Accountable AI! From Singapore’s FEAT principles to Standard Chartered’s AI innovation, he proves governance is a catalyst for trust and progress in fintech. A must-watch for anyone shaping AI’s future—accountability isn’t a burden, it’s a competitive edge! 🚀" I think AI is being far too kind, or maybe its just AI particularly agreeing with the core theme that one can't outsource accountability in AI to AI 😁 Key Takeaways: Justify, Don’t Blindly Trust: AI decisions must be rigorously interrogated to ensure fairness and transparency. Governance Drives Innovation: Embedding accountability in AI processes boosts resilience and value, not bureaucracy. GenAI’s Dual Edge: Harness its potential to fight scams, but only with robust, human-led oversight. Global Lessons: Singapore’s collaborative AI policies (like Veritas) set a benchmark for balancing regulation and innovation. https://lnkd.in/gUVAyrWx

116

8 Comments

#Malaysia Just Dropped the World’s First #AI #Bank. #Ryt Bank, the first fully AI-powered bank that runs on Malaysia’s own large language model (ILMU) #LLM, offers 24/7 #multilingual AI banking assistants and bundles savings, credit, payments, and cards into a single app. #No #paperwork, #no #fees, #no #waittime — just #conversational #banking with daily interest and instant credit lines. This isn’t just a digital front-end tweak; it’s a direct assault on Malaysia’s traditional banks. CIMB, Maybank, Public Bank — all built on legacy infra and human-heavy processes — now face a competitor that onboards in seconds, speaks every language, gives 4% daily interest, and issues credit instantly. https://lnkd.in/gwe37ZjS

38

In most banks, growth performance data is scattered across channels, products, and teams. What if that data could teach your system to make the next plan even more successful? That’s exactly what we’ve built with the Customer Lifetime Orchestrator (CLO). Each plan you run, whether to expand deposits or grow savings, feeds back into the system. CLO learns what message, what image, and what channel drives the best results. So over time, your AI agents don’t just automate — they optimize. Every cycle becomes smarter, faster, and more effective. That’s the power of self-learning, machine-learning-driven agentic architecture. The more you use it, the better it performs. How far are you in turning your growth strategies into self-improving growth engines?

56