Your AI can change overnight

What I’m building with AI + one hot take + five links worth your time.

I spent the whole week using Fable 5, and it has been amazing.

I'm starting with my experience because my feed is full of people saying the opposite: "It's nerfed", "lobotomize", "it's not the same model"

In a benchmark, BridgeMind ran a coding suite on the day Fable was back and saw debugging drop from 86.2 to 25.9.

Both are true; it depends which tasks you run.

Fable is not dumber. The model weights are the same as in the version that the U.S. government forced to be removed in June.

What changed was the safety classifier that sits in front of the model.

When your request is about cybersecurity or biology (even if barely), it's rerouted to Opus 4.8. In the BridgeMind benchmark, only 3 of 12 debugging tasks reached Fable.

The other ones failed because the answer came from a less capable model.

So why am I happy? Well, because I'm not writing exploits. What I do doesn't look dangerous to the filters.

Arena.ai, which runs blind human-preference votes, published results the same day and found Fable to be mostly consistent before and after.

Same model, same week, completely different results. And the difference lies in the types of tasks that matter to the user.

Here's the part we all should think about. You're paying top dollar for the best model, and it can swap out underneath you. Maybe flagged in chat, silent in an agent unless you built for it.

And the filter can tighten again next week: same code, broken on Monday. Sadly, the models aren't stable ground to build on. Keep running your evals.

If you're on Fable or any other frontier model, log which model actually answered. If you're still in the playground, same rule: never trust, always check.

​Most people use Claude Code like a fancy autocomplete. This workshop is about the system around it: the CLAUDE.md, the MCP connections, the guardrails that make the output trustworthy enough to ship.

Gabriela de Queiroz and I are running it live with Packt. 3.5 hours, hands-on, you leave with implementation templates you can use the next day.

40% off for the next 7 days → Spots Are Limited - Register here

This week I met someone who deployed AI as an underwriter. It approves the loan and moves the money without a human signing off.

In a regulated industry, this should be impossible because you can't reject a loan on "a vibe".

You need explainability all the way. But it works here for the reason most people miss when they talk about autonomous agents. It's not the agent, it's the rest.

I lived this at QuickCheck; back then we published 8 papers on credit scoring. And something I learned that is even more important in the LLM era: a lending decision needs a reason; whatever the decision is, it needs to be explained.

For years, this meant you could only automate using deterministic algorithms. Regressions, fancy trees like XGBoost.

So things you can point and say: we rejected because of this number in the formula. Anything ambiguous would go to a human, and a human takes hours.

There's a real cost to those hours. The longer a person waits, the higher the chance they move to a faster lender and you lose them.

At QuickCheck, we cut scoring time from hours to less than 5 minutes by keeping the decision deterministic (brittle Python scripts back then, but explainable ones).

So how to do it in the LLM age? You merge everything: microservices, rules engine, decision engine, then an agent on top that writes the underwriting rationale and releases the funds when the criteria are met.

Humans are there, but only touch the edge cases.

Look at what the non-deterministic layer (the LLM) does: it broadens what you can automate, but it doesn't make the call; the decision comes from the deterministic functions.

The autonomous agents are amazing at coordinating older systems and executing on their decisions.

That's the workflow worth stealing. For any automated decision you ship, ask one question: if a regulator or a customer demands the reason, is your answer the model did it, or can you point to a deterministic rule?

In lending, an approximation is a weak thing to hand a regulator. Keep the decision in the layer you can explain and let the model handle the rest.

AI agent runs first end-to-end ransomware attack​

The same agent you use to build your apps autonomously just went to the dark side. Fully automated exploits in 31 seconds.

​Autoresearch, Claude and Constrained Optimization​

Auto loops only work when the KPI is well defined. In most real-world problems, that's not the case, which is why the claim "I'm 10x faster with agents" isn't always true.

​Product Shape is the Moat​

OpenAI and Anthropic can’t be everything to everyone. If you're building on top of frontier lab tech, fine-tuning won't save you. Your moat is the product experience.

​Claude Code turned every engineer into three​

The bottleneck moved from writing code to deciding what to build. Now the scarce resource is product judgment, not how fast you can type.

​Learning to Replicate Expert Judgment in Financial Tasks​

Small models fine-tuned on proprietary data beat frontier models and cost less. As I've been saying, the future is SLM, small, specialized models./

Got something you need help with? Hit reply and let me know.

Know someone who can benefit from this email? Forward it.

New here? Someone forwarded this to you? → Subscribe to Build What Matters: ​https://luisrodrigues.ai/​

​

Follow me on ​LinkedIn