Michael Brenndoerfer

Private Equity Analytics Operating Partner; At the intersection of data & AI, strategy, finance, and entrepreneurship

[Michael Brenndoerfer @mbrenndoerfer]

@adambcohen93 Fantastic

[Michael Brenndoerfer @mbrenndoerfer]

@adambcohen93 @Standard_Cap Congratulations!

[Wall Street Rollup @WallStRollup]

Allbirds has raised $50mm to pivot its business to AI compute infrastructure

Allbirds now has a long-term vision to become a fully integrated GPU-as-a-Service (GPUaaS) and AI-native cloud solutions provider

[Y Combinator @ycombinator]

"If you want to do good work, what you need is a great curiosity about a promising question."

- Paul Graham, What You'll Wish You'd Known

[Y Combinator @ycombinator]

Paul Graham’s recipe for doing great work:

[Michael Brenndoerfer @mbrenndoerfer]

I break it down in this article: https://t.co/GPVqmxExgu

[Michael Brenndoerfer @mbrenndoerfer]

Curious how and why LLM fine-tuning actually works?

Fine-tuning modifies the model from the inside, updating its parameters. Prompt engineering and retrieval-augmented generation (RAG), on the other hand, operate outside the network.

(1/n)

[Michael Brenndoerfer @mbrenndoerfer]

Understanding the mathematics behind LLMs can feel daunting, but grasping the math behind fine-tuning, provided you have some basic linear-algebra intuition, is well within reach.

[Michael Brenndoerfer @mbrenndoerfer]

That sensitivity is what makes fine-tuning powerful: by updating a carefully chosen slice of parameters, you graft specialized expertise onto the model while preserving its broad, general knowledge.

[Michael Brenndoerfer @mbrenndoerfer]

Because representations are so interwoven, altering even 0.1 % of the weights can ripple through countless computations, producing noticeable behavioral shifts.

[Michael Brenndoerfer @mbrenndoerfer]

Transformer models store knowledge in a distributed way: no single weight or neuron corresponds neatly to “cat” or “legal terminology.” Instead, many neurons collaborate to encode each concept, and each neuron contributes to many concepts.

[Michael Brenndoerfer @mbrenndoerfer]

Picture the LLM neural network as an enormous switchboard: every connection has a weight controlling how strongly signals flow. In Transformer architectures, those weights live in large matrices: giant spreadsheets of numbers the model adjusts during training.

[Michael Brenndoerfer @mbrenndoerfer]

When we fine-tune, we begin with a model whose billions of parameters (weights) were learned from trillions of tokens. Fine-tuning nudges a tiny subset of those weights so the model acts as though it were trained on your domain-specific data.

[Michael Brenndoerfer @mbrenndoerfer]

Fine-tuning modifies the model from the inside, updating its parameters. Prompt engineering and retrieval-augmented generation (RAG), on the other hand, operate outside the network: they steer or supplement the model without touching its weights.

[Michael Brenndoerfer @mbrenndoerfer]

As images can be worth a thousand words, sometimes a bit of code tells the story better. Here's how to get started with agents, covering core concepts in Part 1.

https://t.co/MFY7f7xp3d

[Michael Brenndoerfer @mbrenndoerfer]

Monte Carlo might sound sophisticated, but it’s really just a clever way to model uncertainty and explore possible outcomes by running the same scenario thousands of times with a bit of randomness thrown in.

Here's an interactive walk through: https://t.co/uT237gOA5i

[Michael Brenndoerfer @mbrenndoerfer]

OpenAI released remote MCP servers. Claude 4 can now alternate between reasoning and tool use. It's more important than ever to understand tool calls and the Model Context Protocol (MCP).

Here's a no-fuzz primer.

https://t.co/70OA6ToN3T

[Michael Brenndoerfer @mbrenndoerfer]

Why aren’t LLMs fully deterministic even with temperature set to 0?

I spoke to experts, but only got vague answers.

Dug in, turns out it’s not just floating point precision! It’s distributed systems, hardware quirks and a lot of other interesting reasons https://t.co/3vKw6FpUdY

[Michael Brenndoerfer @mbrenndoerfer]

@mattpocockuk Regular or non-generic to be explicit with respect to the type

[Michael Brenndoerfer @mbrenndoerfer]

@milan_milanovic How do you define deploying? I assume this doesn’t include all tests, but only the rollout of the new image/ start of server ?