King GPT-4 has come, 32k context! The latest news of OpenAI's first developer conference is that the new UI can be customized GPT, and the xAI large model is available on a large scale

Article source: qubits

Editor: Momoko Run

Before the opening of OpenAI's first developer conference, ChatGPT has revealed various revelations, a new UI interface, and everyone can customize GPT, which will lead to the birth of a new profession of "agent engineer". On the other hand, Musk's own xAI large model has also started a large-scale test.

OpenAI's first developer conference has started the countdown!

Remember that on Monday, OpenAI quietly lifted the ban on ChatGPT's two major capabilities: one is to upload PDFs, data documents and other files, and the other is to use all tools without conversion.

No, on the OpenAI developer forum, some netizens have discovered that "GPT-4 (ALL Tools)" has started grayscale testing. And the ChatGPT interface has also undergone different changes.

To put it simply, it is to bring together several individually selected functions of ChatGPT Plus so that it can simultaneously connect to the Internet, call code interpreters, DALL-E 3, and use various plug-ins.

On the other hand, xAI, which is blessed by Boss Ma's traffic, has launched its first AI assistant, "Grok", which has begun a small-scale test.

Netizens on X have posted their "Grok First Experience".

The netizen posted that he had the opportunity to experience Grok through X Premier+, and asked a question about the most popular accounts on X, and the result was that Grok could reply in real time based on the data on X.

32k Context, King GPT-4 (All Tools)

It is worth mentioning that GPT-4 (ALL Tools) is equipped with a 32k context window.

But if we use DALL· E 3, browsers and other tools, only 8k contextual window.

The specifics can be seen from the code diagram:

Netizens who have already turned on the permission started the trial.

Wang Shuyi, an associate professor at Tianjin Normal University, said that GPT-4 (All Tools) is very suitable for use as a teaching assistant after opening up various modes. You can ask it to recognize the slide show and explain it to you, and it can immediately give you examples of your code.

Others said that there are no restrictions on the use of GPT-4 features.

Drawing with a picture is no longer a dream.

Some netizens directly gave tips: find the latest population data of the most populous country, show the value on the board held by people wearing national costumes in the country, and list the source.

ChatGPT then identified India and a population of 1,425,775,850 through "web browsing." HOWEVER, DALL· It took E 3 a few tries to (almost) show the letters on the image.

Here's another example:

Brand new UI, customizable GPT

As above, what we see is that ChatGPT is about to usher in an updated part of its capabilities.

Before the OpenAI Developer Conference started, there were already various styles of news that ChatGPT would usher in a major update, and even the API for developers would be more discounted.

These include:

Brand new UI interface - new ChatGPT prototype Gizmo V8, knowledge deadline April 2023.

Custom chatbot

First up, OpenAI will release a Gizmo tool dedicated to creating, managing, and selecting custom chatbots, with the following features:

• Sandbox testing: Provides an environment for importing, testing, and modifying existing chatbots

• Custom actions: Use the OpenAPI specification to define additional features for your chatbot

• Knowledge File Addition: Add additional files that the chatbot can refer to

• Basic tools: Provides basic tools such as browsing the web and creating images

• Data analytics: View and analyze chatbot usage data

• Draft saving: Save and share drafts for the chatbot you created

• Publish: Publish the finished chatbot

• Sharing: Set up and manage chatbot sharing

• Marketplace: Browse and share chatbots created by other users

「Magic Creator」或「Magic Maker」

It will assist users in defining the chatbot through an interactive interface, identifying user intent, testing in real-time, and modifying the chatbot's behavior through conversational iteration.

• Define chatbots through an interactive interface

• Identify user intent and create a chatbot

• Test the created chatbot in real-time

• Modify chatbot behavior by iterating over dialogs

• Share and deploy chatbots

More detailed information is shown below:

There's also the ultimate toolbox for ChatGPT.

Workspace & Team plans offer new enterprise subscription and workspace management features

The team plan is available for $30 per month and the annual subscription is $25 per month, but requires a minimum of 3 users.

With a minimum monthly cost of $90, it offers additional benefits such as unlimited fast GPT-4 access, longer contexts, and unlimited use of advanced data analysis models.

The workspace feature provides separation between individuals and workspaces and allows for the definition of roles, departments, or other metadata, and is expected to integrate with external ERP and CRM systems.

ChatGPT "Context Connector"

Google Drive and Microsoft 365 can be linked to enable files, tables, or presentations to be attached to or used as context for conversations.

Jim Fan, a senior scientist at NVIDIA, said that I am going to attend OpenAI Developer Day! If the leak is true, it will be a pivotal moment in the AI consumer market:

OpenAI is becoming a full-fledged UGC platform where users can create and share any AI Agents. It will include RPA, role AI, plug-in storage, and much more. The marketplace will sell BEHAVIOR, not software extensions like Wolfram Alpha.

There will be tools to help any user quickly build and test the agent in the sandbox, with access to plugins, browsers, local files, and remote files (via the GDrive/Microsoft connector).

The bot's behavior is not programmed through the SDK, but through a simple iterative chat. I find this particularly interesting because it lowers the barrier to entry considerably.

You don't need to code or write system prompts, but instead ask and refine what you want through conversation. I guess the agent can also ask follow-up and clarification questions because it's being "trained in context". IT'S CONSISTENT WITH THE INTUITIVE INTERFACE OF DALLE·3, SIMILAR TO HOW WE ONBOARD NEW EMPLOYEES AS HUMANS.

Everything seems to be one step closer to Karpath's vision of "GPT as a new operating system". If the market and revenue sharing model really take off, "Agent Devs" will be a new profession, just like "iOS Devs".

There are also big Vs who believe that if the leak of OpenAI's new features is true, "Agents developer" will become a new position after iOS development.

25k context, Grok more internal testing

According to the revelations and summary of a big V on X, so far Grok:

• The size of the context window that supports Super is about 25k tokens
• The response time will be very fast
• Distinctive, funny and witty, far from boring "political correctness"
• The current fine-tuned data source is the 886.03 GB version of the "The Pile" database, as well as the massive data of the entire X platform
• "Real-time" search engine, where data sources are preferentially sourced from X
• API functionality is sure to be available.
• Support voice input prompt words and output replies
• Multimodal forms such as image generation, image recognition, and speech recognition will definitely be supported in the future, and the current model already has some relevant capabilities.
• The lightweight version of Grok can be deployed and run locally on Tesla.

After obtaining the experience qualification, this netizen first had a scolding battle with Grok, and the result was that Grok won a big victory, and Boss Ma personally replied: laugh and cry.

Among the mainstream LLM services that are currently publicly available, it seems that only this one can meet this need.

And most of the melon-eating netizens, Boss Ma's AI assistant that claims to be able to "explore the truth of the world" is still quite optimistic, although no one has used it, but 1/4 of netizens believe that it will become the most leading big prophecy model within half a year.

It seems that netizens still have a "mysterious Thaksin" for xAI, although most netizens are currently unable to experience "Grok", but many netizens are already looking forward to it.

The CEO of a startup company said that as far as the name is concerned, it has already thinned all the current LLMs, and the "epiphany" that Grok represents is very tasteful.

Netizens who have been bored with the political correctness of large models seem to be very much looking forward to getting an AI assistant with strong performance and a less rigid personality.

So, at what cost?

Boss Ma also officially announced the early acquisition method of Grok - subscribing to "X Premium Plus", which only costs $16 per month. However, it is still in the grayscale testing stage, and it will not be fully available to subscribers until the testing is completed.

Musk announced that xAI's Grok AI assistant will be included directly in X Premium Plus, adding that existing X users can spend $16 per month to subscribe.

However, many netizens also expressed their incomprehension of Boss Ma's reversal in fees.

Many users who have spent money to buy the small blue label certification still do not get an early preview of the new feature and ask for an explanation.

Resources: