「Message ChatGPT...」
When we start a topic, ChatGPT defaults to this sentence in the blank input box.
Interaction with ChatGPT is like chatting with friends, you come and go, from top to bottom, the linear structure is very intuitive.
However, is this the best way to interact with AI?
Still scrolling through ChatGPT's chat history? A very new way to interact is coming
When we were students, we should have all seen a whiteboard, where the teacher used an erasable marker pen to write formulas on a white smooth board.
Design platforms such as Canva and Figma provide online whiteboard tools, which follow the characteristics of physical whiteboards, but the canvas is infinite and the form is more free, just like a real desk, which can flexibly place various things, which is very suitable for brainstorming.
▲ Schematic diagram of the online whiteboard tool. Image credit: Canva
If ChatGPT also interacts with us in a whiteboard fashion, what kind of wonderful chemistry will it produce?
Flowith is one such AI content generation tool that recently launched an open beta and added about 20,000 new users a week.
Its biggest highlight in terms of form is that it transforms the vertical chat record into a dynamic canvas, providing a very new way of human-computer interaction.
Flowith supports models such as GPT 4 Turbo and Claude 3 Opus, and is similar in essence to Perplexity, building services on someone else's foundation, but each with its own innovations.
Flowith still has input boxes at the bottom, but you're interacting with the AI on an infinite canvas that resembles an online whiteboard.
A new concept is also introduced here: nodes.
Each time a question is asked, a new node is generated, and the nodes can diverge and refer, strung together into many different lines.
For example, it's more intuitive to ask the AI directly, "Give me 5 AI startup ideas." 」
The answers given by Flowith are: AI medical imaging diagnosis, AI online education platform, AI intelligent investment advisor, AI smart home wearable devices, and AI scenario consumption recommendation.
Then, under the branch of "AI Smart Home and Wearables", we can ask who the main players are at present.
The first time I saw the answer bifurcated and moved in the form of a card, I couldn't help but say "wow".
This example uses the "mind map divergence mode". In addition, Flowith also provides automatic decision-making mode, networked super search mode, and image generation mode, which are suitable for different problems and scenarios.
Most of the open-ended questions are left to the mind map mode, and the online search mode is like a replacement for Perplexity, which can search, summarize, and even open the mind.
When I type in "How is the development of AI search in China?", Flowith's output is divided into three columns.
The left provides real-time search results and images, which can be further queried and asked by the AI to summarize the content of a certain link, and in the middle, several of Google's top-ranking links are summarized, which is equivalent to outputting a short "literature review".
But the most interesting column on the right is that AI gives us a knowledge graph to help us expand our ideas and build a knowledge system.
The "state of AI search development" itself is a very broad question, and Flowith extends to many sub-directions: the integration of AI in search engine functions, the progress of AI search algorithms, and bias in AI search results......
Humans with a strong sense of curiosity can continue to explore. The dynamic effect of selecting the knowledge graph is also a bit magical, QQ bounces, like the texture of jelly.
In the automatic decision-making mode, Flowith will automatically switch to the most appropriate AI model based on your prompt words, determine whether you need to search online, whether you need Wensheng Graph, and choose the speed-first GPT-3.5 or the stronger GPT-4 and Claude 3, which can give a glimpse of the prototype of the Agent.
For example, ask Guangzhou what the weather is like today, and Flowith will do a real-time search.
Overall, Flowith's dynamic UI is also cool, with the entire canvas being dragged, zoomed out, and zoomed in, and nodes referencing, questioning, deleting, and regenerating.
With Flowith, it's a bit like how we used to scribble on scratch paper to think about problems, diverging, extending, deleting, and rambling, not like a logical outline, nor like an intuitive chat with ChatGPT.
But that's what we use the canvas for, and in the process of diverging our creativity, we slowly clear our thoughts like flowing water.
The visualization of canvas AI is itself an explicit expression of thinking.
Form affects thinking, refusing to trap the brain in the "chat box"
The essence of innovation is to enable more efficient interaction between humans and AI.
If ChatGPT is a chat-style AI, products like Flowith can be referred to as canvas AI for short.
Similarly, FigJam, an online whiteboard tool from design platform Figma, also offered a built-in AI widget, Jambot, in August last year. Jambot incorporates ChatGPT's generative capabilities to brainstorm, summarize, rewrite, and more.
Aosheng Ran, a product designer involved in the development, mentioned at the launch:
I think we're stuck in the chat box. It's like we're stuck in Zoom right now.
▲ Jambot
There are limitations to interacting with chat-style AI.
ChatGPT and other chatbots are based on a linear structure, but thinking is often not linear, and large models have a certain degree of randomness, and the generated content, such as copywriting, translation, and code, often needs to be repeatedly modified, compared, and tested.
Even though ChatGPT can be generated repeatedly, if a question has been spoken hundreds of times and the chat history is too long, we may not know where we are talking.
Even if a problem is studied clearly, ChatGPT cannot visually represent the different branches that arise from the problem.
But canvas AI is not the same fireworks.
On the same canvas, we can repeatedly modify the prompts, or try several times with different large models, and the answers are all retained, and we can also complete multiple tasks and view the results at the same time, searching, brainstorming, content summarization and other needs in parallel.
▲ Picture from: X@DerekNee
For example, when the holidays are approaching, I first used the brainstorming mode to ask Flowith to list the 5 characteristics that an AI travel assistant should have, and then used the online search mode to directly find the 3 most popular AI travel assistant apps at the moment.
By analogy, when it comes to creative scenarios that require repeated generation and comparison of content, as well as workflows that give full play to AI's multiple capabilities, canvas-based AI is often more interesting, intuitive, and interactive than chat-based AI.
The team collaboration feature that online whiteboard tools pride themselves on is also what canvas AI excels at, and chat-based AI can only afford to be resigned.
ChatGPT supports sharing chat history, but it can't continue the conversation from a shared link.
Perplexity supports sharing search results, and other users can follow up on the original basis, but after the follow-up question, a new link will be created, which is not on the same page as what we shared.
However, as long as Flowith has the "editable" permission, the other party can participate in the teamwork and ask their own questions in the input box. If it's just "viewable", the other person can also leave a comment even if there is no input box.
Imagine sharing a "flow" with others, directly on the canvas for real-time multi-person generation and discussion, without interfering with the original traces, passion and Party A Battle, a feeling of scolding Fang Xuan in Feishu Docs.
If you don't know how to use a notion-like template more efficiently, Flowith also provides a Discover page to recommend how others can use it. Validated workflows are processed into ready-to-use semi-finished products.
The reason why the difference between canvas and chat is repeatedly emphasized, and even some of them are opposed, is because the way we interact subtly affects the way we think.
If we just see an input box, we may also fall into "conversation mode", talking to the AI as if we were communicating with a human, saying hello to it, asking it how it got rich overnight, or confiding in it our feelings.
This is also a good usage, but when there are more and more free ways to use AI, there are more possibilities for content production, and the boundaries of AI capabilities are closer to us.
Innovation is not the end, and we do not believe in any form of interaction
Canvas AI and chat-based AI are not a life-and-death substitute, but a complement of beauty and beauty.
Chat-based AI has its drawbacks, and canvas AI isn't always the last laugh.
If you're just looking for a simple question or learning about an unfamiliar concept, it's faster to turn to GPT-4 and Perplexity.
As a fledgling product, Flowith is not so easy to use, and the shortcomings are visible to the naked eye.
There are not many models supported by the generative image mode, and currently only Stable Diffusion and DALL· E 3, which previously supported Midjourney, but is no longer an option.
At the same time, the model retrieved by Flowith has 2 echelons, the advanced T1 and the basic T2. While T2 models such as GPT-3.5 Turbo are free for unlimited times after inviting friends, T1 models such as GPT-4 Turbo and Wensheng Diagram models have usage credits.
In personal experience, Flowith has had some strange situations, but is unsure of the prevalence of these problems.
Ask Flowith questions in Chinese, and the answers may also be in English, which may affect the user experience. Ask GPT-4 and GPT-3.5 the same prompt, and the answers are in Chinese.
While Flowith supports reading PDFs, if the uploaded file is too large, it will most likely be unsuccessful.
Flowith's interaction is also not smooth enough, and when you enter a new question, other nodes and the overall interface will always change dynamically, which can sometimes make people lose the visual focus. While the effect is cool, the freshness is only temporary.
In addition, when a certain node of Flowith is questioned, the context is not necessarily coherent, and there are situations where the answer is not asked.
Although the product is not yet mature, Flowith has done something similar to Perplexity: we are looking at how we can apply the capabilities of AI on top of the large language model to better integrate AI into our workflow.
Flowith developer @DerekNee shared his insights on the product at X – you shouldn't just think of Flowith as a whiteboard AI tool.
He is a heavy user of GPT-4 and AI products, but when it comes to writing copy, writing code, and searching for content, the mainstream chat-style UI is not efficient enough to be repeatedly modified, compared, and tested.
In his view, whiteboard is an alternative to the concept of two-dimensional canvas and nodes for the general public, and the core of Flowith is still to help users generate content and explore new and more effective ways of interaction with Human-AI.
This is in line with Figma's philosophy. When Jambot was launched, Figma wrote in his blog that AI is a tool that enhances work and collaboration. Rather than letting AI do things for you, AI is helping you do it better.
Office software such as WPS and DingTalk are AI-based, and PPT can be generated with one click at the user's command, and the AI assistant can read documents quickly, which is actually the same reason.
Flow is meant to be "flowing," and in a way, it's what we want from AI tools: to help the human mind flow.
In fact, the way we think and solve problems has been changed by AI.
AI search, such as Perplexity, has gradually become a product that cannot be separated from every day, and humans ask questions, and they search, summarize, and write, which is already a mature workflow in itself, and we pay less, but are more efficient. Thinking back to the days when only traditional search engines were used, the eternity was like a long night.
Flowith, which hasn't run in well with humans, can't replace ChatGPT, but that's not a problem.
Design platforms like Canva can make a niche in Adobe's space because the underlying layer of building a product is different from Adobe, which is templates and widgets, and the latter is pixel-level tweaking of images.
Canva can't steal Adobe's loyal users, but it does attract users who don't know or can't use Adobe.
Flowith 也是类似，本质是在填补市场的空缺。 Perplexity 不重复 Google，Flowith 不重复 Perplexity。
ChatGPT has led the way in chat-style AI, which is really convenient and simple to use, and after it, many products have followed suit, and users have accepted it by default.
In the future, however, our collaboration with AI will not just be about translation, Q&A, but should allow more interesting ideas to happen, which may not be successful, but will be enough to surprise them.
After all, in the era of AI, changes are always fast, and there must be fresher ways to interact with AI to improve efficiency, innovate thinking, and change workflows.
After playing with more products, we were still teenagers and exclaimed: This is the AI application I want.