Speaker A
Recently, something called Zeb appeared, and it's causing quite a stir, being called a revolution on par with ChatGPT. For instance, when Claude Fable was made to play Pokémon Red, it took 50 hours to clear it, but Zeb, despite
Overview of Zeb, a fast, cheap AI classification model revolutionizing decision-making alongside LLMs, plus a demo of OpenAI's Datz digital assistants.
Key Takeaways
- Zeb offers a revolutionary approach to AI classification with high speed, low cost, and strong accuracy.
- It complements LLMs by handling decision-making tasks, improving overall AI system efficiency.
- Practical use cases include real-time trading bots, browser automation, and AI task delegation.
- Zeb requires sufficient background knowledge input to perform well due to limited reasoning capabilities.
- OpenAI's Datz represents a new interactive AI assistant paradigm, enhancing productivity through natural communication.
What the video covers
- Zeb is a new AI model gaining attention for being faster, cheaper, and more accurate in classification tasks compared to existing AI models.
- Unlike typical LLMs, Zeb specializes in quick, structured decision-making rather than long text generation, acting as a 'system model' for AI workflows.
- Zeb is reportedly several hundred times faster and 1,000 times cheaper than other AI models, with a 0% hallucination rate.
- It can be used for real-time applications like trading bots, browser automation, and hiding AI-generated images during scrolling.
- OpenAI's Datz digital assistants allow users to interact with AI via texting or calling, automating tasks such as calendar management and news summarization.
- Zeb requires providing the current state and questions separately in API calls, supporting multiple answer formats like probability, multiple-choice, and scoring.
- Zeb is a general-purpose classification model that does not require specialized training or data preprocessing for different tasks.
- While Zeb is fast and accurate, it lacks deep background knowledge and reasoning, so it needs contextual information via prompt engineering or retrieval-augmented generation (RAG).
- Zeb currently performs best in English and struggles with Korean slang or casual expressions.
- Research and practical applications of Zeb and similar classification models are ongoing, with potential to improve AI agent speed and token efficiency.
Chapters
- 00:00Introduction to Zeb and its capabilities
- 00:19Real-life applications of Zeb
- 00:41Skepticism and investigation into Zeb
- 00:59OpenAI's Datz digital assistants demo
- 01:39Background on TypeSafe AI and Zeb's model type
- 02:02Zeb's role as a decision-making system model
- 02:50Using Zeb's API and input/output format
- 03:34Summary of Zeb's classification capabilities
- 04:16Comparison to traditional classification models
- 04:39Zeb's advantages, limitations, and future outlook
Full Transcript — Download SRT & Markdown
Speaker A
being a smaller model, cleared it in much less time. The shocking part is that it only cost 1 dollar. They even say that games requiring real-time physical skills can be run with Zeb.
Speaker A
People are also making useful things for real life; for example, someone used Zeb to create an extension that hides AI-generated images in real-time while scrolling. They say you can even make trading bots that require extremely fast decision-making. It
Speaker A
supposedly hit negative 30,000 dollars when run, but hey, at least it was fast. How is this possible? It's because Zeb is several hundred times faster than existing AI models. On top of that, it's 1,000 times cheaper and has a 0%
Speaker A
hallucination rate. If this is true, other AIs might have to go out of business, but when I searched for Zeb, only weird sites came up, so let's find out what this actually is.
Speaker A
Before that, today's sponsor is OpenAI. They have created these furry characters this time. They're called "Datz," and you can think of them as your own digital assistants that you can make work 24 hours a day. So now,
Speaker A
instead of writing prompts, a world has come where you use AI by texting or calling your Datz. If you link plugins, you can even have them handle tasks for you, so I actually tried giving Datz some work. First, you adopt a Datz,
Speaker A
give it a name, customize it, and then whenever you have something to do, you just send it a text. I asked it to create a calendar organizing anime events, and surprisingly, it did a really great job. If you link your
Speaker A
Google Calendar, you can see it handling your schedule automatically. And the thing I find most annoying is looking up news every morning and creating content, but it seems to be doing that very well, too. It reminds me a bit of this, but it’s like a
Speaker A
version on steroids. Currently, it's only available for Pro accounts, but if you're going to use Codex a lot with your Pro account anyway, Datz is free, so you should try using it together.
Speaker A
There's a company called TypeSafe AI, which was reportedly founded by former OpenAI employees. It looks like they are building "type-safe" AI, and that’s where they created Zeb. What kind of AI model is this? It's not like the common LLMs that generate long
Speaker A
sentences, but an AI model that quickly generates structured answers. Looking at this, it doesn't really seem like an AI and looks pretty useless, but where do you use it? Sometimes AI suddenly needs to make a decision, right? For
Speaker A
example, if you ask an AI to build something, it often gets overwhelmed and has to decide whether to search the web or look through local files to implement it. But if you leave that decision-making to something like Astra,
Speaker A
it’s a bit slow. So, if you entrust all these decisions to Zeb, the work speed becomes insanely fast. It even allows for light-speed processing when you need to make various decisions in everyday tasks. Since it doesn’t have
Speaker A
LLM capabilities like reasoning or text generation, you use it alongside those models; it’s an AI model that specifically handles the decision-making role for them, which is fancy-termed a "system model." You can think of it as a decision-making tool
Speaker A
that will become popular in the future. So, I tried using it myself. First, you need to pay a minimum of 5 dollars to use it in API form. Then, after writing the code in Python like this, you get
Speaker A
your API key and insert it here. Next, you ask Zeb questions just like you would with other AIs, but interestingly, you have to input these two things separately. For "state," you just write the current status in plain text. And
Speaker A
for "questions," you just write your questions there. You write the name of the question here, and then you enter the actual question right here. Then Zeb reads the current state and provides an answer to the question. In the "questions" section, you can
Speaker A
specify the format for the answer in advance. If you write "number," it will return a probability between 0 and 1.
Speaker A
So, if you run this, Zeb will answer the question, and it seems the furry probability is 89%. You can also ask multiple-choice questions like this. If you set the question type to "choice," you can make it select from several
Speaker A
provided answers. I asked where I should send this furry person, and it suggests it would be best to send them to option 1. By using the "score" type, you can also have it evaluate levels.
Speaker A
For instance, I described the current outfit and asked it to evaluate how shameless it is; if you provide criteria, it will assess which category it falls into. If you run it, you can see it evaluating that this is
Speaker A
appropriate for a certain level from 1 to 5. To sum up, Zeb is an AI model that’s like a magic conch shell. So, by using this, you can even entrust the act of making judgments to AI. However, veterans in the field might wonder why
Speaker A
this is supposed to be special. That's because using AI as a classification model has been around for a long time.
Speaker A
When you get started with deep learning, there’s a "Hello World" style example where you feed it an image, and it’s a basic model that identifies which digit from 0 to 9 it is. You probably learned by building
Speaker A
classification models like that, and now, in the age of LLMs, these classification models are starting to trend again. And it's not just something Zeb can do; it's a capability that LLMs can perform just the same. Some people have implemented
Speaker A
Zeb in a few lines of Python, showing that you can even achieve it by simply "brainwashing" a local model. If you make a fuss about Zeb without knowing better, you might look like an amateur, so let me point out the exact
Speaker A
highlights: first, Zeb is a classification model, but a general-purpose one. In other words, you don't need to build a specialized AI model for every classification task, nor do you need data preprocessing; you can just input long strings of text to
Speaker A
perform various classifications and judgments. That’s why some are calling it the ChatGPT-level revolution for classification models. Second, it has accuracy comparable to GPT-4, but it’s insanely fast and incredibly cheap, making it a highly practical model. Since the agents we use now
Speaker A
involve many judgment steps, their operational speeds are likely to become much faster. It should also be very helpful for token-efficient engineering. Its biggest drawback is that it’s not very smart. That’s because it lacks a solid base of background
Speaker A
knowledge. So, when using it in practice, you need to provide sufficient background knowledge via RAG or other methods in the prompt to ensure it makes good judgments. Also, it doesn’t understand Korean very well; it’s mostly for English. So, if
Speaker A
you use Korean slang or casual terms, it probably won't have a clue what you're talking about. For this, you’ll need to provide enough background knowledge in Korean or simply wait for the next version of Zeb. OpenAI has also released a
Speaker A
similar model for classification purposes. There are also many open-source "lite" versions available, so you can run them for free on your own GPU. Research into real-world applications of Zeb is also ongoing; for example, some have connected Zeb to a browser to enable automated web
Speaker A
navigation. So I don't think this is something that will be hyped up and forgotten next month, because things like this are actually practical.
Speaker A
Reducing tokens or improving speed are always useful topics, so I expect them to be researched a lot in the future.
Topics:Zeb AIAI classification modelfast AI decision-makingOpenAI Datzsystem model AIAI automationlow-cost AIAI hallucination rateAI real-time applicationsAI prompt engineering
