👶🤖 Introducing `gpt-engineer`
▸ One prompt generates a codebase
▸ Asks clarifying questions
▸ Generates technical spec
▸ Writes all necessary code
▸ Easy to add your own reasoning steps, modify, and experiment
▸ open source:
▸ Lets you finish a
Introducing gpt-engineer App👶
since gpt-engineer became the world's most popular codegen project I have been tinkering with the next step: how to make it practical, ie allow anyone to build and deploy web–apps with plain english
Mission: Reduce barriers to build
shout to
Introducing gpt-engineer App👶
since gpt-engineer became the world's most popular codegen project I have been tinkering with the next step: how to make it practical, ie allow anyone to build and deploy web–apps with plain english
Mission: Reduce barriers to build
shout to
showcase making a copy of Devin, cloning up-and-coming Open Devin, asking gptme to connect with a frontend by gptengineer App
(i.e. making Devin without code)
Approach:
- screenshot Devin
- frontend + github codespaces via gptengineer App
- clone Open Devin, and gptme for the
Insane. It is possible to train a model that beats GPT3.5 in code generation with only:
- Cloud pod (8 GPUs)
- OpenAI API access (for synthetic data)
- $6.5k to spend
This is a breakthrough for open source models. As well as much smaller models for specialized tasks (the
watched a friend prototype her startup idea with gpt-engineer app
-> jaw dropped
more importantly, inspired her to start selling her idea with prototype with aim to quit her job
...
Early this year LLMs started to show that they can reason. I quickly started looking into what this will lead to. And (being a CTO) how the Engineering role will change.
2 half-weekends of coding + one tweet later...
...and now the open-source project from this has a life of
let's anyone:
> specify what you want
> get a deployed web application
> iterate in plain english
crucially we're including the codegen community, as collaboration makes everyone move faster:
//2
GPT Engineer app (<- different from CLI) lets backend devs
- connect to frontend codebase (on github)
- add openapi of the backend
- ask for complex frontned changes, get instant previews, which use github codebase as single source of truth
example shared in our discord by
gpt-engineer has not fully bootstrapped and generated itself from a prompt, but we are making progress 📈
It now learns with each new prompt it attempts
I just connected gpt engineer to Gemini 1.5
then made it edit its own 50k LoC code base (it is pretty neat!)
going to bed — but if you send an open source repo with a given prompt happy to do a recording of the results ❤️
Happy Valentine’s Day everyone!
And happy birthday Dad ❤️(my dad is awesome)
Launching a new AI startup out of Europe today.
Lovable.
Needless to say, we’re very excited about Lovable. We think it will be huge:
We’re building software that builds software.
See website for
Got super valuable feedback on gpt-engineer App
Listening, and shipping 🚢
top requests:
- uploading pictures+designs (-> now experimental feature)
- non-technical-user <> human dev sync 🎉
- shared projects
- "clarifying" chat flow, asking questions
- full backend generation
Thanks
@chrismessina
for adding gpt-engineer App on product hunt – lots of nice comments there (I just noticed from someone DM:ing 😅)
overwhelmed with positive reception thus far ❤
check out...
@LukeGessler
This sounds too good to be true.
How can “gzip distance” possibly handle adding the word ‘not’ well?
The position of that one word can completely change the meaning of a sentence while minimally affecting the bytes after compression.
GPT Engineer has 17k stars and sees no signs of slowing down 🤯🤖
The community forming, and contributing to the repo, is moving fast building...
The platform for developers to tinker with AI code-generation tooling
@goodside
This sounds too good to be true.
How can “gzip distance” possibly handle adding the word ‘not’ well?
The position of that one word can completely change the meaning of a sentence while minimally affecting the bytes after compression.
I’m running a twitter space on coding agents with awesome
@itamar_mar
(founder codium, behind PR agent + auto-generating testing)
What topics would you enjoy discussed?
🤖👶gpt-engineer updates👶🤖
Improved performance 📈
- TDD workflow: generates the tests, then the code
- Self-reflection + prompts that are even Chain of Thought:ier
Feedback loops 💨
- Spins up the project immediately when done
- Fully autonomous mode to allow for...
Today we decided that the ambition of the gpt-engineer community is: To become the most fun and rewarding open-source community to be a part of.
Anyone interested in helping out?
@swyx
Put the pieces together:
Nov 6 - OpenAI devday, with new features of build-your-own ChatGPT and more
Nov 9 - Microsoft cuts employees off from ChatGPT due to "security concerns" [0]
Nov 15 - OpenAI announce no new ChatGPT plus signups [1]
Nov 17 - OpenAI fire Altman
Put the
We set up the gpt-engineer Discord today.
Humbling to read the kind words and compliments and how many want to contribute and build the community and help each other.
Honored to be part of the community.
We used tooling we built internally to compare how GPT-4 -> Claude 3 affect product KPIs of gpt-engineer App. Evaluating on production data test cases (identical for both).
Compared:
- Cost
- Throughput
- Subjective-quality ELO ranking
- + gpt-4 vision evaluated quality
//1
You can ask Copilot to create workspaces for popular project types with the /createWorkspace slash command. Copilot will first generate a directory structure for your request.
Then click "Create Workspace" and it will create the suggested project - files, directories and all.
yesterday I found that Open Source Campus by
@eraqian
et al is potentially the best soon to open coworking space in SF 🫡
then naturally had an impromptu dinner with a few local agent builders with goal to crack the best agent abstractions
Great conversations and demos
gpt-engineer is in its infant stage.
Good developers could have insane impact – and learn a ton – by taking leadership, facilitate structure, unleash hundreds of passionate coders that want to contribute and get shit done.
Hard work will be acknowledged.
overwhelmed by positive response to gpt-engineer App..
A friend asked for a website that monitors wikipedia
pulls real wikipedia data and displays it
takes 1 prompt to customise UI
sign up for waitlist at ❤
Introducing gpt-engineer App👶
since gpt-engineer became the world's most popular codegen project I have been tinkering with the next step: how to make it practical, ie allow anyone to build and deploy web–apps with plain english
Mission: Reduce barriers to build
shout to
listening to product leaders at OpenAI HQ today
ended with
we were pretty candid, please don’t tweet details about what was said
(much alpha I want to but can’t share !)
showcase making a copy of Devin, cloning up-and-coming Open Devin, asking gptme to connect with a frontend by gptengineer App
(i.e. making Devin without code)
Approach:
- screenshot Devin
- frontend + github codespaces via gptengineer App
- clone Open Devin, and gptme for the
As capabilities continue to accelerate, I forsee that the gpt-engineer community can become the epicenter for collaborative and open "red team"ing new approaches for dangerous AI Agent capabilities.
It's imperative that we continue to focus on benchmarks, and are responsible to
Pull request
#500
was just opened to gpt-engineer 🤖👶
Feels great to have such a community honestly!
So many passionate people who want to build the open platform for code generation together 🫶
We're set on strengthening open source in all of this
I have a whole community and many more to thank for support, but special shoutout to
@FabianHedin
@ErikBjare
more info to come 🤖👶
//8
Been building AI startups for 9 years
Lucky to have experienced fun ups and downs, millions in revenue, and most of all: learnings
3 of my fav. AI startup learnings below
Also – next Wed is Valentine's day ❤
Updates on new initiative then, stay tuned
Big shoutout to
@talrid23
@itamar_mar
@DedyKredo
and team for sharing this! 🌟
Very impressive uplift.
Observed similar reduction in errors in GPT Engineer when we added "self-debugging flow" as we refer to it (happy to land on a single terminology for the same thing)
Amazing people behind this. And we're <1 month after team assembly
So far we had a blast using it to create custom animations, interactive landing pages, christmas themed games
//3
today I learned..
GPT-4 (old one) has much more "liquid" intelligence than GPT-4-turbo
(turbo answers better on average but is brain damaged from the fine-tuning)
my friend is swearing by this being true – his legitimacy is that he has been non-stop prompt+flow engineering
Fully agree. Generalizing too much, makes software bad.
The state of langchain was — partly — what drove me to create gpt-engineer. And this is what most people praised (the simplicity!)
Open source projects that are easy to tinker with and fully understand all pieces of are 🤤
🔮 Prediction:
Finetuning is going to be huge.
Most companies training their own foundation models right now will regret it, since fine-tuning APIs will become so much better.
In the same way that companies that hired tons of Deep Learning engineers for computer vision models
Today:
Customers -> raw data -> analytics products -> enriched data -> custom dashboards -> insights
Future:
Customers -> raw data -> AI -> insights
Natural language interfaces will replace SaaS products that built their moat on UIs, and allowing custom data transformation.
If you missed – 314 Billion Parameter Grok-1 Language Model is now available under the open-source Apache 2.0 license - allowing royalty-free access for commercial and private use.
Is it good? My notes:
The release includes the pre-trained model weights and architecture for
One thing I personally like using the app today.
If you're a dev, you can jump in after rapid prototyping. The code is already backed by git and a familiar tech stack.
//7
I had 2 min to prepare a launch at
@AGIHouseSF
Launchaton
Luckily there is a tool to make an app in 1 min (see below)
Launching AGI Quiz
- learn if you are legit AGI bro
- by answering meme:y questions
🧵
And instead of OpenAI... you can fine-tune open-source models
Already exists via API: , or as open source SDK
Impressive stuff
@corbtt
@DavidCorbitt9
(cost figure below)
🔮 Prediction:
Finetuning is going to be huge.
Most companies training their own foundation models right now will regret it, since fine-tuning APIs will become so much better.
In the same way that companies that hired tons of Deep Learning engineers for computer vision models
Github CEO did a thread of the announcements at Universe 2023, but...
Workspace was not listed, and def the most impressive
Copilot writing specs+planning+coding, browser dev env preview, self-debugging, etc all integrated
got “I didn't think you could have an LLM event of this quality in sthlm”
many technically v. detailed tech demos 😍
travelling other cities next for more comparisons
sthlm arguably has the worlds best
- AI medical note taker
- QA agent
- financial assistant (in its niche)
-
agent hackathon with $30k cash prizes today, you all hackers can still sign up
I will be one of the speakers, discord+twitch (in 30 minutes)
sharing some learnings building gpt-engineer 👶🤖
how fast we’re making progress designing the tech-stack of the generated projects for AI to best understand and maintain them, is very exciting
and, letting the system learn from every mistake to not do it again
//6
Maybe others find this helpful too – if using chatgpt and want to save one's brain from parsing tokens unnecessarily, I found these preprompts make it a better writer and more concise
(first part is to not make it ELI5 when asking about code)
this weeks gpt-4-turbo launch didn't comment how good it was
three days after, openai quitely released simple-evals (open source)
screenshot shows the model is a big step up from previous gpt-4 models
been looking at gpt-4-turbo with the
@lovable_dev
team, thread bellow
@swyx
@ScottWu46
Is this rly unbiased comment?
Tried it first hand too:
One task -> 8h -> doesn’t work in the end
(Same task worked with our more narrow tool in 30 sec)
I’m still impressed that Devin is maximally general but from those I talk to with unrestricted access: no it doesn’t work
WizardCoder 34B outperforms all previous models, open source and proprietary, on human evaluation EXCEPT gpt4.
It does outperform the first version of gpt4 (from April), however 🚀
The model is an instruct fine-tuned version of Code Llama.
Exciting (and scary) with the open source progress and Llama 2 publicly available.
It will spark innovation. And be used by scammers and others.🔥
On how good it is?
Still far away from GPT-4 (86.4) on MMLU: Llama 2 scores 68.9.
I needed an instant execute code from LLM – took me 30 sec to spin it up:)
update from us generally, we've been doing some extremely exciting work towards more general agent abstractions for gptengineer (and we're already starting to use it to ship gptengineer itself)
looking
Thanks for the positive reception on `gpt-engineer`. 500 stars and counting.
First PRs already coming in, but we need help to make GPT Engineer "bootstrap" itself:
Check out the `Project` tab in the repo for PRs that would bring us forward, or comment.
@DrJimFan
The rate of 🤯 in this world is definitely accelerating and will continue to do so
Some more arguments here (apart from that top teams are already preparing data to train 10-100x better models)
gpt-engineer shows that everyday these AI powered software development systems get better and more reliable. I know I showed just a website but functional web apps have also been made with this
Very impressive
@antonosika
This video was made with the not-yet-released Sora AI technology just announced from OpenAi. This changes everything. It's 27 seconds from a text prompt.
Here is their prompt:
Prompt: A white and orange tabby cat is seen happily darting through a dense garden, as if chasing
honestly it's crazy that we can sit down and talk to an LLM about e.g. how the human body works and get 10x more concise (at least with my custom instructions:), 10x faster, and on average more correct answers than from a doctor
what a time to be alive?
Great job
@JustinLin610
@huybery
and team!
Someone should sponsor opendevin and others running on the whole dataset though – 30 sample (=swe-bench lite) can very easily be overfit by accident
(note my claim back in March on this topic:)
If you want to be one of the first to use our product in your team's workflow, enabling anyone (also non-devs) to contribute to a codebase with natural language:
We’re white glove onboarding select developers right now to try it out.
Ping me or just go to if interested:
🤖💡Hey Twitter! Happy to put gpt-engineer to toil away on a hack project.
Ideas? The best submissions will be rewarded with a demo video of gpt-engineer in action.
literally anyone can become
@levelsio
with AI and a lot of ambition:)
this content is for early users of Lovable's closed alpha – OG gptengineer contributor
@kkyvik
shows how to build full web app
Our
@kkyvik
shows you how to build a NomadList clone using
@gpt_engineer
and Google Sheets as a database!
Easy as 1-2-3!
1. Start with data in Sheets
2. Transform it into an API with SheetDB
3. Use GPT Engineer to create a dynamic frontend
🔗
AgentOps and
@antonosika
hosted an invite-only dinner for ███ █ codegen agent engineers.
We brought in YC founders, OpenAI ████, and ████ for ████ ██ + tech demos.
Thesis: If we share learnings, on average we all win and ███ █
Here's what we saw (🧵):
Codgen writeup
RepairLLaMA, new approach for Automated Program Repair (APR). Uses 7B model and outperforms much larger models like GPT-3.5 and GPT-4 for bug-fixing and program repair
//1