[ad_1]
For some, there is something behind the term artificial intelligence still a really big question mark. At first glance you can see what the tools can do, but how they work is mostly a mystery.
To better illustrate AI, Ishan Anand, VP of Product at Edgio, created a special project. With “Spreadsheets are all you need“You should only understand AI using a large Excel spreadsheet.
How does the AI Excel spreadsheet work?
Anand has the basic functions of GPT 2, one of the predecessors of ChatGPT 4, transported to an Excel table. You can easily download the table from his website. At 1.25 gigabytes, the file is significantly larger than average Excel tables. Because there are 124 million AI parameters in a table.
In the Excel table you will find several pages – including a page for prompts and many pages with data. They give you insight into the processing steps of GPT:
- Tokenization
- Text & position embeddings
- Multi-head attention
- Multilayer Perceptron
- Language Head
If you enter a prompt on the “Type Prompt Here” page, the Excel table converts it into so-called tokens. Each word is assigned one or more tokens to identify it precisely. Anand uses “Mike is quick” as an example. He moves” to give the AI the task of completing the second sentence based on the information given.
In the next step, the Excel AI finds the “text & position embeddings” for the tokens. This means: Each token has a corresponding list of 768 numbers that reflect its meaning. At the same time, the AI keeps a list of the positions of the words in the prompt. So if there are duplicate words in the prompt, they get different embedding codes.
The final result
The AI now looks for connections within the sentence in the “Multi-head Attention” step. Based on the codes, the AI recognizes that the word “He” refers to “Mike” and “moves” is related to “quick”.
In the “Multilayer Perceptron” step, the tool finds out what meaning the words have. The English word “quick” could mean “clever” or even “lively”, but in the context of “moves” it best describes a quick, physical movement.
Editorial recommendations
So it looks for words like “fast”, “quickly” and “around” from its own database in order to meaningfully close this gap. The “Attention” and “Perceptron” steps occur several times in a row to increase the probability of the answer.
“Language Head” finally selects the word that completes the clause. The AI selects from the word pool that was created in the previous step. In most cases, the AI will choose the answer that is most likely based on the data. However, the other answers can also be used to suggest alternatives. In the case of the example, the AI initially suggests “quickly”.
No ChatGPT yet
The Excel table is primarily used to illustrate AI. You won’t find a real chatbot that spits out answers to your questions here. This is because programs like ChatGPT still need human feedback to optimize machine learning models.
By the way, the project initially started in a Google Sheets spreadsheet. However, the model with 124 million parameters quickly became too large for the program. That’s why Anand switched to Microsoft Excel.
[ad_2]
Source link