Revolutionizing AI: OpenAI’s GPT Builder Transforms the Tech World Overnight

Explore the seismic shift in AI with OpenAI’s GPT Builder. Dive into the details of the GPT-4 Turbo, the pioneering GPT Store, and the groundbreaking Assistants API that’s redefining AI development.

In a mere 45 minutes, OpenAI managed to captivate the AI industry once more, potentially causing a night of restless excitement for AI professionals around the globe. November 6, local time, the OpenAI Developer Conference kicked off with founder Sam Altman and colleagues taking the stage. In just 45 minutes, they unveiled the team’s latest breakthrough, the GPT-4 Turbo. This new iteration isn’t just faster and equipped with longer context capabilities, but it also boasts enhanced control features.

Simultaneously, OpenAI announced a nearly threefold reduction in API pricing, dropping it to 1,000 inputs per penny, which was met with roaring applause from developers in attendance. But perhaps even more significant was the launch of “GPTs” – a new offering that enables people to build customized GPT versions using natural language. And you guessed it – these can then be uploaded to the soon-to-be-launched “GPT Store”!

If the GPT-4 Turbo is the improved “iPhone” of AI, the GPT Store may well be the pivotal step in positioning OpenAI as an “Apple-like” titan in the industry. While competitors are still tinkering with “AI alchemy,” OpenAI has begun constructing what appears to be a grand ecosystem.

01: GPT-4 Turbo Unleashed: The Speed and Cost-Efficiency Breakthrough in AI

The conference opened with Sam Altman announcing a major upgrade to GPT-4, introducing GPT-4 Turbo, simultaneously released for both ChatGPT and the API versions. Altman shared that the team has been keenly collecting developer feedback, leading to six significant upgrades addressing developers’ concerns: longer context lengths, stronger controls, knowledge model updates, multimodality, model fine-tuning customization, and higher throughput limits.

The first four upgrades primarily enhance the new model’s performance, while the last two are targeted solutions for enterprise developers’ pain points. Alongside performance improvements, OpenAI also announced a substantial price reduction for its API, a move akin to offering more for the same price.

Among the six upgrades, the first is context length.

OpenAI’s previously offered maximum context length was 32k, but GPT-4 Turbo has now expanded this to 128k, surpassing competitor Anthropic’s 100k context length. To give you an idea, a 128k context length is roughly equivalent to the amount of text covered in 300 pages of a standard-sized book. Beyond accommodating longer contexts, Sam also noted that the new model maintains higher coherence and accuracy within these extended narratives.

Secondly, the update provides developers with stronger control mechanisms for better API and function calls.

The new model introduces a JSON Mode, ensuring model responses in a specific JSON format for more convenient API interaction. Additionally, the new model allows for simultaneous multiple function calls and introduces a seed parameter to ensure consistent model outputs when needed. Over the coming weeks, the model is also set to introduce new features, including visibility into log probabilities.

Third is the update to both the internal and external knowledge bases of the model.

Approximately a year after ChatGPT’s launch, GPT’s knowledge base has finally been updated to April 2023. Sam Altman has committed to ongoing updates to the knowledge base to prevent it from becoming outdated. “We share your frustration – even more than you do – about GPT’s knowledge being stuck in 2021,” Altman remarked.

In addition to the internal knowledge base upgrade, GPT-4 Turbo has also improved how it updates external knowledge bases. It now supports uploading external databases or files to supplement GPT-4 Turbo’s external knowledge.

Fourth, and perhaps least surprisingly, is multimodality.

The new model supports OpenAI’s visual model DALL·E 3 and has introduced a new text-to-speech model – developers can choose from six preset voices to find the one that suits their needs.

GPT-4 Turbo can now create images from text. On the topic of image queries, OpenAI has introduced a safety system to prevent misuse. OpenAI also announced that it would cover the legal costs related to copyright issues for all customers.

Within the voice system, OpenAI claims its voice model far surpasses similar market offerings and has announced the release of the open-source voice recognition model Whisper V3.

Fifth is the model fine-tuning and customization.

In August, OpenAI had launched the fine-tuning service for GPT-3.5 Turbo. Early tests indicated that the fine-tuned version of GPT-3.5 Turbo could surpass GPT-4 in certain tasks, albeit with a relatively high price tag.

This time, Sam announced that the GPT-3.5 Turbo 16k version is also available for fine-tuning customization, with pricing set lower than the previous generation. The fine-tuning customization for GPT-4 is also currently under application.

Moreover, OpenAI has started accepting individual enterprise customizations of the model. “This includes modifying every step of the model training process, conducting additional pre-training for specific domains, and post-training, etc.,” Altman stated. He also mentioned that OpenAI could not undertake many such customizations and that they would not come cheap.

The last of the six upgrades is a higher throughput limit.

GPT-4 users could immediately enjoy a doubled rate limit per minute after the conference. Additionally, if unsatisfied, further rate limit increases could be requested through the API account.

Beyond the six upgrades is an across-the-board price reduction for the API system. The newly released GPT-4 Turbo has seen input prices slashed to one-third of GPT-4’s cost and output prices reduced by half, with OpenAI stating that overall usage prices have been cut by approximately 2.75 times.

The new model is priced at one cent per thousand input tokens and three cents per thousand output tokens. The price reduction was welcomed with cheers from developers onsite.

Sam also mentioned that after addressing pricing, the next focus will be on speed issues, and developers will soon find that GPT-4 Turbo will be much faster.

02: GPT Store Launches: A New Era of AI Customization and Accessibility Begins

Back in May, OpenAI had already rolled out a plugin system, debuting with 70 applications related to their large-scale models, encompassing domains such as word guessing, translation, and stock data retrieval. The style of the GPT Store bears a striking resemblance to the “App Store” – a design choice by OpenAI that signifies its ambitions.

At that time, the feature was met with high expectations, with many in the media comparing it to the launch of Apple’s App Store, believing it would transform the ecosystem for large model applications. However, despite an increase in plugins, the system did not achieve the same level of impact as the Apple App Store.

At this conference, OpenAI redefined its app store framework, expanding it into a new realm where anyone can create AI Agents based on their knowledge base using natural language. These can then be integrated into the OpenAI app store and earn revenue shares. The applications released by OpenAI are no longer referred to as plugins but have taken on a somewhat unusual name: GPT. The overall app store, named GPT Store, is set to officially launch later this month.

According to Sam Altman, each GPT is a customized version of ChatGPT designed for a specific purpose. To highlight the new GPT applications, there will be slight adjustments to the ChatGPT page. Beneath the ChatGPT on the top left, the applications released this time will be featured.

During the demo, complex plugins, such as Zapier, which was among the first batch when OpenAI launched its plugin system, remain present in the app store and may continue to be a significant part of the future ecosystem. Jessica Shay from OpenAI used Zapier to link her calendar and text messages, arranging her schedule and notifying colleagues directly by chatting with the Zapier application.

However, despite Zapier’s robust capabilities, such applications were not the focus of this release. Data from Glassdoor indicates that Zapier employs between 500-1000 people, and according to Fortune, Zapier is valued at 5 billion dollars. Relying on such applications to enrich OpenAI’s still-developing app store to create a vibrant ecosystem is not very realistic.

Therefore, in this release, OpenAI made a major announcement: allowing individuals without coding knowledge to easily define a GPT. Sam Altman conducted a live demonstration for this purpose.

“Having worked at YC for many years, I’ve always encountered developers seeking business advice,” said Sam Altman. “I’ve always thought it would be great if one day a robot could answer these questions for me.” He then opened the GPT Builder, typed in a definition for this GPT aimed at helping founders of startups think about their business ideas and receive advice. During the conversation, the GPT Builder generated a name and icon for the GPT, and through dialogue with Sam, it inquired whether he wanted to make adjustments to the generated name and icon.

Following this, the GPT Builder proactively asked how the application should interact with users. Sam suggested choosing appropriate and constructive responses from his past speeches and uploaded a segment of one such speech. Including the explanation, the entire application was completed in just three minutes. Visitors to this GPT are greeted with an auto-generated conversation starter and can discuss startup-related queries, receiving responses akin to those Sam Altman himself would provide.

Sam indicated that creators could also add actions (dynamic interactions) to their GPTs. Essentially, the customization features users can define for a GPT are not extensive: predefined prompts, external knowledge bases, and actions. Yet, seamlessly integrating these elements to enable those without coding skills to create applications is indeed an innovation by Open AI.

After a GPT is released, the application can be set to private, exclusive to a business, or publicly accessible. Open AI has stated it will share profits with popular applications.

Clearly, OpenAI’s release does not aim for ordinary users to create complex applications through natural language alone. The real value lies in the potential for individuals and businesses to upload their knowledge bases to OpenAI and construct bespoke applications with a single click.

For instance, a shipping agent with a freight rate sheet could upload the file to OpenAI and deploy their pricing assistant instantly – a streamlined and smooth application deployment that didn’t exist before. If such releases gain user approval, they could fill OpenAI’s app store, turning it into a treasure trove of diverse information.

03: Introducing Code-Free AI Agent Creation with OpenAI’s Assistants API

If you found the zero-code GPT impressive, OpenAI has now introduced an even simpler way for developers to leverage the OpenAI API – the Assistants API. Sam Altman highlighted the remarkable experiences created by API-based agents in the market. For instance, Shopify’s Sidekick allows users to take actions on the platform, Discord’s Clyde assists administrators in creating custom characters, and Snap’s My AI serves as a customizable chatbot that can be added to group chats to offer suggestions.

However, building these agents can be complex, often requiring months of work by teams of engineers to handle tasks including state management, prompt and context management, extending capabilities, and retrieval.

At the OpenAI Developer Conference, these tasks have been API-fied – with the launch of the Assistants API, developers can now build ‘assistants’ within their applications. With the Assistants API, OpenAI clients can construct assistants that execute tasks using specific commands, leveraging external knowledge, and calling upon OpenAI’s generative AI models and tools. Use cases for such assistants range from natural language-based data analysis applications to coding helpers and even AI-powered holiday planners.

The capabilities encapsulated by the Assistants API include:

Persistent threads, eliminating the need for users to manage long conversation histories.
Built-in retrieval, enhancing the assistants created by developers with external knowledge, such as product information or documentation provided by company staff, and offering a new Stateful API for context management.
An integrated code interpreter that can write and run Python code within a sandboxed environment. This feature was introduced for ChatGPT in March, enabling the creation of graphics and charts and the processing of files, thereby allowing assistants created with the Assistants API to iteratively run code to solve coding and math problems.
Improved function calls, enabling assistants to invoke programming functions defined by developers and integrate the responses into their messages.
The Assistants API is currently in beta and available for all developers starting today. Developers can visit the Assistants Playground to try out the beta version of the Assistants API without writing any code.

Demo: With the Assistants API, developing agents requires no coding | Source: OpenAI

The Assistants API is seen by OpenAI as the first step in helping developers build ‘agent-like experiences’ within their applications. With the Assistants API, building agent applications has become much easier. OpenAI has stated that over time, they will continue to enhance these capabilities. Additionally, plans are in place to allow customers to offer their own copilot tools to complement the platform’s Code Interpreter, retrieval components, and function calls.

The product upgrades announced at the OpenAI Developer Conference once again bring us closer to a future where everyone can have one or even multiple personal assistants, develop software using natural language, and browse, purchase, or access popular personal assistants for free or for a fee.

In just over half a year, from GPT-4 to GPT-4 Turbo and the GPT Store, OpenAI has made rapid strides. Within these six months, the global technology and AI industry has undergone a transformative shift.

As OpenAI continues to innovate with multimodality, longer text inputs, more affordable options, and personalization – features designed to compete against large-scale AI models – it’s unclear how global AI competitors feel. But the technological advancements of OpenAI are thrilling, and the business strategies employed by the team are mature beyond those of an average startup. We are witnessing the pulse of an industry and perhaps the birth of a titan.