gabby

Literally 100% of the recruiting emails I get are for AI companies. 100% Get a real product idea, folks!

Equally harmful: pressuring someone to only spend five minutes reviewing the same code The common failure mode in both cases is micromanaging how engineers do their jobs. Engineers have natural checks and balances against misbehavior which are steamrolled when management sets blanket expectations

For example, if someone asks me to review their slop PR the natural check on that is that I can spread out my review over a longer period of time to protect my time and to discourage that sort of behavior However, if my manager pressures me to review on a timely schedule then that check is lost

I sure am! In fact, I've been in there since at least GPT-4. I actually have a fun story to share about this So, back in the day, I had a friend who had access to gpt-4-base because she and some others were contracted by OpenAI to red team the model. She invited me to a party in the Bay Area …

There also wasn't any hidden metadata either (I was using someone else's laptop to run this experiment) That means that gpt-4-base had fingerprinted me as the author purely based on the subject matter and writing style. Although, let's be honest, who else would write a blog post with that title?

I sure am! In fact, I've been in there since at least GPT-4. I actually have a fun story to share about this So, back in the day, I had a friend who had access to gpt-4-base because she and some others were contracted by OpenAI to red team the model. She invited me to a party in the Bay Area …

So with Loom's help I begin drafting a blog post I had on the backburner for a long time (and still do; never got around to finishing it). The working title of the post I began with was "The fundamental difference between imperative and functional programming"

I sure am! In fact, I've been in there since at least GPT-4. I actually have a fun story to share about this So, back in the day, I had a friend who had access to gpt-4-base because she and some others were contracted by OpenAI to red team the model. She invited me to a party in the Bay Area …

If you're less familiar with LLMs, gpt-4-base is what's known as a "base model" meaning that it has no safeguards in place and also hasn't been tuned to act in appropriate ways So, needless to say, they were getting high and drunk and doing irresponsible shit with the base model at this party

To understand why I’m so down on agentic coding I should explain what I think LLMs are good at LLMs are good if you want *consistency* because they are statistical models with extra steps. If you ask them what comes next they’ll bias to what came before … but consistency is *BAD* for programming

Recently I've begun thinking about the "derivative" of LLMs, meaning how changes to the input produce changes to the output. Specifically, even if an LLM were deterministic, that doesn't necessarily mean that you can easily reason about how changes to the input produce changes to the output

However, even if you accept chat as the source of truth, you still have the same problem: the lack of an intelligible derivative. You can't easily reason about how a change to the input (a new conversation turn) will translate to a change to the output (the code)

Recently I've begun thinking about the "derivative" of LLMs, meaning how changes to the input produce changes to the output. Specifically, even if an LLM were deterministic, that doesn't necessarily mean that you can easily reason about how changes to the input produce changes to the output

Now suppose you want to make changes. Would you modify the original prompt and ask the LLM to rebuild the program from scratch? Probably not. You might have grown attached to specific behaviors that were underspecified by the original prompt and you might not know how to capture those behaviors

Recently I've begun thinking about the "derivative" of LLMs, meaning how changes to the input produce changes to the output. Specifically, even if an LLM were deterministic, that doesn't necessarily mean that you can easily reason about how changes to the input produce changes to the output

Compare with coding, where I can in principle build a mental model of how a change to my code leads to a change in the program's behavior Sometimes a small change to my code can produce a very large change to my program behavior, but these rules are formal and learnable