Ever since Gen AI grew to become a buzzword, many tech firms are whitewashing their merchandise as Gen AI. It appears each startup has pivoted to embrace this development, particularly prior to now 12 months after the discharge of Chat GPT. As a researcher having labored at Lyft I’ve seen the evolution from AI/ML fashions to Gen AI. It’s shocking how usually I hear or learn pitches the place merchandise primarily based on commonplace AI/ML, and even hard-coded rule-based workflows, are being labeled as Gen AI.
As somebody who labored at Lyft the place AI/ML fashions have been important to how the enterprise was run from rider-driver matching, pricing to ETA estimation to routing and many others, I really feel that it’s necessary to grasp what are the similarities between Gen AI and AI/ML fashions. If you wish to study extra about AI/ML fashions which were utilized in rideshare, try the blog from Lyft’s Information Science crew.
This publish is focused in direction of a non-technical viewers, and it’s meant to be a primer to grasp Gen AI and primary variations between AI/ML and Gen AI.
Let’s break it down in easy phrases.
What’s Generative AI?
- Creation of latest information is vital to Gen AI. If no new content material is created then it’s simply any clever system not Gen AI. The brand new content material that’s created may very well be textual content (ChatGPT), photographs (Midjourney), video (Runway), and extra. Consider it like an artist that learns from present works after which creates one thing fully new. Most Gen AI fashions are skilled on Giant Language Fashions (LLMs) utilizing a course of known as unsupervised studying on huge quantities of textual content information from the Web, books, articles and many others. If you wish to study extra on this topic, try this video by Andrej Karpathy who labored at Open AI.
- Transformers: Gen AI makes use of superior fashions known as transformers that could be a kind of deep studying structure launched within the paper “Attention is All You Need” by Vaswani et al. in 2017. Vaswani and his crew labored at Google DeepMind after they created this structure. It has revolutionized pure language processing (NLP) and different fields by enabling the coaching of very massive fashions effectively. The most important breakthrough of a transformer mannequin is the self-attention mechanism that helps perceive context extra successfully by weighing the significance of phrases no matter their place in a sentence. Allow us to take the sentence: “Mani, who loves know-how, lives within the San Francisco Bay Space.” Transformers can perceive that “Mani” “lives” and “Bay space” are key phrases within the sentence. Though “who loves know-how” is within the center, the mannequin connects “Mani” and “San Francisco Bay Space” as a result of they offer us the important thing details about Mani’s location.
- RAG: Gen AI fashions usually depend on Retrieval-Augmented Generative (RAG) to cut back hallucinations which is a flowery time period for Gen AI producing believable however incorrect info. So subsequent time whenever you write a immediate specify that you simply don’t need content material that’s false. Why is RAG necessary? RAG is required as a result of it supplies extra correct and dependable outcomes that reduces hallucinations by retrieving updated and related information (it may be from an organization’s personal LLM and up-to-date info from the Web) on which the LLM was not skilled.
What are Conventional AI/ML Fashions?
Conventional AI/ML fashions are extra like analysts. They classify, predict, or analyze information fairly than create new content material. They’re utilized in areas like fraud detection, advice programs, and predictive upkeep. They’re constructed to make selections primarily based on patterns within the information they’ve been skilled on.
- Evaluation of Current Information: These fashions concentrate on duties like predicting outcomes primarily based on earlier behavioral information. (e.g., recommending kind of trip (precedence, Lux, Shared and many others primarily based on riders earlier habits) or classifying information (e.g., detecting rider fraud primarily based on a number of accounts utilizing the identical cost info). The Lyft DS crew has a blog on how they use the advice engine in case you are focused on studying extra.
- Leverage Statistical Fashions: Conventional AI contains algorithms like resolution bushes, logistic regression, and neural networks. Whereas all these strategies are highly effective and predictive, they don’t have the artistic talents of Gen AI.
Conclusion
In conclusion, one can argue that Gen AI fashions are constructed on basic AI/ML ideas resembling neural networks and deep studying. However there are key variations in output and mannequin structure. Conventional AI/ML fashions use less complicated algorithms that are constructed on statistical strategies like regression, K-means clustering or random forest, Gen AI makes use of complicated transformer structure. Conventional fashions usually require structured information like spreadsheets whereas Gen AI can work with unstructured information like written textual content.
So subsequent time when somebody discusses Gen AI options that they’re engaged on, ask them these important questions: what’s the underlying structure, what new content material is being created, what’s the LLM that the mannequin is skilled on, are they utilizing RAG to cut back hallucination?
PS:
In case you are focused on attempting new Gen AI instruments try these firms: Perplexity (Information Engine utilizing LLM) Inflection (Like ChatGPT) Anthropic (Like ChatGPT) Groq (Quick AI processing service) Runway (Video) Beatoven (Music) Tavily (analysis) Midjourney (photographs)