Crates is especially challenging -- the build process is very expensive on all resources (cpu, ram, and disk), and packages are very hard (impossible?) to cache. It works super well on the Pro plan.
It’s communication meant to existing users and it’s a nuanced picture worth elaborating. But if you are not an active user it’s best if you just look at our pricing page for a complete picture: https://replit.com/pricing
Yes, in fact we recently made the free plan more powerful with non-preemptible VMs, premium networking, and package caching for everyone. And working on an AI free tier coming soon. Will likely invest more in the free plan once the abuse from unlimited hosting go away.
Also — the title is slightly misleading. Static hosting is still free, and autoscale (with scale to zero) will effectively cost next to nothing for most users. Especially new programmers get very little traffic that according to our historic data it will cost $0.2 / month.
Broadly finetuning is any post pretraining training. Most of the time it is used in the context of fitting a more narrow task. In our case, it was the same training objective as the pretraining but meant to be more representative of what Replit users like to code. However, we were surprised by how well it boosted overall performance. Best guess: it's a) novel data and b) the model could take even more training!!
The model is not RLHF'd or instructed. It's an inline autocomplete model so it will get confused if you talk it like you're talking to a person. Altho it is possible to finetune it this way. To get better full function completion try giving it the function definition and a descriptive docstring as a prompt.
Can't find it now but pretty sure BigCode said somewhere they explicitly looked for it and removed it. Also subjective measure does match up to the benchmark. Our finetuned model performed +50% on HumanEval and then when using it felt at least that much improved.
1- We trained on languages that are most popular on Replit. Markdown is important because you need some amount of natural language in the data, and it will act as a sort of "natural language label" for code.
2- I like how portable it is being a single small model doing a lot of languages. Single code models are an approach that models like Salesforce/Codegen did that, but I believe we beat (or get very close) to their mono models on benchmarks.
LLMs generally but more so small models will keep going and generate seemingly unrelated things. On the frontend tools like Copilot and Ghostwriter do a lot of things like use stopwords or simply not show completions outside a single block.
As for your prompt, it's following your prompt a little too closely and generating just the function. You can however condition it that this is the start of the program it will do the import, e.g.
# python function that returns a random integer between min and max
import
A lot about this project was surprising. We knew it was going to be good, but didn't expect to be this good -- especially surprising was the finetuned performance boost, and the fact that the model is decent at language tasks and reasoning (in some cases much better than much larger general-purpose models).
It feels like there is a lot more to do with this model, and I have a suspicion you can even make a half-decent chatbot (at least one focused on code) by finetuning it on conversation (and/or instruction) datasets.
Will follow up with a more comprehensive technical report and the UL2R version (fill-in-the-middle support).
Yes we have stories from all over the world where people are using mobile to build real things. There is a kid in Egypt, for example, that built discord bots for a living on his phone.
Additionally there are also lots of cases were fairly well off folks with access to desktop computers and wifi that find themselves with a phone and want to prototype an idea or make a quick change to a project. I’ve been using it and I find it relaxing to lay back on the couch and do some fun coding. Especially with Ghostwriter (our Copilot-like thing: https://blog.replit.com/ai) it’s super usable. I built large part of my toy Lisp in Python in the app while in the park, waiting at the doctor’s office, or simply relaxing after a long day: https://replit.com/@amasad/Lisp-in-Python
We've had Code Explanation (based on Codex)[1] in production for over a month and while it leaves a lot to be desired it was surprising that 75% of users thought the explanation were useful to them.
Perhaps there is some weirdness if you’ve signed up with GitHub. Feel free to email me and we can take a look: [email protected]