Eric J Ma's Website

In this blog post, I share my experience running the Mixtral 8x7b-Instruct model on my old Linux GPU tower. I used the 4-bit quantized model and was pleasantly surprised that it worked. I generated keywords for a paper on protein engineering and machine learning using the model, and the results were comparable to GPT-4. Although the model was slower than running mistral-7b, it was still functional on older hardware. Have you tried running large language models on older hardware? Read on to find out more about my experience.

In this blog post, I share insights from a career panel at MIT where we discussed advice for Ph.D. students about to graduate. We covered the importance of studying job postings, networking effectively, maintaining a public profile, and understanding business needs when applying for jobs. We also touched on the value of a publicly viewable portfolio and the challenge of balancing work and home life. What other advice have you heard?

In this blog post, I share advice for those starting a new job, focusing on the first 90 days. I discuss the importance of automating your calendar, recording your accomplishments, building a company committee, and choosing a manageable project with tangible impact. These strategies can help you gain control over your career direction and make a positive impression in your new role. How can you apply these tips to your own career journey?

In this blog post, I share my strategies for maintaining technical skills as a data science team lead. Balancing management duties with technical tasks, I use strategies like performing lower-level tasks, pair coding, code reviews, asking questions, and prototyping. These methods help me stay sharp and credible with my team, while also fostering personal growth. How do you keep your technical skills up-to-date?

In this blog post, I share how I breathed new life into my idle GPU tower by running an Ollama server on my home's private network. I connected all my devices via a Tailscale virtual private network and installed Ollama on my GPU server. I then used LlamaBot to build bots that utilized the Ollama server. This turned out to be an effective way to extend the usable life of my GPU box. Curious about how you can do the same with your idle GPU? Read on!

In this blog post, I discuss the overlap between dashboard-ready and machine-learning-ready data. I share an example from a chemical screening campaign, where the same data used for a dashboard can also be used for machine learning models. I explore the reasons behind this from both a statistical and business perspective. How can you gain leverage in your data for both purposes?

In this blog post, I shared my insights from the SLAS 2024 conference on how a data science team can deliver long-lasting impact in a biotech research setting. I discussed the importance of bounding work, identifying high-value use cases, possessing technical and interpersonal skills, and having the right surrounding context. I also shared some personal experiences and lessons learned from my work at Moderna. How can these insights help your data science team become successful agents of high-value delivery? Read on to find out!

In this blog post, I explore the rapidly evolving landscape of large language model (LLM) tooling, discussing APIs, self-hosting, API switchboards, Python-based LLM Application SDKs, vector-based retrieval, prompt experimentation, evaluation, UI builders, and command-line interfaces. I share my experiences building LlamaBot and offer principles for making smart tech stack choices in this ever-changing field. How can you navigate this dynamic ecosystem and make the best decisions for your LLM projects? Read on to find out!

In this blog post, I challenge the traditional approach to exploratory data analysis (EDA) in data science. I argue that EDA should be directed and purposeful, not aimless. I share key principles for effective EDA, including falsifying our assumptions, having a clear end purpose, and embracing iteration when purposes are invalidated. I also emphasize the importance of practice and domain expertise in developing this skill. How can we make EDA more purposeful and effective in our data science work? Read on to find out!

In this blog post, I debunked the misconception that embedding models must match the text generation model in retrieval augmented generation (RAG). I explained how these models are decoupled, with the choice of embedding affecting only the quality of content retrieved, not the text generation. I also shared my preference for SentenceTransformer due to its cost-effectiveness and performance. Finally, I updated LlamaBot to reflect this understanding, allowing for more flexible model composition. Curious about how this could change your approach to RAG? Read on!