Last Updated on 1 year by Christian Kaiser

GPT-4
GPT-4-mobile

OpenAI’s GPT-4 Language Model: The Latest on the Latest AI

The latest AI version of OpenAI’s GPT-4 will be able to do more than its predecessor GPT-3, that much we already know. And that makes it exciting. Already today, GPT-3 based text generators can automatically create texts that can hardly be distinguished from human-written content. The best text generators at the moment can be found >>here.

What will be the next step in the development of Artificial Intelligence and how can we potentially use it.

Everything we know so far about GPT-4 (Generative Pre-trained Transformer) is summarized in this article.

OpenAI’s GPT-4 Language Model: The Latest on the Latest AI


When will GPT-4 be released?

There is no official release date for this yet. However, if you look into the past, we should be able to expect the latest version from the GPT series soon.

GPT was introduced to the world by OpenAI on 11/06/2018 through the publication: ‘Improving Language Understanding with Unsupervised Learning’.

Less than a year later, the paper for GPT-2 was published on February 14, 2019.

Then on May 28, 2020, the AI breakthrough with GPT-3 came through this publication.

Consequently, we can assume that a new study will probably be published in 2022 or spring 2023.

The rumor mill is already bubbling, too.

GPT4-Tweet

What improvements have already been released since GPT-3?

07/11/2020 Kicked off commercial use through an API that enabled the development of text generators like Jasper.ai*.

25.03.2021 OpenAI could announce that already more than 300 applications are based on GPT-3 and published further improvements.

10.06.2021 Study on finetuning and related quality improvements (this option was also made available to developers on 14.12.2021 )

09/23/2021 GPT-3 can now merge books

29.10.2021 simple math problems can be solved

12/16/2021 WebGPT as milestone in improvement to have truthful content generated by AI.

Web-Browsing-Model

In 2022, GPT-3 got better at math and as of 03/15/2022, can now add new content to existing sentences and text blocks.

The Best Text Generators and AI Software for Generating SEO Optimized Texts

Jasper_Logo
NeuroflashLogo
Logo-surfer
Logo-Closerscopy

Jasper.ai

Best AI Text Generator Test Winner

Boss fashion from Jasper is ideal for those who expect the best quality.

> Read more

Website*

Neuroflash

Best German AI Text Generator

If you are looking for a German AI text generator with a German interface and German support, Neuroflash is the right choice.

>> learn more

Website*

SurferSEO

Better rankings through AI Onpage Optimization

SurferSEO helps you to understand how the content or your structure should be changed to be in the best position to rank in the Top 10 in Google.

> learn more

Website*

ClosersCopy

Best Lifetime Deal

Qualitatively still a lot of potential but therefore still a super cheap lifetime deal.

No risk no fun.

>> Read more

Website*

How will GPT-4 differ from GPT-3?

GPT-4 facts based on statements made by Sam Altman, CEO of Open AI in September 2021:

  • GPT-4 will be a text-only model and not multimodal like Dall-e
  • Better finetuning as well as better algorithms and higher quality sources should improve the quality of GPT-4
  • GPT-4 will not be built on 100 trillion parameters, over 500 times larger than the current GPT-3 as Andrew Feldman, founder and CEO of Cerebras, the maker of the World’s Largest Computer Chip announced in August 2021. This has already been denied by Sam Altman, CEO of OpenAI.

GPT-4 assessments in the original by Alberto Romero:

  • Optimality: GPT-4 will require more computational power than GPT-3. It will provide new optimization (parameterization) and scaling capabilities (the number of training tokens is as important as the model size).
  • Sparse After GPT-2 and GPT-3, GPT-4 will be a dense model (all variables are used to process a given input). Frugality will become more and more prevalent in the future.
  • Alignment: The new GPT-4 will be more aligned with our values and norms than the current GPT-3. It will build on the findings of InstructGPT, which has been adapted with human feedback. Andrew Feldman, founder and CEO of Cerebras, the maker of the World’s Largest Computer Chip announced in August 2021. This has already been denied by Sam Altman, CEO of OpenAI.

GPT4-Tweet2

GPT-4 Alternatives: These AIs are being developed concurrently:

GPT-NeoX-20B from eleuther.ai

On 20/02/2022, a language model was published by eleuther.ai, named GPT-NeoX-20B, where 20B stands for the 20 billion parameters that make up the AI. This makes the AI the largest publicly available AI already comes very close to the quality of DaVinci from GPT-3.

The new GPT-4 will be more aligned with our values and norms than the current GPT-3. It will build on the findings of InstructGPT, which has been adapted with human feedback. Andrew Feldman, founder and CEO of Cerebras, the maker of the World’s Largest Computer Chip announced in August 2021. This has already been denied by Sam Altman, CEO of OpenAI.

Accuracy-Vergleich

Fairseq 13B from Meta AI

Published on 01/23/2022 Fairseq 13B is an English language model with 13 billion parameters. The model is used by AIs such as NovelAi or Holo AI.

Cohere

Is another platform on which an AI can be built. Unfortunately, the number of parameters is not known here. Since 28.02.22 a beta version“Extremely Large” is available

JI Jumbo from AI21

Or even Jurassic-1 Jumbo from Israeli startup AI21 has 178 billion parameters, making it larger than GPT-3 with 175 billion parameters. In addition, it has a vocabulary of 256,000 words, 200,000 more than GPT-3, and processes text 1.4 times faster. The study shows that JI-Jumbo can generate more information from a wide range of sources, potentially leading to even better text quality.

The development remains exciting.

Open-Pre-trained-Transformer” (OPT) from META

Published on 02.05.2022 Marc Zuckerberg has also entered the AI race with Meta. The model should be accessible to everyone. Meta will offer different sizes of language models from 125 million to 175 billion parameters, which will also differ in the number of layers (from 12 in the smallest to 96 in the largest version).

Summary

I am very excited about the development of GPT-4 and what it might mean for the future of text generators. Although there are few precise statements about GPT-4, the potential impact is enormous. I look forward to following the progress closely and keeping you posted as it becomes available.

Please note:

I make my money through affiliate links, which are highlighted with an *. Should you purchase a product through one of these links, you are not only an absolute hero or heroine, but you are helping me continue to produce informative content that will hopefully help you along. And the best: There are no additional costs for you. I only get a small referral fee from the respective provider.

Many thanks!

Other articles that might interest you:

Improve image quality online with the 3 best AI tools

Jasper AI: Blog up to 10x faster thanks to artificial intelligence

Surfer SEO: 9 reasons why you should use the tool

The Future of Blogging Today: How GPT-3 Artificial Intelligence Will Change Everything

The Best AI Text Generator/ SEO Text Generators 2023

Create explainer video 2022: The best AI to save money and time

Similar Posts