Open source · MIT licensed · Free to use

Explore the frontier.

A workspace for researchers. Discover what to read next, then read it alongside an AI assistant that knows the whole text.

Free usage every day · No key required to start · Self-host if you like

Share
arXiv:2310.06770 · cs.CL · Oct 2023

SWE-bench: Can Language Models Resolve Real-World GitHub Issues?

JIMENEZ · YANG · WETTIG · YAO · PEI · PRESS · NARASIMHAN

SWE-bench evaluates language models on 2,294 software-engineering problems drawn from real GitHub issuesthread · 4 and corresponding pull requests across 12 popular Python repositories. Given a codebase and an issue, a model is tasked with generating a patch that resolves the described problem. Resolving issues requires understanding and coordinating changes across multiple functions, classes, and files simultaneously.

Our evaluations show that both state-of-the-art proprietary models and our fine-tuned model SWE-Llama can resolve only the simplest issues…

01Discover

Know what to read next.

Describe a topic, a method, an open question. A research agent searches the literature and the open web, reads what it finds, and hands back a ranked shortlist — papers and lab blogs together, each with a one-line reason to care.

speculative decoding for long-context inference
1

Medusa: Simple LLM Inference Acceleration Framework

arXiv:2401.10774

Multiple decoding heads, no draft model — the closest thing to a drop-in answer.

2

EAGLE: Speculative Sampling Requires Rethinking Feature Uncertainty

arXiv:2401.15077

Explains why naive draft–verify breaks down exactly in long-context regimes.

3

How speculative decoding works in vLLM

vllm.ai · blog

The production tradeoffs you'll actually hit, from the team running it at scale.

Toolformer · p. 3 · §2.1

Given just a handful of human-written examples of how an API can be used, we let the model annotate a large language-modeling dataset with potential API calls. We then use a self-supervised loss to determine which of these calls actually help the model predict future tokens.

Thread on this highlight
Isn't this just distillation from the API responses?
Not quite — the filter keeps an API call only if it lowers perplexity on the following tokens Eq. 4, so the supervision signal is the model's own future predictions, not the API output itself.
02Review

Highlight a sentence. Ask about it.

The assistant has the full text in context. Ask for a derivation, pull definitions side-by-side, or highlight any passage to open a thread that stays pinned to it. Every answer cites its source — click the chip, land on the passage.

03Share

Send a colleague your read on a paper.

One link carries the whole bundle — every highlight, every margin note, every chat and deep dive. A colleague imports it into their own workspace and keeps going where you left off.

Share · SWE-bench reviewLink ready
withartifact.com/share/r-7f3e9a2bCopy link
14 highlights6 margin notes3 chat threads2 deep dives
Everything travels
Annotations, notes, chats, and deep dives arrive intact.
Import and continue
One click pulls the review into their own workspace.
Principles

Free, open source,
and yours to fork.

Artifact is built in the open. Read the code, file an issue, send a patch — contributions are very much welcome.

View on GitHub
01

Open source, MIT-licensed

Read the code, run it yourself, fork it, send a patch. Use the hosted version free, or self-host the whole thing.

02

Free usage, every day

Every account includes a daily allowance on the house — sign in and start chatting. No key, no card, no paid tier.

03

Bring your own key

Need more than the free allowance? Add your own OpenRouter key and keep going on any model it offers.

04

Trust every answer

Every answer your assistant gives traces back to the PDF or page. Click a citation chip, land on the passage.

Explore the frontier. Capture what you learn.

Free usage every day · No key required to start · MIT licensed