DEV Community: SamyMe

All AI trough a single API - Open Sourced

SamyMe — Wed, 11 Jan 2023 10:00:35 +0000

Artificial Intelligence is about allowing a machine to mimic intelligent human behavior. Now, machine and deep learning make it possible to teach computers with data what we fail to describe in an algorithmic approach. This unlocks new capabilities for our programs to :

📖 Read a text and understanding concepts like : the topic of the text, the sentiment of the writer and the key concepts it talks about
🗣️ Hear a voice , understand the words it says and responding to it (Siri and Alexa : speech to text and text to speech)
🎥 Look at an image or watch a video and understand the big picture and the details of what it contains
🔮 Predict the future : Based on historical data, predict malfunctions, miss payments, growth ...etc

But in order to achieve these objectives, you have to train large models with millions of labeled data.
Now, with the rise of AI as a Service , a lot of companies provide off-the-shelf trained models that you can access directly through an API. These companies are either the tech giants (Google, Microsoft , Amazon) or other smaller, more specialized companies, and there are hundreds of them. Some of the most known are : DeepL (translation), OpenAI (text and image analysis), AssemblyAI (speech analysis).

There is a large panel of AI companies each one providing new and innovative AI models with wide variety of applications and an important mission would be to :

1/ Empower any stack developers with easy access to on-demand intelligence.

Intelligence is the new essential resource that’s usually compared with energy (AI is the new electricity).

Simplifying access to intelligence by making it an on-demand self service is important to reduce friction and encourage innovation.

And what a better way of presenting this than simple and intuitive API with clear and complete documentation and a great graphical tool for simple exploration!

2/ Discover and showcase the best AI can offer.

Anyone should know how easy it is to moderate any type of content (images, videos, text), to automatically transcribe an audio file and understand its content or to translate an entire word, pdf or ppt document while keeping its structure intact.

With the rise of generative models, it's also important to know what applications it could have. With the right prompt, generative models can be extremely useful. You can use them to classify documents, summarize them or ask questions to retrieve valuable information from them.

But it's very hard to know what company does what and at what price.

So we believe that it's important to clarify all of this innovation by bring it all together so that any new applications AI companies create can be easily discovered in one place.

3/ Bringing it all together for greater flexibility

One needs to keep up with the rapid advancements of artificial intelligence by switching to new AI providers as soon as they outperform old ones. So having the provider as a simple API parameter can be a huge advantage.

Combining multiple providers can be very important too : You need to extract specific information from a document but AWS's API doesn't support your language ? It's OK you can translate the document with DeepL and query AWS with a single API and with a single account.

Finally, in order to make AI at the hand of everyone, we're working on getting AI everywhere by integrating all AI APIs into different solutions : ERP (like oodoo and sap ), CRMs (like salesforce), NoCode platforms (make-integromat, zapier bubble…etc) and other tools like Data Platforms (Alterix, Dataiku).

Open Source :

Now we’re open sourcing our aggregation layer as a python module that you can ⭐ find on github ⭐ .

Join us on our discord and let us know what you think of it. We'd love to have your feedback and/or contribution.

OpenAI GPT-3 vs Specialized Models [Benchmark] - Should AI companies be really worried ?

SamyMe — Fri, 06 Jan 2023 14:37:41 +0000

OpenAI’s Objective is to create Artificial General Intelligence - “highly autonomous systems that outperform humans at most economically valuable work” .

They are planning on combining text, image and speech models.

In this article, we focus on one of their actual milestones, GPT-3 . Can it really achieve state of the art performance on any language task (compared to specialized models)?

OpenAI’s GPT3 : Open AI is challenging these companies (especially NLP models) with their GPT3 model.

Let’s ask ChatGPT (GPT-3 optimized for dialogue) what GPT-3 is ?

Specialized AI Models : There are a lot of AI companies training specialized models for specific tasks and providing access to them through APIs. Among witch big tech companies (Google, Amazon, Microsoft and IBM) as well as smaller companies focused on specific tasks (DeepL(translation) , Deepgram (speech) , Clarifia (vision) )

Large Language Models like GPT-3 should be able to compete with specialized models on a lot of Natural Language Processing (NLP) tasks without fine tuning. That's referred to as Zero Shot Learning . Let's verify that!

Benchmark

In order to verify that, we chose four tasks : Keywords Extraction, Sentiment Analysis, language detection and translation to benchmark GPT3 against other state of the art proprietary models from different companies.

We've done that using a single API : Eden AI. Code snippets will be available for each task so that you can reproduce the predictions yourself on your own data.

There is an Open Sourced version of EdenAI that you can ⭐ find on github ⭐ as a python module !

Please refer there for more information on the different APIs used in the benchmark.

Language Detection:

Language detection is simply the task of returning in what language a text is written.

1/ Dataset : We used an interesting dataset from huggingface with 20 languages : arabic (ar), bulgarian (bg), german (de), modern greek (el), english (en), spanish (es), french (fr), hindi (hi), italian (it), japanese (ja), dutch (nl), polish (pl), portuguese (pt), russian (ru), swahili (sw), thai (th), turkish (tr), urdu (ur), vietnamese (vi), and chinese (zh)

2/ Evaluation
We compare Open AI to Google GCP, Amazon AWS, and IBM.
We took a few hundreds of examples and evaluated the models performance on them with the Accuracy metric.

3/ Results
The results are shown below with OpenAI ranking 3rd out of the four AI providers we chose.

Amazon 96%
Google 95%
OpenAI 89%
IBM 87%

Sentiment analysis :

This tasks is about understanding the sentiment of the writer when writing a specific piece of text. It can be Positive Negative or Neutral .

1/ Dataset: Most of the datasets we've found did not include a “neutral” sentiment except for the Twitter Sentiment Analysis dataset from kaggle.

2/ Evaluation:
We compare Open AI to Google GCP’s, Amazon AWS’s and IBM’s APIs. using the Accuracy metric

3/ Results :
As shown below, once again, OpenAI gets the 3rd place.

Amazon 76%
Google 66%
Openai 61%
IBM 56%

Keyword extraction :

Keyword or Keyphrase Extraction is about being able to extract the words or phrases that most represent a given text.

1/ Dataset: we selected our datasets from the public github repository AutomaticKeyphraseExtraction
Most of the datasets listed there were too long for the 4k token limit of OpenAI so we had to go with the Hulth2003 abstracts dataset.
Since the different providers are trained to return keywords and keyphrases present in the original text, we did some cleaning to remove all keywords that were not present in the abstracts. We ended up with 470 abstracts.

2/ Evaluation
We compare Open AI to Microsoft, Amazon and IBM.
And we measured their performances using the average precision metric.

3/ Results : This time, OpenAI's GPT-3 was ranked last.

1/ Microsoft 0.6513312046679187
3/ IBM 0.6022276518997
2/ Amazon 0.4954784007523
4/ OpenAI 0.2598775421

Translation :

Automatically translate a text from a language A to a language B.

1/ Dataset : we chose a dataset from the Language Technology Research Group at the University of Helsinki’s Tatoeba Translation Challenge .
We took 100 of examples from different latin languages pairs : deu-fra, eng-fra, fra -ita, deu-spa , deu-swe which constitutes a 500 example test dataset.

2/ Evaluation
We compare Open AI to DeepL, ModernMT, NeuralSpace, Amazon AWS and Google GCP.
A lot of metrics exist for automatic machine translation evaluation. We chose COMET by Unbabel (wmt21-comet-da) which is based on a machine learning model trained to get state-of-the-art levels of correlation with human judgements. (read more on their paper ) .

3/ Results:
The scores are not interpretable but are used to rank machine translation models. And here again, OpenAI ranks last in this task.

DeepL : 0.19001633345126925
ModernMT : 0.17788391513374424
Amazon : 0.16483921567053203
Neuralspace : 0.163133354485786
Google : 0.16280640903935437
OpenAI : 0.15934198508564865

Conclusion

OpenAI's GPT-3's results are quite impressive. We’re getting closer to a kind of general (zero shot) nlp models multitasking without fine tuning.

But in practice, for the moment, if you need an API for one of the tasks we presented in the benchmark, OpenAI's GPT3 wouldn't be the best choice .
In addition to the fact of it not being the best performant model, the 4k input tokens limit can be problematic for reasonably long texts.

We still do need to closely watch the new models OpenAI is working on. As Sam Altman talked about in an interview, they are And implementing a continuous learning approach which would make their model constantly improving by feeding on the internet.

They are also planing on unifying their models to deal with multiple input types which resulting in a single model capable of analyzing any type of data.

Code Snippets :

Language Detection :

import json
import requests

headers = {"Authorization": "Bearer 🔑 Your_API_Key"}

url ="https://api.edenai.run/v2/translation/language_detection"
payload={"providers": "google,amazon,openai,ibm", 'text': "this is a test"}

response = requests.post(url, json=payload, headers=headers)

result = json.loads(response.text)
print(result['google'])
print(result['amazon'])
print(result['ibm'])
print(result['openai'])

Sentiment Analysis :

import json
import requests

headers = {"Authorization": "Bearer 🔑 Your_API_Key"}

url ="https://api.edenai.run/v2/text/sentiment_analysis"
payload={"providers": "google,amazon,ibm,openai", 'language': "en", 'text': "this is a test"}

response = requests.post(url, json=payload, headers=headers)

result = json.loads(response.text)
print(result['google'])
print(result['amazon'])
# ...

Keyword Extraction:

import json
import requests

headers = {"Authorization": "Bearer 🔑 Your_API_Key"}

url ="https://api.edenai.run/v2/text/keyword_extraction"
payload={"providers": "microsoft,amazon,ibm,openai", "language": "en", "text": "this is a test of Eden AI"}

response = requests.post(url, json=payload, headers=headers)

result = json.loads(response.text)
print(result['microsoft'])
print(result['amazon'])
#...

Translation :

import json
import requests

headers = {"Authorization": "Bearer 🔑 Your_API_Key"}

url ="https://api.edenai.run/v2/translation/automatic_translation"
payload={ "providers": "deepl,modernmt,neuralspace,amazon,google,openai", 
'source_language':"en", 
'target_language':"fr", 
'text': "this is a test"}

response = requests.post(url, json=payload, headers=headers)

result = json.loads(response.text)
print(result['deepl']['text'])
print(result['google']['text'])
# ...

OpenAI (GPT-3 API) for WebDevs - Concrete applications

SamyMe — Tue, 20 Dec 2022 13:16:45 +0000

What is OpenAI’s GPT3 and how is it different from other AI models?

Open AI is a (well funded) research and development company that focuses on Artificial Intelligence. They’re developing a large language model, a type of AI model that is trained on a very large datasets of textual data including web pages, books, articles, and more.

How can you use OpenAI’s GPT as a WebDev.

Content generation : Apart from generating code and threatening developers' jobs, a lot of companies are finding a use in GPT3 to create software around automating content creation. It can be blog articles, social media posts or personalized email campaigns. Notion has just published their AI assistant showcasing that pretty well.

Make your app understand texts : GPT3 is great at understanding natural language. In a previous article I talked about AI use cases in web dev and it can do a lot of that. Like : sentiment analysis, text classification and keywords extraction. But you can also do more advanced things like summarizing a text.

Summarization example :


import json
import requests

headers = {"Authorization": "Bearer 🔑 Your_API_Key"}

url ="https://api.edenai.run/v2/text/summarize"
payload={"providers": "openai", 
         "language": "en", 
         "text": "Linux is a family of open-source Unix-like operating systems based on the Linux kernel, an operating system kernel first released on September 17, 1991, by Linus Torvalds. Linux is typically packaged in a Linux distribution..."}

response = requests.post(url, json=payload, headers=headers)

result = json.loads(response.text)
print(result['openai']['result'])

Semantic Search : one of the (limited but evolving) use cases is searching for information (Google vs ChatGPT3). This could also be used to search content inside an app. It could be used as an internal search engine where your users can search for information in your documentation or your app. (Illustration ref)

You can easily do this with Eden AI's API.

import json
import requests

headers = {"Authorization": "Bearer 🔑 Your_API_Key"}

url ="https://api.edenai.run/v2/text/search"
payload={
    "providers":"openai",
    "texts":["In Roman mythology, Romulus and Remus (Latin: [ˈroːmʊlʊs], [ˈrɛmʊs]) are twin brothers whose story tells of the events that led to the founding of the city of Rome and the Roman Kingdom by Romulus.","In ancient Roman religion and myth, Mars (Latin: Mārs, pronounced [maːrs]) was the god of war and also an agricultural guardian, a combination characteristic of early Rome.",
        "Proto-Indo-European mythology is the body of myths and deities associated with the Proto-Indo-Europeans, the hypothetical speakers of the reconstructed Proto-Indo-European language","The Ashvamedha (Sanskrit: अश्वमेध, romanized: aśvamedha) was a horse sacrifice ritual followed by the Śrauta tradition of Vedic religion.",
        "Purusha (puruṣa or Sanskrit: पुरुष) is a complex concept whose meaning evolved in Vedic and Upanishadic times. Depending on source and historical timeline, it means the cosmic being or self, consciousness, and universal principle."],
    "query":"Rome"}

response = requests.post(url, json=payload, headers=headers)

result = json.loads(response.text)
print(result['openai']['items'])

Automated customer service and support : ChatGPT3 is GPT optimized for dialog and it’s leading the Chatbots renaissance. After all the attention chatbots got a few years ago, we saw their limitations very rapidly. But now, with the stunning power ChatGPT showed these bots are going to be upgraded into a new level.

Find the right prompt : All of these applications are simple examples of what a LLM can do. Now it’s up to you to find new applications by finding the right input prompt.

Conclusion :

In addition to gpt, OpenAI have DALLE-E their generative image model as well as the whisper speech recognition model. They are working on unifying their different models to be able to process text, images and audio at the same time which will lead to incredible new use cases !

Other alternatives to OpenAI exist, either retraining and fine-tuning gpt3 or creating new models like Google’s PaLM, china’s huge Wu Dao or GPT-Neo open source alternative.
At Eden AI, we’re working on bringing the capabilities of all these different models to your application using one simple unified APIs with a very intuitive developer portal.

AI APIs for Web Devs - Concrete applications

SamyMe — Mon, 19 Dec 2022 16:38:24 +0000

Have you ever wondered how you, as a web developer, can grasp the power of AI ?

But first, what’s AI ?

Artificial Intelligence is about allowing a machine to mimic human behavior. Computers were made to do that by the power of elaborated algorithms. Now, machine and deep learning make it possible to teach computers with data what we fail to describe in an algorithmic approach. This opens the way for computers to understand and manipulating non structured data :

📖 Reading text and understanding concepts like : the topic of the text, the sentiment of the writer and the key concepts it talks about
🗣️ Hearing your voice, understanding the words you say and responding to you (Siri and Alexa : speech to text and text to speech)
🎥 Looking at an image or watching a video and understanding the big picture and the details of what it contains

But in order to achieve these objectives, you have to train large models with millions of labeled data.
Now, with AI as a Service, a lot of companies provide off-the-shelf trained models that you can access directly through an API. These companies are either the tech giants (Google, Microsoft , Amazon) or other smaller, more specialized companies, and there are hundreds of them. Some of the most known are : DeepL (translation), OpenAI (text and image analysis), AssemblyAI (speech analysis). If only there was one global API for accessing all these models…

Usecases :

The very easy access to this intelligence through simple APIs lead to the creation of tons of use cases :

Automatically parse unstructured documents into a structured form : If your app handles scanned documents or pdfs (example: resumes, invoices, ID documents …etc), you can parse them automatically to extract the data it contains and store it into a structured DB for easier manipulation.


import json
import requests

headers = {"Authorization": "Bearer 🔑 Your_API_Key"}

url="https://api.edenai.run/v2/ocr/identity_parser"
data={"providers": "amazon,microsoft,mindee"}
files = {'file': open("🖼️ path/to/your/image.png",'rb')}

response = requests.post(url, data=data, files=files, headers=headers)

result = json.loads(response.text)
print(result['amazon'])
print(result['mindee'])

Reaching out more people by automatically translating your contents : There are tons of great text translation and language detection apis. Their performance varies depending on the language pair you need.


import json
import requests

headers = {"Authorization": "Bearer 🔑 Your_API_Key"}

url ="https://api.edenai.run/v2/translation/automatic_translation"
payload={"providers": "google,deepl", 
         "source_language":"en", 
         "target_language":"fr", 
         "text": "This is a new test."}

response = requests.post(url, json=payload, headers=headers)

result = json.loads(response.text)
print(result['google']['text'])
print(result['deepl']['text'])

Analyze users' feedback and concerns by (semantically) understanding textual content : text classification, extracting key concepts, and understanding the writer’s sentiment . Natural language analysis can have tons of other applications.

import json
import requests

headers = {"Authorization": "Bearer 🔑 Your_API_Key"}

url ="https://api.edenai.run/v2/text/sentiment_analysis"
payload={"providers": "emvista,microsoft", 
        'language': "en", 
        'text': "this is a great test"}

response = requests.post(url, json=payload, headers=headers)

result = json.loads(response.text)
print(result['emvista']['items'])
print(result['microsoft']['items'])

Images and videos tagging and classification : for easier storage and retrieval.


import json
import requests

headers = {"Authorization": "Bearer 🔑 Your_API_Key"}

url = "https://api.edenai.run/v2/image/object_detection"                 
data={"providers": "clarifai,amazon"}
files = {'file': open("🖼️ path/to/your/image.png",'rb')}

response = requests.post(url, data=data, files=files, headers=headers)
result = json.loads(response.text)
print(result['clarifai']['items'])
print(result['amazon']['items'])

Content moderation : You can detect explicit content present in an image or a video, detect racist messages or insults in a given text.

import json
import requests

headers = {"Authorization": "Bearer 🔑 Your_API_Key"}

url=" https://api.edenai.run/v2/image/explicit_content"
data={"providers": "picpurify,microsoft"}
files = {'file': open("🖼️ path/to/your/image.png",'rb')}

response = requests.post(url, data=data, files=files, headers=headers)

result = json.loads(response.text)
print(result['picpurify']['items'])
print(result['microsoft']['items'])

Data anonymization : You can remove Personally identifiable information (PII) information from a text before making it public. Same thing with an image by blurring faces and license plates for example.


import json
import requests

headers = {"Authorization": "Bearer 🔑 Your_API_Key"}

url ="https://api.edenai.run/v2/text/anonymization"
payload={"providers": "openai", 
         "language": "en", "text": "My name is Jeremy and this is a test"}

response = requests.post(url, json=payload, headers=headers)

result = json.loads(response.text)
print(result['openai']['result'])

Work with audio content : You can easily transform an audio or a microphone input to a text and analyze it's content. The inverse would be text to speech which gives your app a voice to interact in new ways with your users.

Conclusion :

As said earlier, there are tons of companies producing new APIs each day. You can find a great variety of performances at different prices. So, at Eden AI, we want to give access to all the power of AI through one standardized API. One interface for you not to bother about how to access intelligence, and only focus on inventing great apps out of it !