Shulou(Shulou.com)11/24 Report--

[guide to Xin Zhiyuan] in just 3 days, the world ushered in the outbreak of GPT applications, and the popular style of the times has taken shape.

A new GPT is born in 1 minute!

In less than a week, all kinds of customized GPT broke out all over the world, and the growth rate has completely exceeded everyone's imagination.

This precisely confirms what Altman said at the developer conference, "We are breeding new species, and they are proliferating rapidly." "

The whole network is pessimistic about Agent startups, and even the research direction of developers is about to change dramatically.

On the other hand, a Karpathy tweet added fuel to the flames and exposed OpenAI's ambition to make the GPT model the operating system for large models.

In a teasing way, he also gave the specific "technical specifications" of the large model OS:

LLM:OpenAI GPT-4 Turbo 256core (batch size) processor @ 20Hz (token / s)

RAM:128K token

File system: Ada002

Incidentally, Sam Altman has just announced that the new version of GPT-4 Turbo has also been launched on ChatGPT, which means that the 128k context can be used by everyone.

After the opening of GPTs, people's creative enthusiasm is instantly aroused, and its potential lies in the outburst of genius creativity, this form of UGC will always appear the fashion of the times.

The most complete series of customized GPT summed up that ChatGPT is actually just a trial version, and GPTs is the real AI product launched by OpenAI.

After the release of GPT-4, it took only a little more than half a year for OpenAI to become the world's best provider of AI applications from the best provider of basic large models in the world.

Now the number of GPTs applications on various platforms on the Internet has reached 3000 million!

In GPTs, the most creative people in various industries in the world are using their own experience and data to help OpenAI develop AI applications.

From the accurate information service in the professional field to the question and answer of life.

As for the powerful assistant of academic research, the GPTs which appeared in just 3 days covers almost every field you can imagine.

Unexpectedly, the famous saying of AI boss, "all applications need to be reconstructed by AI", has already happened in just 3 days.

We chose the most popular GPTs on the Internet to show you how deep the rabbit hole is!

C-RPG games generate GPT a GPT application called Retro Adventures. It can make a pixel-style C-RPG word game for you on the spot.

Just give him a specific theme and it can generate a specific game. We chose "20,000 miles under the sea", which is the beginning of its generation:

After introducing the background of the game through a picture and a paragraph of text, the game is finished. Play is very much like C-RPG or mud games, through dialogue to experience the game content.

After choosing to drive a submarine to explore mysterious waters, the game continues:

Players can choose one of the three options generated each time to promote the game, or they can directly enter what they want to do. Retro Adventures can generate very interesting game content for you.

This time, instead of following its advice, I chose to push the game according to my own ideas and replied: I found a hot pot restaurant at the bottom of the sea and went in to try.

As a result, it can really follow my way of thinking and take me in to experience what it is like to eat hot pot for 20,000 miles under the sea!

Then we re-opened a "Westworld" and the game experience was very interesting.

Another dream sticker maker is ChatGPT's official "Dream sticker" GPT, which allows users to turn the contents of their dreams into usable stickers and deliver them directly to their homes!

According to the introduction, we created a very "wild" dream:

The ideal boyfriend came to the famous blogger @ dotey teacher to answer the call of the majority of netizens and became a "boyfriend GPT".

This boyfriend is a man with a strong tolerance.

Features: humor does not lose depth, innocence does not lose discretion, idealism does not lose simplicity, the most important thing is to always maintain stubbornness and kindness in front of fate.

Https://chat.openai.com/g/g-IlNu7BVYQ-your-boyfriend-alex/c/b9d330d5-a4e5-4b76-a0ba-7a7690446f75

All come to test whether this boyfriend is qualified or not.

GPT, Buy Buy No worries engineer Quinn Leng from Databricks made a "chopped hand GPT".

Its strength lies in that any item can be automatically identified, compared with the price of the whole network, and can also provide purchase advice.

For example, you can enter a picture of fitness equipment, and then "chopping hands GPT" will open a web-wide search to sum up a specific list for you.

Scientific research weapon, AI paper embellishment expert online "AI Paper Polisher Pro" is simply an AI scientific research weapon.

It can provide direct and clear suggestions for improving AI conference papers, focusing on the structure of the paper, technical accuracy and visual elements of the LaTeX code.

It can even analyze screenshots of the paper and provide feedback at different levels, including the overall layout and structure, as well as detailed writing suggestions.

Some netizens said, this is what I want!

There are too many GPTs, and a domestic developer said that thousands of GPTs have sprung up, and I really don't know how to start them, so I have built a special GPT that "recommends GPT".

Https://chat.openai.com/g/g-iD7sLuO9S-gptofgpts

For example, if you want to learn English, don't bother to find it, just ask "GPTofGPTs".

The Chinese Little Chef God generates all kinds of classic Chinese food with one click.

GPT address: https://chat.openai.com/ g / g-VrmZ5hQPP-zhong-hua-xiao-chu-shen

In addition to telling you how to cook, I can also recommend various dishes to you according to your lifestyle.

Want to be your own GPT, but not? Let GPT teach you.

GPT address: https://chat.openai.com/ g / g-Z0f6pPPGc-gpt-builder-builder

Master of 3D printing, 5 minutes entry 3D printing

GPT address: https://chat.openai.com/ g / g-W0lCzVAZH-3d-print-master

After a simple search based on the content, we found that the four recommended models are indeed quite reliable.

GPT-4 Turbo performance card on the first developer Day of 73K, Altman upgraded GPT-4 API to GPT-4 Turbo, updated to April 2023. Most importantly, the maximum length of the context is 128k, equivalent to 300 pages of text.

Its price is about 1Compact 3 of GPT4, and its speed has soared almost fourfold.

How fast is it? feel it intuitively.

But what is the real performance of GPT-4 128K? an engineer tested it under pressure.

It was found that the recall performance of GPT-4 began to decline when prompt exceeded 73K token.

The performance of the recall is related to the location of the recall fact, and the performance is low when it is between 7% and 50%. If the fact is at the beginning of the document, it is recalled regardless of the length of the context.

This means that GPT-4 retrieval of facts can not guarantee 100% accuracy, do not assume that it will retrieve all the facts.

So what's a good solution?

As we all know, less context = higher accuracy. Providing less context input can improve the accuracy of GPT-4 recall facts.

In addition, the location of the facts in the document is also important. Adding the facts to be recalled at the beginning and bottom of the document makes it easier for the model to recall.

The engineer Greg Kamradt's test process is as follows, using Paul Graham's 218 articles as a "background" token.

Then, put a random sentence at different depths of the document, actually using "the happiest thing in San Francisco is sitting in Dolores Park eating sandwiches on a sunny day."

Ask GPT-4 to answer this question using only the context provided, and use another LangChainAI evals model (GPT-4) to evaluate GPT-4s 's answer.

Next, repeat this process for 15 document depths (from the top of the 0% document to the bottom of the 100% document) and 15 context lengths (from 1K to 128Ktoken).

Two tests were run under a larger context length to achieve higher performance.

Finally, Kamradt said, "while I think this is the right direction, more testing is needed to gain a clearer grasp of GPT4's capabilities."

The cost of this test is about $200 (1457 yuan), and the cost of a single call to 128K input token is $1.28.

The API call cost for this test is ~ 200USD.

The source code of this test project has been open source, interested can be hands-on testing.

Https://github.com/gkamradt/LLMTest_NeedleInAHaystack

Does GPT-5 turn on grayscale test? Some netizens have found that OpenAI has begun to test two new functions in grayscale: magic creation (Magic create) and Gizmo.

The name of the model Gizmo, which is no stranger to many people, is the legendary GPT-5. Its training period ends in January 2022.

What do you think?

Reference:

Https://twitter.com/rickawsb/status/1723217817531801992

Https://twitter.com/dotey

Https://twitter.com/lencx_/status/1723226804671135958

This article comes from the official account of Wechat: Xin Zhiyuan (ID:AI_era)

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.