#149 - Reflecting on 2023, Midjourney v6, Anthropic Revenue, Unified-IO 2, NY Times sues OpenAI

6 snips

Jan 7, 2024

Reflecting on the pivotal changes in AI during 2023, the hosts highlight the rise of open-source alternatives and shifting safety concerns. Midjourney version 6 impresses with its text-to-image capabilities, though some spelling challenges remain. The competitive landscape heats up as Baidu's ErnieBot gains traction against ChatGPT. Anthropic forecasts substantial revenue growth, while Microsoft's new Copilot app for Android makes waves. Plus, semiconductor innovations showcase the ongoing tech race amidst regulatory hurdles.

Ask episode

AI Snips

Chapters

Books

Transcript

Episode notes

ANECDOTE

Listener Correction on FunSearch

Jeremy received an email from Chris Cox, CIO of the Bank for International Settlements, correcting a nuance in his explanation of DeepMind's FunSearch.
Cox, with a background in evolutionary algorithms, clarified that new information in FunSearch also comes from mixing and matching cached functions.

INSIGHT

Mainstreaming of AI

2023 witnessed the mainstreaming of AI safety, self-driving cars, and AGI timelines.
Short AGI timelines, once ridiculed, are now discussed seriously, with frontier labs predicting AGI within 5-10 years.

INSIGHT

AGI Timelines Shorten

AGI timelines have shortened significantly, with some predicting it within the next five years.
This shift brings AGI discussions into the mainstream, highlighting the increasing relevance of AI in public discourse.

Get the Snipd Podcast app to discover more snips from this episode

Get the app

Improving Text Embeddings with Large Language Models

Linjun Yang

Nan Yang

Furu Wei

Liang Wang

Xiaolong Huang

Rangan Majumder

This paper introduces a novel method for enhancing text embeddings by leveraging large language models and synthetic data. The approach involves generating diverse synthetic data and fine-tuning open-source language models to achieve high-quality text embeddings without relying on labeled data. The method demonstrates strong performance on competitive benchmarks and sets new state-of-the-art results when combined with labeled data.

Unified-IO 2: Scaling Autoregressive Multimodal Models with Vision, Language, Audio, and Action

Derek Hoiem

Christopher Clark

Ryan Marten

Sangho Lee

Jiasen Lu

Aniruddha Kembhavi

Zichen Zhang

Savya Khosla

Unified-IO 2 is a groundbreaking autoregressive multimodal model that integrates vision, language, audio, and action into a unified framework. It uses a single encoder-decoder transformer model to process diverse inputs and outputs, achieving state-of-the-art performance on several benchmarks. The model is trained from scratch on a large multimodal dataset and fine-tuned on over 120 datasets to enhance its capabilities.

DeWave: Discrete Encoding of EEG Waves for EEG to Text Translation

A Novel Framework for Open-Vocabulary EEG-to-Text Translation

Yiqun Duan

Zhen Wang

Yu-Kai Wang

Charles Zhou

Chin-Teng Lin

DeWave is a pioneering framework that uses discrete encoding sequences to translate EEG waves into text. It integrates a quantized variational encoder with pre-trained language models to address issues like individual differences in EEG signals and order mismatches between eye fixations and spoken words. This approach enables the translation of entire EEG signal periods without word-level markers.

TinyGPT-V: Efficient Multimodal Large Language Model via Small Backbones

A Novel Open-Source Multimodal Large Language Model

Lichao Sun

Zhengqing Yuan

Zhaoxu Li

TinyGPT-V is a novel open-source model that integrates a compact language backbone with pre-trained vision modules, requiring minimal computational resources for training and inference. It is designed for tasks like image captioning and visual question answering, making it suitable for devices with limited resources.

Task Contamination: Language Models May Not Be Few-Shot Anymore

Jeffrey Flanigan

Changmao Li

This paper investigates the impact of task contamination on the performance of large language models (LLMs) in zero-shot and few-shot learning scenarios. It highlights that LLMs often perform better on datasets released before their training data creation date, indicating task contamination. The study employs methods like training data inspection and membership inference to detect contamination.