Milan Vasarhelyi

That's not a pivot; the product is still the same.

Just some extra AI feature 99% won't use but looks good for marketing.

(AI fields are actually useful btw, for now the app builder is not)

Rockets are something where the smallest failure leads to something very clearly and visibly catastrophic outcome.

So I'm not sure if this is a good analogue with AI. I think smaller "catastrophic" failures aren't really a thing here.

It's pretty interesting for now but it won't be for long.

For sensitive questions its surprisingly open about what it shouldn't talk about in the CoT.

In coding, when it misunderstood my prompt it gave me, I could find out what was unclear in my prompt and correct it.

That's fair

You can usually send a "Sorry we don't have capacity" email though.

I'd argue this is more than that. Because an LLM on its own cannot use tools, so an agent is different in a meaningful way

Today I realized that FancyZones, the built-in tiling window manager of Windows can actually do 90% of what i3 can.

Honestly Windows, at least for dev, came a very long way when I last used it, and I don't regret switching back from Linux a while ago at all.

A lesson for anyone doing development as a freelancer / agency:

If progress is slower than expected, sometimes saying an "I didn't forget you" email to a client is more important than grinding an hour more.

LLM coding errors are usually user errors.

Specifically errors with not having enough context.

If you use a recent/obscure library, you should include the API reference in the prompt. And also the relevant parts of your code.

And then either o1 or Claude will give first-time perfect code 90%.

Cool!

The sad truth is that a model like this is / would be so much behind that no one would actually use it.

The two most important things that LLMs depend on are

1) Massive datasets (where obtaining consent from everyone is already a cost too high)
2) Massive amount of compute

For sure, this is a very hard problem to solve while keeping to the decentralized nature in principle

On the other hand, I trust them to solve hard problems based on how far they have come with this (very difficult) model.

The GIF quality is terrible, it didn't when I chose it 🤷

rick and morty are sitting at a table talking about their purpose ALT: rick and morty are sitting at a table talking about their purpose

When I ask o1 Pro to reformat some text for me

That I could probably do with "find and replace" as well

Yeah that's true. It's a very B2B price for something packaged as a consumer product.

The way everyone searches the web will definitely change over the next few years.

For some users it already did.

The question is: will Google be leading this change or will it become irrelevant?

Every time I have to copy a password from one screen to another, I mix up a capital i with a smal L.

I get the hipocrisy of the govt using them, but wouldn't all employees just register as independent contractors without these rules?

So I feel like some constraints are necessary to be able to give any tax breaks to independents

My favourite one was early Bing AI calling someone a "Bad user" for correcting a mistake

Yeah I should try that out again as well, pretty sure I didn't use the last version.

I copy-paste the full documentation of the framework I use in the beginning of the convo. It's obscure enough that it makes methods up otherwise.

I tried with the API at one point with Omnimodel for better UX than playground, but even with 4o I was above $10 spending per day :D

Don’t always automate processes immediately.

First, take some time to document and optimize them.

If you automate them, you freeze inefficiency.

And you’re missing a big opportunity to improve.

I got the $200 ChatGPT Pro.

But I'm thinking about going back to the $20 next month.

I use it 4 hours a day for writing code, I consider it my main tool, so I don't see the price unreasonable.

But the truth is that right now the benefit of o1 pro over o1 is marginal.

It's benchmarking better, unfortunately that hasn't correlated much with actual use in some cases.

You can access the best AIs for $20/mo plans

But that changes very soon.

As OpenAI launched ChatGPT Pro, for $200/month.

With exclusive access to o1-pro.

Will you be getting it?

Results of ChatGPT on Advent of Code (daily programming quizzes) in 2023 vs 2024, so far.

AI shines in small and isolated problems while humans are still much better at working in a complex system, which is the bulk of SW engineering.

But for tasks like this, AI had insane progress since 2023.

I went from "I don't care about spending four more I hours on my config" to "I want to start working on this project in 5 minutes" as I went from learning to actually making money haha.

"do not, I repeat, ***DO NOT*** use size="lg" in your repeated code as it's not valid in this component library!!!"

I work with no/low-code, would appreciate an add!

Currently mainly working with:
1) Airtable - automated internal systems
2) Voiceflow - Marketing chatbots

Turbo delphi (in high school class 😄) -> emacs -> spacemacs -> neovim -> vscode