- Blog
- Pathways to Becoming an AI/ML Product Manager
Pathways to Becoming an AI/ML Product Manager
- By Leslie Cook
- Posted on
Welcome to our Product Newsletter, a biweekly email highlighting top discussions, and learning resources for product managers.
What We Will Cover In This Edition:-
Top Discussions:
1) In practice, how is artificial intelligence actually used by companies?
2) Getting ready for the role of AI/ML PM
3) How does one become an AI/ML PM?
Top Learning Resources:
1. What is a neural network? Chapter 1, Deep Learning
2. Gradient descent, how neural networks learn. Chapter 2, Deep Learning
3. What is backpropagation really doing? Chapter 3, Deep Learning
4. Backpropagation calculus. Chapter 4, Deep Learning
5. Intro to Large Language Models
__________________________
Top Discussions
Question 1) In practice, how is artificial intelligence actually used by companies?
Hi there, I want to talk about artificial intelligence (AI) in particular, despite all the hoopla, I don’t see a lot of companies using it.
Could you kindly tell us (roughly) how AI was employed if you have worked on an AI project or product?
Thanks in advance.
– Amy Walker
Discussion
A] In reality, companies are increasingly leveraging artificial intelligence to gain a competitive edge and improve various aspects of their operations. AI is being used in diverse industries, such as healthcare, finance, retail, and manufacturing, to automate processes, enhance customer experiences, optimize supply chains, and make data-driven decisions.
However, implementing AI solutions requires careful planning and understanding of the specific business needs and challenges. It is important for companies to invest in the necessary expertise and resources to effectively harness the power of AI for their unique business goals. This includes hiring AI experts or partnering with AI-driven companies to develop and customize AI algorithms and applications that align with their objectives. Moreover, companies should ensure that they have a robust infrastructure in place to collect, store, and analyze the vast amount of data required for AI operations.
Additionally, businesses must consider the ethical implications of AI, including ensuring fairness, transparency, and accountability in its decision-making processes, and protecting customer privacy and data security. By addressing these considerations, companies can successfully integrate AI into their operations and gain a competitive edge in today’s fast-paced and data-driven business landscape.
– Naomi Nwosu
B] AI has been employed in various industries and sectors, such as healthcare, finance, and retail. In healthcare, AI has been used to analyze medical data and assist in diagnosing diseases, leading to more accurate and timely treatments. In finance, AI algorithms have been utilized for fraud detection and risk assessment, improving the overall security of financial transactions.
Additionally, in the retail industry, AI-powered recommendation systems have enhanced customer experiences by personalizing product suggestions based on individual preferences. These examples highlight the diverse applications of AI across various industries, demonstrating its ability to revolutionize processes and drive innovation. Moreover, in the field of transportation, AI has been employed to optimize traffic flow and reduce congestion, ultimately improving commuting times and enhancing overall efficiency.
Furthermore, AI has also made significant strides in the field of agriculture, where it has been used to monitor crop health, predict weather patterns, and optimize irrigation schedules, enabling farmers to make informed decisions and maximize their yields. These diverse applications of AI underscore its potential to transform and improve countless sectors of the economy.
– Tina Greist
C] Within the EdTech industry, artificial intelligence (AI) is utilized to grade students’ written answers to prompts. Additionally, this same technology provides students with real-time feedback on their written responses, allowing them to refine their responses further. This innovative use of AI in grading and providing feedback not only saves teachers’ time but also ensures consistent and objective evaluation.
Moreover, the integration of AI technology in education fosters personalized learning experiences by tailoring feedback to individual students’ needs and helping them improve their writing skills effectively. As AI technology continues to advance, it has the potential to revolutionize education by providing even more personalized and adaptive learning experiences.
For example, AI chatbots can simulate conversations and provide immediate language practice for students, enhancing their speaking and listening skills. Additionally, virtual reality (VR) technology can create immersive learning environments, allowing students to explore subjects like history and science in a more engaging and interactive way. The integration of AI in education holds great promise for empowering students and teachers alike to reach their full potential.
– Nathan Endicott
__________________________
Question 2) Getting ready for the role of AI/ML PM
Unlike many other disruptive technologies (like virtual reality), AI is growing too large to ignore.
It appears that all college students are currently studying AI, either for coursework or research. I’ve had conversations with master’s degree holders in chemical engineering and data science, among other fields.
How can I become a forward-thinking tech PM by learning enough technical material about AI now?
I’m thinking of suggestions for free or paid courses, news sources to follow, or both, but anything is helpful. I truly don’t want to return to school. I’ve earned three degrees already. I assumed a six-month program would be sufficient for me.
What do you think guys?
– Michelle Plowman
Discussion
A] PM currently, former ML engineer. I would continue to concentrate only on the user if I were you. It’s just as difficult to define the interface for human interaction with these technologies as it is to advance AI research. The success of AI systems heavily depends on how well they can understand and cater to the needs of users. By prioritizing user-centric design, you can ensure that AI technologies are not only advanced but also effectively serve their intended purpose. Additionally, understanding user behavior and preferences can provide valuable insights for further enhancing AI research and development.
Learn a little bit about unsupervised learning, regression, and classification at a high level before considering how this might lead to improved experiences. Even for me, it’s not a simple task, but that’s where time is best spent. By familiarizing yourself with unsupervised learning, regression, and classification, you can gain a deeper understanding of how AI algorithms work and their potential applications. This knowledge will enable you to make informed decisions when it comes to designing AI systems that deliver more accurate and personalized experiences for users. Investing time in learning these concepts will ultimately pay off in creating more effective and efficient AI technologies.
– Dhiraj Mehta
B] Absolutely agreed. As a general AI PM, I frequently advise other PMs that experience is more important than parameters or tokens. Since everyone will ultimately have what they call the “best” LLM, experience is our opportunity. Experience allows PMs to understand the nuances and complexities of managing AI projects, enabling them to make informed decisions and navigate challenges effectively. It helps in identifying potential pitfalls, anticipating user needs, and ensuring the successful implementation of AI solutions. Additionally, experience equips PMs with valuable insights that cannot be solely derived from parameters or tokens, ultimately contributing to the overall success of their projects.
– Gerard Kolan
C] If you’re more interested in high-level information than in actually creating and deploying the models, Stratechery and Lenny have some excellent articles about the present state of the LLM explosion. These articles provide valuable insights into the current landscape of machine learning and its impact on various industries. They offer a broader perspective that can complement the technical knowledge gained from the Google Crash Course. Additionally, staying informed about the latest developments in the field can help you make more informed decisions regarding your solution’s fit within the rapidly evolving machine learning ecosystem.
A course on statistics introduction would be very beneficial. Understanding statistics is crucial for effectively analyzing and interpreting data in the field of machine learning. It provides a solid foundation for making accurate predictions and understanding the significance of results. By incorporating a course on statistics, you can enhance your understanding of the underlying principles and improve your ability to make informed decisions when developing machine learning solutions.
– Lawrence Martin
__________________________
Question 3) How does one become an AI/ML PM?
I want to start learning more about AI and ML because I’ve been interested in them. I believe that gaining knowledge in AI and ML will not only satisfy my curiosity but also open up new opportunities for me in the future. Additionally, understanding these technologies can help me stay updated with the latest advancements and contribute to solving real-world problems using AI and ML techniques.
Which challenges do AI/ML PMs answer, and what use cases do they create?
What prerequisites exist for PMs working in the AI field?
What are the general features of your roadmaps?
– Flavia Bergstein
Discussion
A] At the established tech company I work for, I am currently the PM for an ML algorithm (think Search, Recommendations, Relevance, Risk/Fraud Detection, etc).
The use cases aspect interests me since, in contrast to “normal” PMs, I have a lot of difficulty with it. Because it’s a mature organization, the algorithm’s functionality was established before I joined the team as PM, and our goal result metric is well specified. However, the issue arises when various stakeholders want the algorithm to do different things (for example, suppose the reddit ranking algorithm was optimized for “most entertaining” and now a stakeholder requests that it show new posts or self-posts more effectively). My team is trying to either push back against this or must perform a great deal of analysis work to determine whether or not there is a problem. For me, a data science relationship is far more important to product discovery than traditional user interviews.
Prior to joining my algorithmic team, the majority of PMs had experience managing another algorithmic product. Initially, I pursued a career in data analysis (on the DS track rather than the business track). However, I ultimately concluded that I wasn’t interested in DS enough to pursue a PhD, so I obtained my MBA and transitioned to project management. Having the background in DS is incredibly helpful since it allows me to discuss the model with my DSes in an informed manner and explain the changes I would like to see made (e.g., modifying the loss function, Thompson sampling). Additionally, I can spot when someone is trying to con me.
The roadmap is basically a list of ways to improve algorithms. For example, adding new data from team x to see if it improves $targetMetric, changing the algorithm target to see if it improves $relatedMetric while keeping $targetMetric the same, and so on. Additionally, some tooling work (e.g., pipeline hardening; alarm if data from team Z is absent) or upgrading the backtesting package to include x metric are needed.
– Pankaj Jain
B] Is it necessary for a non-technical PM to become as technical as a data scientist in order to become an AI/ML PM?
– Marty Ross
C] You should be aware of the components of the overall algorithm, including the data pipeline, the prediction model itself (including its training and scoring cycle), the model’s optimization and generalization to actionable units, and the deployment and push pipeline.
To learn about the inner workings of the algorithm, the significance of data collection and cleanup (and the value of a high-quality data pipeline), and the fundamentals of machine learning and neural networks, I would suggest enrolling in some DS-focused online courses. A possible place to start would be Google Developers, which offers some ML crash courses.
Disclaimer: That course covers a wide range of topics, including machine learning engineering, but I haven’t taken it.
– Bobby Duncan
__________________________
Top Learning Resources
The video serves as an introductory guide to understanding neural networks in deep learning. It delves into the fundamental concepts and components of neural networks, addressing queries such as the nature of neurons, the significance of layers within these networks, and the mathematical principles underlying their functionality. Overall, the tutorial serves as a comprehensive primer for beginners seeking to understand the foundational aspects of neural networks, offering insights into their structure, mathematical foundations, and practical applications through various examples and explanations.
The video elucidates the pivotal role of gradient descent in enhancing neural networks’ learning efficiency, showcasing its function in parameterizing a handwritten digit recognition network through a cost function and weights/biases evaluation. It emphasizes the necessity of a smooth cost function output for effective minimization using gradient descent, elucidating the algorithm’s mechanics in finding local minima. The discussion extends to research findings on neural networks trained with random versus correct labels, underlining the importance of accurately labeled datasets. Ultimately, the video underscores the effectiveness of gradient descent in optimizing neural network training while briefly mentioning the support Amplifi partners offer to early-stage company founders.
The video explores backpropagation as the underlying algorithm facilitating neural network learning. It operates by computing the gradient of the cost function, sensitive to the network’s weights and biases, subsequently utilizing this gradient to adjust these parameters. It’s highlighted as a supervised learning algorithm, boosting neuron activation in deep learning networks by modifying weights and biases in the preceding layers proportionally to their respective values. Backpropagation facilitates the propagation of desired alterations in the weights and biases of subsequent layers. Additionally, it’s described as a stochastic gradient descent algorithm, enhancing computational speed by updating weights and biases based on mini-batches of data rather than the entire dataset, converging towards local minima of the cost function efficiently.
The video delves into the calculus behind backpropagation in deep learning, focusing on a straightforward network with one neuron per layer defined by weights and biases. It introduces the chain rule to comprehend the impact of minor weight changes on the cost function, determining sensitivity via derivatives of the cost, activation function, and weighted sum. Sensitivity calculations involve iteratively determining the impact of prior activations in the chain rule expansion to compute sensitivity to preceding weights and biases. This concept remains consistent even with multiple neurons in a layer, necessitating additional indexing to track weights’ positions within the layer. The section concludes by discussing sensitivity in terms of neurons firing together and wiring together, emphasizing the derivative’s averaging across all training examples for accuracy.
The hour-long talk introduces large language models, composed of parameters and run files, self-contained systems operable offline. It delves into model training, employing vast text data for compression, and the primary goal of predicting the next words for data compression. Highlighting pre-training on internet data and fine-tuning for task specificity, it outlines model capabilities and stages in detail. The video also discusses challenges like security risks—prompt injections, shield-breaking, and data poisoning—and emphasizes the ongoing development of defenses against these threats. Additionally, it touches upon the models’ dependence on data volume and computational resources, driving a computing gold rush for more extensive, powerful models.
__________________________
If you enjoyed this newsletter, please consider sharing it with a friend by asking them to sign up here.
Who’s Prowess? We are optimist product managers, engineers, and educators working on creating a world where merit meets opportunity. On Prowess, aspiring and experienced product managers showcase skills, learn from the community, and connect with employers to advance their careers.
How can you grow with Prowess?
-
- Learn from curated learning resources and community
- Work on curated projects or join expert-guided group projects
- Receive personalized feedback from product leaders
- Build a portfolio to stand out as a product manager
- Access and apply to curated jobs
- Prepare for interviews with Q&A from top companies
