march 2025

ai skeptic

Why am I skeptical about generative AI?

1. catalyst for doubt

In mid-March, I read a coherent, convincing take-down of Generative AI. It’s titled There is no AI revolution, written by this guy Ed Zitron.

Despite the colourful language and general slander, the point he makes still resonated with me. It boils down to:

Is this whole thing overblown?!

I had been harbouring feelings of doubt for a while, but this article put a fine point on it. Also, the timing felt especially pertinent since:

• the company I work for is very confident about ai.
• much of the stock market (that I invest in) seems certain about radical advancements in generative ai

Unfortunately, I don’t feel as confident. The following is an effort to figure out why.

2. how do i use ai?

One reason I might not feel as optimistic about AI is because I’m using it wrong.

Here’s what I do. It mostly fits into two categories:

a. A pseudo-Google search

GPT has virtually replaced Google in my daily workflow. Things I used to find by clicking links and reading StackOverflow I find through the chatbot now.

It’s nice because it doesn’t care about spelling or grammar and it can recall context from previous messages in the thread. It’s very good at understanding what I want.

• e.g., “is map faster then forEach in JS” -> “how can I filter some items from the array?”

b. Copy editing

The majority of my published writing gets checked by GPT.

It will notice illogical sentences and disconnected ideas. It can distill down important points well and offer coherent, useful suggestions.

It’s simply very good at understanding my writing.

• e.g., “go over this github reply”
• e.g., “are there any mistakes in the following text?”
• e.g., “summarize these paragraphs”

I mean, I like being able to talk to the computer in English! It’s faster than Googling something and much less effort than asking a human to read my writing.

It actually seems to understand meaning in language.

However…

3. what is it missing?

It’s missing opinions.

I rarely ask it for advice. When I do, I get back a generic, lowest-common-denominator opinion. Its opinion is the internet’s opinion.

As in, it gives me the highest-ranked-website-on-Google’s answer. Because it’s so good at generalizing! But this is not always useful.

It’s missing a memory.

It doesn’t seem to remember things. Past a certain point, it will forget what we’re talking about or go into a loop. The deeper I get into a thread, the weirder it gets.

This incentivizes starting from scratch for each interaction, which makes long discussions nearly impossible.

It can’t do math.

Well, it can do simple math. But it doesn’t understand math. Apple did a study—when a word is changed in the math problem, it fails up to 65% more often. That can’t be right! I don’t get 65% worse when the “apples” turn into “oranges”!

4. it’s an interface

So what should we do with a machine that can understand language, has the internet’s opinions but can’t remember things and can’t do math?

Well, like I said 2 years ago, use it as an interface around software.

Instead of using programming languages or interfaces or buttons or Excel or a UI…just use English!

This article puts it pretty well. You want the LLM on the outside as the interface. This relates quite directly to MCPs.

Like yes! Give me a chatbar in Photoshop. Not to generate creepy fake images, but to create a masking layer around an object automatically. Or to add a percentage change column to my Excel spreadsheet. Do the clicking around for me while I speak in English. Like Jarvis! Right?

So sure, there are myriad ways to monetize this. Speaking English to software is indeed a paradigm shift.

But, unfortunately, I don’t think this is the path to AGI. From this article:

Everyone now seems to be admitting you can’t just use more compute and more data while pretraining large language models and expect them to turn into some sort of all-knowing digital god.

I don’t deny that there could be another path out there to AGI, but today’s models aren’t it.

misc

Silly me spent all my time writing about AI 🤦♂️. I’ll try not to do that again for a while.

Here are some other things that happened in March:

• Ashley and I hosted an Oscars watch party
• tried freediving for the first (and second) time and loved it
• I attended the helicopter logging safety forum
• got into hifi audio and I set up my first system
• decided I want a Mazda CX-5
• decided I actually don’t want a Mazda CX-5
• started running regularly again
  • noticed my feet going numb after 5km
• hit a new squat PR! (245x6)

end

thank you kindly! Bye 👽

view on substack

← all issues