Why we won't be replaced by AI
There’s been a lot of talk about AIs replacing programmers recently. The National Post: “AI is coming after the tech bros and their easy money”. Emad Mostaque, CEO of Stability AI: “There Will Be No Programmers in Five Years”. Some people are
Five years ago I created TabNine, the first commercial code completion tool to use deep learning.
Today I'm releasing Supermaven, the first code completion tool with a context window exceeding 100,000 tokens.
Time to share what I've been working on for the past couple months: a visual search engine called Same Energy. It's early stages but you can try the beta here:
I've decided to leave OpenAI to work full-time on creating a visual search engine. I'll miss my coworkers but I'm very excited about what I'm going to make.
Supermaven not only provides the best completions of any tool but it's also the fastest. In the video below I test the speed of Supermaven compared to several alternatives.
Supermaven's response time is 250ms, over 3x faster than any other tool in this case.
Many people asked if Supermaven trains on your code. To be clear: Supermaven does not train on your code.
In the future we might let you get Supermaven at a discount if you allow training on your code. That will always be a clear opt-in: you won't end up there by accident.
Unlike existing tools which are all based on the Transformer architecture, Supermaven uses a new operation which is better than self-attention at integrating information across a long context.
@zack_overflow
There’s a length limit of 400 KB currently. We’re going to raise the length limit or at least add better communication about it. Thanks for giving it a try
This allows us to expand the context window to 300,000 tokens (30 times longer than any other tool) while keeping cost and latency low.
As a user, that means Supermaven will understand APIs and conventions that are unique to your codebase.
Supermaven sees not just your code but also the sequence of edits you've made since your previous commit. That helps it understand your intention when you're making changes.
We've trained Babble, a new model with a 1 million token context window.
Babble is 2.5x larger than the previous Supermaven model and upgrades our context length from 300,000 to 1 million tokens.
We are deploying the new model to all Supermaven users today.
Some early feedback from Supermaven users was that they found the speed distracting.
That confused us at first - we believe faster is always better. Were the users wrong? No.
Say hello to GPT-4o, our new flagship model which can reason across audio, vision, and text in real time:
Text and image input rolling out today in API and ChatGPT with voice and video in the coming weeks.
We've added a chat interface to Supermaven. It's a really clean way to access models like Claude 3.5 Sonnet and GPT-4o while staying in your editor - check out the thread for videos.
Introducing Supermaven Chat: the best way for developers to use GPT-4o, Claude 3.5 Sonnet, and other chat models.
If you've ever copy-pasted code into ChatGPT, then Supermaven Chat is for you. Save time by using Chat to upload your code while staying in your editor.
A nice little thing that I like about Supermaven is using it to bump versions. It's seen enough diffs to know that when you're editing a line with a version number, you usually want to increase it by 1.
A great example of the power of Supermaven's 300k token context.
Left: the completion, right: the context that allowed the model to produce it.
Nowhere in the repo does the phrase "Discord server" appear.
I'm experimenting with a new section on Same Energy for finding art prints you can buy. It's unlisted on the site for now, but you can try it out here:
We're excited to announce support for Jetbrains IDEs!
Thanks to all our beta testers who have helped work out issues with the plugin over the past weeks.
With our update yesterday, we stopped your text from moving around due to suggestions. After this update, a suggestion will never cause your text to move around unexpectedly.
You get the best of both worlds: speed and distraction-free coding.
Supermaven 0.1.23 (VSC) makes a significant change to suggestions in the middle of the line. They will now display off to the right so your text doesn't move around as you're typing:
This is one of many changes we've made to the core systems behind Supermaven in the last few weeks - the product is greatly improved compared to a few weeks ago
We made changes to Supermaven's handling of large diffs that should significantly improve quality and latency when making many changes or adding large files.
on your screen are two clubhouse rooms. in one, strangers are arguing about light vs dark theme; in the other, in-n-out vs shake shack. you try to return to the home screen but it doesn't work. you're trapped
I was playing around with the gradients of GAN loss functions, and ended up finding a view that made GANs feel a lot simpler to me.
I wrote it up in a technical post:
The real problem was that the suggestions displayed in the middle of the line. That meant that because the suggestions were fast, your text moved around a lot in response to the suggestions - very distracting!
I switched to
@SupermavenAI
a few weeks ago and haven't looked back.
Copilot is great but damn is Supermaven just way faster and better contextually at understanding my code.
As soon as I switched I reached out to
@jbfja
and we hopped on a podcast to talk all about AI,
@kevin_jordan__
We'll release some benchmarks on recall next week.
You can use Cursor together with Supermaven. Just disable Copilot++ so they don't conflict
hi, Million (
@milliondotjs
) has $1.3M in gpu credits that expires in a year.
we are looking to fund experiments for:
- determining the most optimal training curriculum, reward modeler, or model merging combination with evolutionary algorithms (any domain ok)
- a diffusion text
@DanielleFong
I made one a couple years ago but never posted it anywhere. Here it is - I have the code to generate it in any resolution if you're interested
Over the weekend, we added a couple changes to make Supermaven more usable on Windows. If you're on Windows and you had issues with Supermaven, it may be worth giving it another try.
The UK has phenomenal AI talent and a long established culture of responsible AI development.
Today I’m proud to be opening a new office: Microsoft AI London. If you’d like to join us, get in touch. We’re hiring!
We’ve developed two neural networks which have learned by associating text and images. CLIP maps images into categories described in text, and DALL-E creates new images, like this, from text.
A step toward systems with deeper understanding of the world.
A novice came to the master as he was finishing his meditation.
"Master, what should I tweet today?" he asked.
The master smiled and said nothing.
At that moment, the novice was enlightened.
Feels like the Capitol had an aura of sanctity about it that was harmed today. The people in Viking hats are gone, but what will people plan next, using today's events as proof it can be done?
We're organizing a meetup in SF next Thursday July 18 at 5:30pm!
Come to meet our team, talk about the future of AI for coding, and give us feedback on Supermaven.
Link below-