Agustin Caverzasi, AI and how to leverage it
Hosts: Mak Gutierrez (Magma), Agustin Caverzasi (Founder of AnyoneAI)
Agustin’s path to starting Anyone AI:
Dropped out from PhD in 2014 to build a Deep Vision AI, startup in computer vision space.
Deepvision AI built a team of 30 AI experts. Provided visual intelligence services for partners. Exit in 2020.
Trained many people in AI, realized the challenge of training people in AI, and the talent shortage in the industry.
• He started Anyone AI in 2021.
• AnyoneAI invests in software developers in latam, trains them and connect them with companies in the space:
- 1500 applicants per month, they accept 1-2% per month.
- Two models: 4 months full time program, 6 months part-time
- Very intensive program with the idea of learning by doing. They give students real problems that real companies give them.
The basics: what are the terms we should know?
- Supervised learning: You give a model annotated data with the correct answers. The model optimizes parameters to minimize output error with respect to the training data given.
- Unsupervised learning: Uses deep learning to arrive at conclusions and patterns through unlabeled training data. Example is ChatGPT. It is trained to predict the next word in a text. Based on Transformer (paper from Google in 2017). Transformer was built to predict the next word in a text. With a corpus of data (social media, websites, etc), they use Mass Language Modeling. You take some words, mask the key terms, and the algorithm trains on that data. The model automatically learns to predict these words. The goal is to provide contextually relevant text to user prompts.
Evolution from Original Machine Learning to LLM, AGI, Image creation:
AI is going from mostly supervised learning to mostly unsupervised. Unsupervised is much less expensive to train since it needs no annotated datasets.
Examples of supervised learning capabilities:
- Image classification, sales forecasting, risk assessment.
Examples of supervised learning capabilities:
- ChatGPT: the self-supervised algo is what they are doing. Can infer knowledge, goals, summarize, provide assistance, recommendations, can do more advanced tasks.
How are these being used in real life? Use cases:
- Integrating models to internal automation tools (Zapier, Bardeen, etc) to increase productivity levels.
- Sales: Helping with the sales outreach process, classifying leads, creating personalized messages for potential customers.
- Pricing models. Eg: Delivery apps that have to sell a combo deal. Price in Cordoba should be very different from price in a small village 20km away. With AI, you can take a lot of variables into consideration and optimize pricing.
- Marketplaces: Route optimization for the riders.
- Fintech: Fraud detection, credit risk analysis, risk management. Eg: understanding probability of default.
- Physical world: Deploying a model to detect anomalies rather than putting people in security in front of screens.
When can you start to leverage AI? When does it make sense?
- If you've been using data a lot, and things are taking a lot of man power, that is probably the time you could leverage AI
- You should start by gathering the most amount of data possible. It’s very hard to know what data will help you solve problems in the future.
What are the steps to start using AI? How do I as a founder start to think about implementing it into my startup?
There are many large pre-trained models. You should leverage those and fine tune them with data. No need to create models from scratch.
- Collect all the data possible: It’s very hard to know what data will help you solve problems in the future.
- Create data infrastructure.
- Perform EDA (exploratory data analysis).
- Start with simpler machine learning models.
- Go to more complex deep learning models.
Watchout: Don’t send models to production too early. They might have big error rates. Best implementations are AI models with humans in the loop. AI is an assistant to the person making the final decision.
How Anyone AI is leveraging AI to run more efficient processes:
- Understanding quality of linkedin profiles and level of attractiveness in the market.
- AI transcription of meetings. Summaries and action items.
- Sales: Getting linkedin profiles, making personalized messages for automated massive outreach.
- Models to make better selection for their programs based on applicant backgrounds.
How to build a team to maximize AI in my startup:
- Depends on the stage of the company and how large your organization is.
- Independent ML teams: Mostly for applied machine learning and research teams. Then feed engineers team
- Mixed teams: Full stack or back end developers working with AI or Machine learning developers
- Starting small: Start with a data analyst doing simple reports, then add ML engineers to do more complex analysis.
Where can I find the best AI tools to leverage today?
