Building an AI product is hard. Engineers who understand AI are expensive and difficult to find. And there's no way of telling who's legit and who's not.
That's why companies around the world trust AE Studio. They'll help you craft and implement the optimal Al solution for your business with a team of world class Al experts.
OpenAI formed a Safety and Security Committee after announcing the training of its new foundation model. This committee will be tasked with issuing recommendations to the board about actions to take as model capabilities continue to improve.
Jan Leike, a former OpenAI researcher who resigned over AI safety concerns, has joined Anthropic to lead a new "superalignment" team focusing on AI safety and security. Leike's team will address scalable oversight, weak-to-strong generalization, and automated alignment research.
Microsoft unveiled new features for Copilot at Build 2024, including Team Copilot for team collaboration, custom AI Agents to automate workflows, and Copilot Extensions and Connectors for easy customization. These enhancements aim to improve productivity and business process efficiency. The updates are currently in limited private preview, with broader availability expected later in 2024.
Scaling laws are a way to predict how well models will perform at certain sizes with a certain amount of data. They are expensive to obtain. This paper explores using gzip compression ratio as a strong signal to predict a data-dependent scaling law.
Researchers have improved transformer performance on arithmetic tasks by adding embeddings that encode each digit's position relative to the start of the number.
A few weeks ago, a new optimizer from Meta made the rounds as a promising alternative to Adam. This paper outlines the method in more detail, including the online updates portion of the method. In general, it seems like a promising result, especially when the total number of desired training steps is unknown at the beginning of training.
Andrej Karpathy has released an update to LLM C, a single self contained GPT-2 implementation aimed at reproducing the 2019 suite of models. With this most recent update, the library is able to train the smallest of these models in just 90 minutes. It runs end to end and has minimal dependencies.
One of the best depth models is now available as a pipeline in diffusers. This tutorial discusses how to use the model, what you can do with it, and a trick to make it work seamlessly with videos by conditioning on the first frame's latents.
Key members of OpenAI's safety team, including leaders of the superalignment team tasked with ensuring AI alignment, have departed partly due to lost faith in CEO Sam Altman. These departures follow internal conflicts and concerns about OpenAI's safety priorities and the company's direction under Altman's leadership. OpenAI is updating its offboarding documents after criticisms regarding potential equity cancellations for those who refuse to sign non-disparagement agreements.
This article discusses different approaches to AI regulation, contrasting model-level, use-level, and conduct-level frameworks. It argues that conduct-level regulation, which applies existing laws to new technology without excessive specificity, is preferable to use-level regulation, which could result in impractical restrictions and unnecessary complications for AI deployment. The recent EU AI Act's restrictions on AI's ability to infer emotions is cited as an example of the pitfalls of a use-level approach.