What is Amazon PartyRock and How Did AWS Launch a Marketplace Before OpenAI's GPTs?
The rise of no-code generative AI-fueled personalized experiences
Amazon’s PartyRock enables anyone to build no-code generative AI-powered apps. The company said tens of thousands have been created since its launch two weeks ago.
It may seem like Amazon’s answer to OpenAI’s GPTs, but PartyRock is more of an educational precursor to a product similar to Microsoft’s Copilot Studio. Curiously, on Sunday, Amazon announced the arrival of a “Discover page, where top community-created apps will be curated by AWS.” In other words, AWS introduced a “skill” store for PartyRocks (or maybe we can just call them Rocks) before the GPT store launched.
This is an important development in the evolution of the generative AI market. Up through October of this year, there have been three key roles in the generative AI ecosystem, and now there is a new type of participant:
Creators of AI foundation models
Software developers that use those models to create applications
Application users
New: assistant creators that offer customized experiences
GPTs lowered the technical barrier to participating in the generative AI “app” customization market with its no-code builder. PartyRock is formatted differently but essentially provides a similar no-code entry point for anyone to create a customized experience on the back of generative AI foundation models.
Comparing PartyRocks to GPTs
I have made several GPTs and decided to recreate one in PartyRock. You will see they are not exactly the same in terms of output or experience and are quite different from the creation perspective. PartyRock is currently more about fun and engagement. You can do that with GPTs, but you also have additional control and can develop more robust experiences. Check out some examples by clicking the buttons below (N.B. GPTs currently require a ChatGPT Plus subscription to access, so I will also include screenshots here, and you can watch the video of 8 GPT demos to see them in action).
You can see these are fundamentally different experiences. GPTs enable you to have an open-ended conversation with the app and create your own SuperHero. You can even upload an image, and it will create a SuperHero in your likeness. Partyrock only enables you to enter a name, and it then creates a backstory, generates an image, and presents a trivia option. It will also enable you to chat with your new superhero.
I like the PartyRock widget option that enables creators to build their preferred layout and customize the on-screen user experience. GPTs force you into the same familiar ChatGPT experience. That approach also has merit, particularly in terms of user familiarity, but predetermines more UI constraints.
Other than that, you will find that GPTs are superior today. This is probably customizable, but PartyRock always wants to force me into starting with a name and requires more design decisions to customize the application. After I enter the SuperHero name, PartyRock does all of the work. It’s fun but not as engaging and lacks depth, unless I spend more time tuning a series of prompts tied to specific widgets.
GPTs are more robust. The SuperHero GPT can just create everything from scratch or start with a name. It is more open-ended. Depending on how much information you provide in your prompt, the GPT will ask for additional guidance.
After my response of one-word answers to this list of questions, the GPT responded:
I can also upload a selfie, and it will (sort of) create an image in my likeness as the SuperHero character. Overall, GPTs are less prescriptive and more engaging. This is just one entertainment-oriented example. You may find a different experience with expert and task-oriented Rocks, but I doubt it.
In theory, I like the finer level of control in assembling a number of widgets, each with its own task. In practice, the black-box style GPTs with fewer configuration options produce better results with less effort. I suspect this means they will be good for different use cases. And I hope GPTs begin offering additional features related to threading different functions together and offering new layout options.
11 Things I Learned Using PartyRock
A few additional PartyRock observations are also worth sharing.
English only: PartyRock only supports English today.
Claude is best: For each widget (i.e., the boxes on the PartyRock screen), you can choose your preferred LLM. Options include Claude and Claude Instant (smaller model) from Anthropic, Jurassic 1 and 2 from AI21 Labs, and soon Amazon’s Titan. I found Claude to be the best for my Rocks so far. Let me know if you have had a different experience.
SD XL is not very good in PartyRock: I am not impressed with the image generation in PartyRock, which is a custom implementation of Stable Diffusion XL from Stability AI.
Don’t expect persistence: PartyRock has a cool snapshot feature to capture your session interaction, but it does not save your conversations. So, the snapshot is your only option.
There are guardrails: I guess PartyRock didn’t like something in the name oxymoron or its own generated output, but I could not create a snapshot of one of my sessions. I was offered a dialogue box saying my snapshot request was rejected. 😟
Remixing is interesting: The remix idea is great for creators who want to take an existing idea and build on it. It has the downside of being unable to protect your PartyRock development ideas unless you keep your Rocks private, but it is an interesting feature for accelerating development and learning.
PartyRock is free up to a point: While PartyRock is in preview, it is free to use, but it is based on credits. Once you run out, you must add a credit card to continue using PartyRock and generating tokens. Still, the credit allowance appears to be generous.
Data sharing default is “opt-in”: Note on the image above that the “Share PartyRock Data with AWS” permission is checked by default. This may not be a privacy best practice, but now you know.
Run duplicate widgets with different settings: You can select different models, change prompts and other parameters, and view the result side-by-side. Do this by setting different variables for copies of a widget. This may be PartyRock’s best feature. Why doesn’t everyone do this?
PartyRock is unsupported: PartyRock is in preview and technically is an unsupported application. So, don't build anything critical and expect it to be around and supported. That likely will change as Amazon Bedrock intends to enable users to implement Rocks into their generative AI application portfolio.
PartyRock is a learning tool: While GPTs are designed to be fully formed applications, PartyRock documentation stresses its function as an educational tool. It is a more user-friendly version of the playgrounds that are commonly offered by LLM makers.
PartyRock and Skill Stores
Many will recognize the skill store reference above as pointing to Amazon Alexa. It was a key factor in Alexa’s early adoption patterns. After a user had acquired an Amazon Echo, they could browse through the wide variety of Alexa skills and find applications to try. The store was curated by Amazon. That led to Amazon picking the Alexa skill winners and losers, many of which persist to this day.
One early Alexa skill developer showed me that getting featured by Amazon’s skill store curators in the weekly email or online would guarantee an eight-fold increase in users for that week, even for well-established skills. The impact was even bigger for new skills. This approach worked well when the majority of Alexa users were motivated to explore the voice assistant’s potential. Later, when the “normies” arrived, many didn’t want to go to the web or open an email to find out the wonderful things that their device could do. So, the impact waned.
Bringing this back to PartyRock, the Discover “Rock” store could become a great asset, though it is hardly that today. Today, it only highlights a few Rocks created by Amazon employees. These are best viewed as templates to serve as a frame of reference for new creations.
The “remix” button is the most intriguing in this regard. It enables users to make a copy of an existing Rock as a starting point for a new creation. The user just needs to click “edit” to make changes. This will undoubtedly lead to a lot of throw-away copycat Rocks. It also may be a disincentive for some creators. Why invest in creating something when everyone else can just copy it without a second thought?
Still, I like this idea of remixing if it were listed as an option. Some Rock creators could share their templates for others to get started, while others would want to keep their secret sauce secret. This may remind some people of Alexa Skill Blueprints, and it will both help people create Rocks and lead to a lot of low-quality options. Thus, Rock store curation may not just be a useful feature but a requirement for a decent user experience.
With all of this said, I expect OpenAI’s initial GPT store to be more extensive and more useful. There are over 10,000 GPTs now, and some are likely to find an audience. Even if OpenAI doesn’t wow everyone with its online store features, there are a lot of motivated ChatGPT users who will want to explore GPTs that could enhance their application experience.
What’s Next
PartyRock’s larger significance is that it represents a third meaningful offering in the no-code generative AI application development category and the first by a company that has previously built a developer ecosystem and application store. The first meaningful example in this category was Poe, the generative AI chabot builder from Quora, led by OpenAI board member Adam D’Angelo. Then GPTs were launched and received with great enthusiasm. Now we have PartyRock, which is flying under the radar.
I suspect two things are about to happen. First, other companies with generative AI-powered chatbots are going to move into this space to maintain competitive parity and offer an easy path for users to customize their experience. Google is the most obvious, as it doesn’t want Bard to follow the path of Google+. Anthropic will probably not go that route in the near term as it appears laser-focused on business applications, even though Claude has a consumer or prosumer bent.
Second, while PartyRock is a product of the business-oriented AWS Bedrock service and Amazon teams are notoriously siloed, Amazon may soon offer a ChatGPT competitor. Maybe this will be a re-launch of Alexa, or it could be something new. Microsoft has the tooling in place for personal copilots but is currently marketed to business users, so it is more likely to enable those creations to be used through ChatGPT than to create a competing offering.
Amazon is in a different place. The PartyRock name is derived from AWS’s Bedrock offering for business users. It is an alternative to Microsoft’s Copilot Studio. In that way, PartyRock points to a goal of competitive parity with Microsoft, not OpenAI.
However, Amazon’s consumer business and Alexa experience are likely to motivate the introduction of a consumer-oriented competitor to ChatGPT. I suspect this will be a relaunched Alexa with some PartyRock-style features. PartyRock is a similar approach to Skill Blueprints that made it easy for anyone to create customized, no-code Alexa skills.
Let me know what you think.
Check out our amazing sponsor, Dabble Lab, to learn more about bespoke AI and application development powered by AI scheduling and tracking.