Answer.AI - Practical AI R&D – Answer.AI

Answer.AI is a new kind of AI R&D lab which creates practical end-user products based on foundational research breakthroughs. To learn more about the research agenda and background behind our work, see our launch post and an independent analysis of our lab. See below for a complete list of published articles.

Finally, a Replacement for BERT: Introducing ModernBERT

Blog

This blog post introduces ModernBERT, a family of state-of-the-art encoder-only models representing improvements over older generation encoders across the board, with a 8192 sequence length, better downstream performance and much faster processing.

Benjamin Warner, Antoine Chaffin, Benjamin Clavié, Orion Weller, Oskar Hallström, Said Taghadouini, Alexis Gallagher, Raja Biswas, Faisal Ladhak, Tom Aarsen, Nathan Cooper, Griffin Adams, Johno Whitaker, Jeremy Howard, Iacopo Poli

nbsanity - Share Notebooks as Polished Web Pages in Seconds

Transform your GitHub Jupyter notebooks into beautiful, readable web pages with a single URL change. No setup required.

ShellSage Loves iTerm

This is a quick note on a convenient way to use Nate Cooper’s ShellSage, one of the coolest pieces of tech to come out of AnswerAI recently.

Alexis Gallagher

ShellSage - Your AI Bash Buddy

ShellSage saves sysadmins’ sanity by solving shell script snafus super swiftly

Building an Audience Through Technical Writing: Strategies and Mistakes

Writing

What I’ve seen work and what doesn’t.

A New Chapter for fast.ai: How To Solve It With Code

fast.ai is joining Answer.AI, and we’re announcing a new kind of educational experience, ‘How To Solve It With Code’

How To T̶r̶a̶i̶n̶ Synthesize Your D̶r̶a̶g̶o̶n̶ Data

The art and science of crafting synthetic data for AI training

Concurrency For Starlette Apps (e.g FastAPI / FastHTML)

Concurreny fundamentals for Python web apps

rerankers: A Lightweight Python Library to Unify Ranking Methods

Re-ranking is an integral component of many retrieval pipelines; however, there exist numerous approaches to it, all with different implementation methods. To mitigate this, we propose rerankers, a Python library which provides a simple, easy-to-use interface to all commonly used re-ranking approaches.

Benjamin Clavié

Learn GPU Programming in Your Browser

Blog

WebGPU has arrived, opening a direct pipeline from the web browser to your local GPU. We’ve built WebGPU Puzzles to help you try it and explore the possibilities. It’s a…

Sarah Pan, Austin Huang

/llms.txt—a proposal to provide information to help LLMs use websites

We propose that those interested in providing LLM-friendly content add a /llms.txt file to their site. This is a markdown file that provides brief background information and guidance, along with links to markdown files providing more detailed information.

Small but Mighty: Introducing answerai-colbert-small

Say hello to answerai-colbert-small-v1, a tiny ColBERT model that punches well above its weight.

Benjamin Clavié

How HTMX is Revolutionizing Web Development — Interview with Carson Gross

We recently sat down with Carson Gross, the creator of HTMX, to discuss how it’s changing the web.

Elizabeth Creason

FastHTML: Modern web applications in pure Python

FastHTML is a new way to create modern interactive web apps. It scales down to a 6-line python file, and scales up to complex production apps.

JaColBERTv2.5🇯🇵: Optimising Retrieval Training for Lower-Resources Languages

Introducing JaColBERTv2.5🇯🇵, the new best Japanese retrieval model. Through this release, we present a thorough analysis to better understand what helps in training a good multi-vector retrieval model.

Benjamin Clavié

Cold-Compress 1.0: A Hackable Toolkit for KV-Cache Compression

Blog

Answer.AI is proud to publicly release Cold Compress 1.0, a toolkit for creating and benchmarking state-of-the-art KV cache compression methods. Built on top of GPT-Fast, it…

Griffin Adams, Faisal Ladhak

Faith and Fate: Transformers as fuzzy pattern matchers

Are GPT-like models thinking? Unclear. But the Faith and Fate paper (Dziri, 2023) points out they are often “just” pattern matching.

Alexis Gallagher

gpu.cpp: portable GPU compute for C++ with WebGPU

We’re thrilled to announce the release of gpu.cpp. gpu.cpp is a lightweight, open-source library that makes portable GPU compute with C++ simple.

Politicians can’t change math

It is not possible to provide assurance that a released model can’t be used to cause harm, because a released model can be changed.

A little pooling goes a long way for multi-vector representations

We’re releasing an early version of a simple token pooling trick for ColBERT. This allows for considerable memory&disk footprint reduction with very minimal retrieval…

Benjamin Clavié

A new kind of literate program – Claudette

Claudette is the first ever ‘literate nbdev’ project. Its source is a rendered Notebook with callout notes and tips, HTML tables and images, detailed explanations. It explains how and why the code is written the way it is, along with how to use the libraries it depends on.

Introducing Claudette, a new friend that makes Claude 3.5 Sonnet even nicer

Today, Anthropic launched the most powerful language model available: Claude 3.5 Sonnet. And today, we are making it ever better, with the launch of Claudette.

The definition of ‘Artificial Intelligence’ in SB 1047 was actually meant for systems, not models

SB 1047 defines “AI Model” using a definition designed for “AI Systems”, and as a result fails to cover anything.

What policy makers need to know about AI (and what goes wrong if they don’t)

Policy makers need to understand how AI really works. Otherwise bills like SB 1047 totally fail to achieve their stated goals.

More=Better?

In this post I want to share a few quick experiments to show something that is both obvious and also perhaps underappreciated: you can get a sizeable quality boost from a…

Jonathan Whitaker

SB-1047 will stifle open-source AI and decrease safety

By imposing the restrictions on open-source AI, SB-1047 could reduce AI safety, through reducing transparency, collaboration, diversity, and resilience.

Efficient finetuning of Llama 3 with FSDP QDoRA

We’re releasing FSDP QDoRA, a scalable and memory-efficient method to close the gap between parameter efficient finetuning and full finetuning.

A few tips for working on high-surface-area problems

What do you do when there are too many pieces to fit in your head?

A few tips for working on things that involve a lot of moving pieces, that have a high “surface area”.

Enabling 70B Finetuning on Consumer GPUs

A Technical Deep Dive into FSDP+QLoRA

A detailed guide for adding FSDP and QLoRA support to quantization libraries and training frameworks.

Benjamin Warner, Johno Whitaker, Kerem Turgutlu

You can now train a 70b language model at home

We’re releasing an open source system, based on FSDP and QLoRA, that can train a 70b model on two 24GB GPUs.

Lessons from history’s greatest R&D labs

A historical analysis of what the earliest electrical and great applied R&D labs can teach Answer.AI, and potential pitfalls

A new old kind of R&D lab

Answer.AI will figure out the fundamental research needed to tame AI, and the development path needed to make it useful in practice