prompt engineer @browsercompany | I make apps. My code runs on 100mil+ phones | former iOSdev @Twitter | indie dev @mindGoblinStdio Download Hivemind⬇️
8 subscribers
Jul 13 • 6 tweets • 1 min read
There’s a new type of coder / tech entrepreneur that’s similar to TikTok painters
The final product isn’t necessarily a sellable painting or profitable app, but a piece of media showing how it was made
Turning the creation & documentation of creation, not the output, into the artform
I suspect we see huge huge growth in this area over the coming years
Ai coding tools -> lower barrier to entry & easier to understand, and importantly it makes it much easier to witness -> greater casual viewing interest
Oct 1, 2024 • 23 tweets • 8 min read
OpenAI dev day live tweeting
Let’s see what they cooked!
T minus 90 seconds
The State of Coding with Ai
Tier Lists for Best Base Model & App
(August 2024)
Agree? Disagree?
Anything missing from this list?!?
What would you add?
May 25, 2024 • 36 tweets • 14 min read
Does Devin live up to the hype!?
@cognition_labs new ai developer put to the test!
Follow along as I fuck around and find out
live 1st impressions, reviewing(sorta) & tryin to make stuff:
First things first, getting setup was super simple
Got an email, clicked the link, filled in my info, and im in!
I had poked around some open source devin clones, and didn't bother. Plz don't give me 20 steps to install something. Glad to see this is more of a product
May 14, 2024 • 8 tweets • 2 min read
New ChatGPT mac app hacked!!
The code wizard 🧙♂️ is behind enemy lines
How you can get early access:
Hack is a strong word for
"override the ab test feature flag response"
I would probably not recommend this for the average non-technical person
But its super easy if you know your way around charles or proxyman. All we are doing is flipping a feature flag in their config
Apr 15, 2024 • 5 tweets • 2 min read
Google is building a GPT store!
Vertex ai agent builder, 1st look
Exact same builder Ui, straight up clone
Actions via openAPI schema
Code interpreter!
Notable Improvements:
-Call tools & other agents via system prompt
-example conversations. Mini training data sets
A few more screenshots
Feb 8, 2024 • 4 tweets • 4 min read
First impressions of Google Gemini:
-a lazy coder
Peculiar that both GPT and gemini are running into this.
Gemini seems worse. Particularly because it refuses to fix it and fill in
-double check response, modify response, see drafts, and try again without extensions, are very intriguing and nice touches
-still randomly refuses to do simple tasks...
Absolutely infuriating experience. It can draw a goblin, but a goblin with a sword is impossible apparently. Can't draw images, even though it did that 2 minutes ago
-really really bad at conversation turns.
often ignores instructions, repeats patterns, refuses to do things. Consistently takes things in super broad contexts and basically forgets what we were talking about. After writing incomplete code, I asked for "all code" and it gave me a disclaimer that it cant write all code in the universe...
-extremely disclaimer heavy. more annoying that GPT4 by far. But probably a good idea given their huge customer base.
-being able to access email and docs is nice, but kinda broken because it assumes I have 1 google account. But I have like 5 google emails, and I'm not signing up both my work and personal for 20/mo
-editing your prompt destroys chat history. No forking 2/3 chats to arrow through.
-generate more images button doesnt work half the time
-hate that it pulls in my location. with no option to turn off
-fast, but the load anim is weird. I don't like that it scrolls my message to the top after entering. (tbh I hate the aesthetics of Google's material design in general.)
-left bar makes no sense. Why is there a drop down button when theres plenty of room to show my history?
-image quality is hit or miss. Far behind midjourney, dalle & playground. Especially faces. Gets progressively worse as you ask for changes.
---
Overall:
Decent release,
Feels far behind GPT-4 imo
Uninspired. Nothing inherently new. Playing catch up.
Except perhaps the double check search? which feels like table stakes for GOOGLE.
Gonna keep playing with it, but as it stands after 30 mins of testing, I likely won't sign up for the subscription after the free trial is up.
Might make more sense for people with lots of google docs?
By far my biggest issue is it randomly decides it can't help with things sometimes, when it clearly can. I would much rather it try and fail, then just say "I cant do that"
Trying to get it to write code, and not give commentary is basically impossible
Jan 23, 2024 • 8 tweets • 3 min read
Building gpts tip:
-Don't use knowledge files
They destroy prompt performance and remove's the ai's ability to coherently follow the system prompt
Why?
When you insert docs, the system copies snippets of your docs into the system prompt & context window
So 10 docs. Each around 4000 letters. FLOODS the entire context window
Which means your ai is now too stupid to remember your last message. Less memory than a goldfish
And now the system prompt is 10x as long, so your instructions get ignored
Workarounds & fixes:
Pink knowledge docs are available in the context and via RAG
Blue are only via code interpreter, and work MUCH MUCH better, especially as things scale up
-Use .md files, which are not compatible with knowledge retrieval, and open them by writing code, with code interpreter. This keeps them out of the system prompt context window, only fetching when needed. Dramatically improving overall performance
(Downside here is doc search becomes way worse. Better when you want the full doc RAG'ed in. Not good for summarizing.)
-DO NOT USE ALL 20 FILES, or if you do. DO NOT MAKE THEM LARGE. I had 10 files at ~4000 characters which destroyed all multi step reasoning. It would literally forget and repeat the first message in the conversation over and over again. Swapping from .txt to .md instantly fixed this
-Only include docs which include working system prompts for the ai. Do not simply chuck a doc in here and expect it work
More info on supported file types: platform.openai.com/docs/assistant…
Jan 4, 2024 • 4 tweets • 1 min read
This is not a real person
This not real handwriting
Introducing
Evolution Chamber
A custom GPT to mutate your custom GPTs
-Build GPT actions easily
-OpenAPI schema generator, specific for GPTs quirks
-Validation tools with better error messages
-A template to build your own backend server to handle actions
Advanced prompting ai dev trick w/ @cursor_ai
How to make Ai self-talk debug flows
1. make a custom debug logger, with toggle-able log levels. Example in swift 2. tell Cursor's Rules for Ai to use your logger 3. Generate code. Error messages are created & auto sorted
If your code breaks
comment/uncomment the shown log levels
to make a debuggable prompt based
"stack trace"
made from 100s of ai print line breadcrumbs
which you can then paste back into the chat
making ai give debug hints to itself
You can do this in ~5-15 minutes max
To refactor your codebase for a never before seen type of RAG-lite based debugging (??? idk what to call this)
Instead of ai deciding what code snippets are loaded into the prompt context
you do
the old school way
lots of debug print lines
Aug 12, 2023 • 15 tweets • 6 min read
State of the art Ai coding stack:
Cursor
VSCode fork. Best Ai editor by miles
100x GitHub copilot
Aider chat
Best agent for actual dev work in existing codebases