6/6 Buyers are getting smarter fast. They'll learn to evaluate agencies, define success metrics, and scrutinise solutions. If we want to deliver on the AI agenda, we need real ML depth: experimentation methodology, mathematical foundations, production infrastructure, and domain expertise.

5/6 Here's the real issue: LLMs are a communication layer between user and system. If you want real intelligence, you build the AI system behind the LLM. Too many agencies just wrap an LLM and call it a recommendation engine. That's not AI. That's a shortcut.

4/6 Traditional ML: long lead time, high upfront cost, but cheap and reliable in production. LLMs invert this. Connect Copilot to Fabric, looks like magic. Time-to-demo: near zero. But production cost and complexity? Enormous. The demo creates expectations that reality can't match.

3/6 I've worked with many AI agencies, local and international. Most are just Python dev shops. Developers who evolved into ML roles without a deep understanding of AI methodologies, algorithms, math, or proper tooling. They ship code, not intelligence.

2/6 Before COVID, selling ML was brutal. The market was tiny. Only large corps with internal ML teams bought in.

LLMs changed everything. Now every company has AI on its agenda. The market exploded.

But a dangerous gap is forming between C-level expectations and actual delivery capability.

6/ The real lesson from 35 Snapchat launches: at massive scale, the AI model is rarely what matters most. Data quality, smart infrastructure, and asking the right questions drive the wins. GiGL is open-source if you want to explore: github.com/Snapchat/GiGL — What would you build with it?

5/ The wildest finding: instead of asking "what does this user like?", they asked "what do this user's FRIENDS like?" No new model. No retraining. Just a smarter question over the same data. Result: +13.9% on key metrics. Sometimes the breakthrough isn't a better answer, it's a better question.

4/ They tested every graph AI variant available. The winner is the attention-based models because not all your friends matter equally for predictions. Some connections are strong, others noise. Attention learns the difference automatically. Which of YOUR friends would an algorithm weight highest?

3/ But here's what really surprised me. The biggest gains had nothing to do with smarter models. They just changed which relationships they looked at, from "who's friends" to "who actually talks". Same AI, different data. +8.9% lift. One data cleanup step boosted accuracy by 38%.

2/ Their trick? Treat it like a boring data pipeline, not a cutting-edge AI system. Pre-compute everything, store it, then train as you would in a normal ML job. No exotic infrastructure needed at training time. This one shift cut costs by 80%. Sometimes the boring solution wins.

(7/7) The psychoanalyst argued that "good enough" parenting, allowing manageable frustrations rather than perfect responsiveness, builds more resilient children.

Maybe AI is the same: don't tell it what's perfect. Just help it avoid what's wrong.

📄 arxiv.org/pdf/2506.01347

(6/7) Whether training reasoning models or building recommendation engines, diversity preservation isn't a nice-to-have.

It's essential for long-term performance.

There's something almost Winnicott-ian here...

(5/7) Netflix discovered that negative signals (skips, abandons, dislikes) can be more informative than likes.

They tell you what to avoid rather than what to obsess over.

The deeper principle is that entropy is a resource. Once collapsed, it's nearly impossible to recover.

(4/7) This maps directly to recommendation systems.

We've battled this tradeoff for years: exploitation vs. exploration.

Too much positive reinforcement creates filter bubbles, trapping users in increasingly narrow content loops.

(3/7) When you suppress a wrong answer, probability redistributes across alternatives proportionally to what the model already believes.

Rewarding correct answers creates overconfidence. The model locks onto specific paths and loses its ability to explore alternatives.

(2/7) When training LLMs on math problems, penalising failures, without ever rewarding successes, matches or beats standard reinforcement learning.

At higher attempts? It actually outperforms everything.

The hypothesis is: punishment preserves diversity.

Nadella’s AI strategy is just a pitch for the shareholders. There is no idea, no details, no plans or tangible actions. The most embarrassing thing is that his own team occasionally refutes him. And his Director of AI? His posts are just low-quality chatter about AI.

Don’t tell this to MAGA people, they will interpret it as he is 38x better than Jesus 🤭