Skip to content

Latest commit

 

History

History
76 lines (70 loc) · 4.29 KB

2024-08-27.md

File metadata and controls

76 lines (70 loc) · 4.29 KB

AI for Software Engineers (Part 1): Deep Learning & Neural Networks

Details

Join Codesmith CEO & Co-Founder, Will Sentance, for a high-level map of deep learning and neural networks where we will dive into the following:

  • Prediction: Datasets, weights & forward propagation
  • Learning: Gradient descent
  • Backpropagation

https://linkedin.com/in/willsentance https://linkedin.com/in/mcphail-alex

Notes

  • The full stack is expanding to include productized prediction
    • Monitoring user behavior
    • Looking for fraudulent refund claims or legit ones
      • Predict via prior data so automatic responses can be implemented
    • Contributing to Tensorflow (2nd most popular interface to neural networks)
      • Some of this was built by CodeSmith students
  • Cognitive software engineer is another name for what they're covering today
    • Distince from ML engineer somehow
  • Other trends
    • Stacastic/nuanced software AI
    • Hardware questions have become software questions (like building 5G infrastructure, robotics)
    • More of our default mode of operating is now online
  • Software is taking it's next stage in the centrality of our lives
    • What are the pieces we can invest in to be on the cutting edge of this
    • Interest rates are dropping now, in the next 3 months there will be more spending on software
      • The past 2-3 years shifted the dynamics of the industry
      • Firms that weren't tech firms have invested in it (finance, healthcare)
        • Leaders from tech companies have roles in these companies
        • Were brought in from outside to lead
        • No longer outsourcing tech, have more internal candidates now
      • The tools that are available for these companies are becoming more suited
        • Many are AI tools
        • Can increase the output of your programmer teams using these nuanced tools
          • Ex: software can provide legal advice
          • Ex: software can help provide care to patients by finding providers
  • What does this mean for software engineers
    • Five things CodeSmith looks for
      • Capacities, problem solving, technical communication
    • Look for ways to infuse software into places where it wasn't being used before
    • Engineering approach/principles have changed
      • Additonal mental models (model pipelines, LLMs)
      • The nature of prediction and statistics
    • Domain knowledge
      • Needs to stay up-to-date
      • Model/data pipeline, API/Model deployment, infrastructure, UX, LangChain
      • Have to be able to work with machine learing engineers
  • It's all prediction
    • Taking prior patterns that we've taken from our data (historic data patterns)
    • Using that as probable identifiers of future behavior
    • Using datasets to make predictions, narrowing down the sample over time
      • Updating datasets if customer behaviors change over time
    • Using this data to make automated responses (like for refund cases to protect from fraud)
    • The model training part of this is working out what is the best patterns
    • This all translated directly into prediction using neural networks
  • Neural networks predictiveness is based on brain patterns
    • Have been around since the 1940s - 1950s
    • Computational power has made neural networks dominate/outperform other approaches
    • They are the technique behind LLMs
    • They're now analagous with how the brain processes data
    • Backpropagation is the algo that says we can reweight and make new predictions
      • It doesn't mirror how the brain works, but still has a lot in common
  • Predictions that humans make have to be reflected in data for computers to mimic them
    • This included image data, like the dataset that the USPS used in the 80s to read handwriting
      • Many values of greyscale recognizing nuances of human handwriting
    • Use subsets of weighted data to build out samples
    • Single-layer neural network represented by a grid with weights for the data
    • Neural networks become powerful when they use multiple layers
      • Capture sub-shapes or "edges"
  • Prediction tool has to be integrated into a product or its just research
  • Rise of alternative tools to improve productivity as software engineers
  • Statistics and prediction are a different model from software engineering