Why Software Engineers Need to Understand Machine Learning
And how ML helps software engineers in their daily work
The biggest concern I hear vocalized about AI is that it’ll take someone’s job. I see this disproportionately often in the software engineering community where we see tools like Devin and chat bots like ChatGPT, Gemini, and Claude that can code impressively well. Here’s my take: we hear about AI taking the job of software engineers because we are the community that is up-to-date on AI advancements and interested in them.
My response to any software engineer concerned about losing their job to AI is to understand machine learning and how to use it. As AI becomes more capable the role of a software engineer will become less about code and more about efficiently solving problems. These problems will be solved with machine learning.
Software engineers are taught to think mathematically when solving problems. Conditionals and assertions are used to create a set of logical rules for programs to execute. Machine learning turns problem solving from a logical science to a natural science where statistics and experimentation are used to understand an uncertain scenario.
Let’s use an example: The problem you need to solve is taking video footage or an image and identifying the facial expressions of the people within it. Defining a set of logical rules for a program to accomplish this task would be very difficult and time-consuming. With machine learning, we can provide a program the data needed to understand the task and it can determine those rules for itself. We can use a subset of machine learning called computer vision to make this task much more trivial.
An understanding of machine learning helps software engineers think like a scientist. This is will become the way software engineers solve problems.
Using machine learning in their tasks also provides the following benefits for software engineers:
It can help reduce time spent programming: There are tasks that can be solved quickly with an out-of-the-box machine learning solution that would take months with a traditional software engineering approach. A great example of this is the computer vision task mentioned above.
It can make it easy to customize products: Machine learning changes the approach to iterating on and customizing products from rewriting code to recollecting data. A great example of this is adapting a product to a different language by re-training it on data in that language.
If you’re interested in understanding machine learning, I’ve set up a road map to help anyone learn ML entirely for free. It includes prerequisites, machine learning math, and more advanced topics. Make sure to star the repo to keep tabs on it as I’ll be further streamlining it soon.
Machine Learning Resources and Updates
Here’s today’s list of machine learning resources and updates. Don’t forget to follow me on X for more updates:
Processing 2 Billion Images for Stable Diffusion Model Training - Definitive Guides with Ray Series
The overwhelmed person’s guide to Google Cloud: week of Oct 23
Interesting experiment from Google that creates an NPR-like discussion about...
Computational cost reduced by 350X for datacenter-scale AI by @nvidia...
AI, AI and more AI: A deep dive into Nvidia’s announcements at Computex 2024
Nvidia announces Project G-Assist, an AI assistant for gaming with GeForce GPUs
Sakana AI is proud to sponsor the LLM Merging Competition:...
LLM Research Insights: Instruction Masking and New LoRA Finetuning Experiments
LLM Merging Competition: Building LLMs Efficiently through Merging
FineWeb: decanting the web for the finest text data at scale
LLM Research Insights: Instruction Masking and New LoRA Finetuning Experiments
Here's my conversation with Roman Yampolskiy (@romanyam), AI safety researcher...
OpenAI revives its robotic research team, plans to build dedicated AI
UPDATE: Sam Altman finally responds to Helen Toner's revelations from...
Processing 2 Billion Images for Stable Diffusion Model Training - Definitive Guides with Ray Series
This guide teaches you how to build scalable data pipelines for training Stable Diffusion models using Ray. It explains how to download and process the LAION Aesthetics dataset, including image and text transformations. You learn to encode data efficiently with GPUs and optimize performance and cost.
This is an underappreciated point
People are either overly pessimistic or overly optimistic about AI's future impact on their lives. Current AI won't replace the need for long-term financial planning like saving for retirement. Despite high productivity, scarcity issues like housing still exist in the US.
sammcj/gollama
Gollama is a Go-based client for managing Ollama models with an interactive TUI for listing, sorting, selecting, and deleting models. It can link Ollama models to LM-Studio and provides various configuration and logging options. Installation involves downloading from GitHub, and contributions are welcome under the MIT License.
The overwhelmed person’s guide to Google Cloud: week of Oct 23
This blog curates the best updates from Google Cloud. Highlights include a new IP indemnity policy for AI, Cloud Spanner's upgrades, and Google's defense against a massive DDoS attack. Join the Google Cloud Innovators for weekly updates and exclusive content.
The Top ML Papers of the Week (May 27 -...
This week's top ML papers introduce innovative methods like Contextual Position Encoding (CoPE) for transformers and Symbolic Chain-of-Thought for logical reasoning in LLMs. Abacus Embeddings achieve high accuracy in arithmetic problems, while GNN-RAG combines LLMs and GNNs for improved question answering. Additionally, new advances include a vision-language model introduction, a parallelizable attention mechanism, and a multilingual language model, Aya23.
Interesting experiment from Google that creates an NPR-like discussion about...
Google has created an experiment that generates NPR-like discussions about academic papers. This tool offers exciting possibilities for science communication. The voices and pauses make it sound just like public radio.
Computational cost reduced by 350X for datacenter-scale AI by @nvidia...
Computational cost reduced by 350X for datacenter-scale AI by @nvidia over the last 8 years. 🤯
From Jensen Huang Keynote at COMPUTEX 2024 finished just now.
Chips Act is attracting insane amount of investment
The Chips Act is attracting massive investment in US electronics manufacturing. This year's investment may surpass the total from 1996 to 2020. The advanced manufacturing tax credit could cost much more than the estimated $24 billion.
Build Your First AI Agent in 5 Easy Steps (100% local)
The article explains how to create AI agents using CrewAI and Ollama to run them locally. It details setting up a local LLM, creating tools, defining agents, assigning tasks, and running the agents to read a PDF, write a blog post, and generate a title.
AI, AI and more AI: A deep dive into Nvidia’s announcements at Computex 2024
Nvidia announced new AI advancements at Computex 2024, including AI laptops, Blackwell platform, and Spectrum-X Ethernet network, shaping the future of AI and computing.
Teaching LLMs to Express Confidence
"SaySelf" trains language models to express accurate confidence and self-reflective reasons. It uses supervised fine-tuning and reinforcement learning to improve predictions and reduce overconfidence. The method keeps performance stable and provides insightful rationales.
A good short blog post on how to train reward...
The blog post explains how to train reward models using different preference datasets. It discusses aspect-based rankings and the use of MoE gating. The focus is on balancing correctness and helpfulness in models.
GNN-RAG
GNN-RAG combines language understanding from LLMs with reasoning from GNNs for better question answering. It extracts relevant graph information and uses it to improve LLM performance on knowledge graphs. GNN-RAG matches or outperforms GPT-4 with a smaller, tuned LLM.
In my monthly research write-up, I am covering 3 new...
The research write-up covers three new papers on finetuning large language models (LLMs). It explores how prompt masking affects performance, finds that LoRA is good for instruction-following but not for instilling new knowledge, and introduces an alternative to LoRA that improves the uptake of new knowledge. The goal is to provide practical insights from recent studies.
🥇Top ML Papers of the Week
This week's top ML papers cover innovative methods in position encoding, logical reasoning with symbolic expressions, and improving accuracy in numerical tasks. They also explore vision-language models, combining LLMs and GNNs for better question answering, and efficient attention mechanisms. Additionally, they discuss multilingual language models, short-LLMs for long-context tasks, financial analysis with LLMs, and a new approach for preference optimization.
Introducing @middayai Assistant V1
Midday released the first version of their assistant, @middayai Assistant V1. They appreciate the @vercel AI SDK team for their tools. Users are encouraged to try it and give feedback.
MLX LM LoRA Fine Tune.ipynb · GitHub
This GitHub page shares a Jupyter Notebook for fine-tuning MLX language models using LoRA. Users can clone the repository or download it as a ZIP file. Sign up or log in to GitHub to join the conversation and comment.
We are (finally) releasing the 🍷 FineWeb technical report!
The 🍷 FineWeb technical report is now available. It details processing decisions and introduces the 📚 FineWeb-Edu dataset, focused on educational content. Special thanks go to @HKydlicek and the team for their hard work.
Nvidia announces Project G-Assist, an AI assistant for gaming with GeForce GPUs
Nvidia introduces Project G-Assist - an AI helper for gamers using GeForce GPUs. It provides in-game advice and optimizations for better performance.
Sakana AI is proud to sponsor the LLM Merging Competition:...
Sakana AI is sponsoring the LLM Merging Competition at NeurIPS 2024. This competition focuses on efficiently building large language models through merging. For more details and a submission starter kit, visit their provided links.
AI Overviews: About last week
Google introduced AI Overviews in the U.S. to improve search accuracy and user satisfaction. Some users reported odd or erroneous results, leading Google to make technical improvements. Google continues to monitor feedback and refine AI Overviews to enhance search quality and reliability.
NVIDIA CEO Jensen Huang Keynote at COMPUTEX 2024
NVIDIA founder and CEO Jensen Huang will deliver a live keynote address ahead of COMPUTEX 2024 on June 2 at 7 p.m. in Taipei, Taiwan, outlining what’s next for the AI ecosystem. Tune in to watch it live. https://nvda.ws/3UXATRe
LLM Research Insights: Instruction Masking and New LoRA Finetuning Experiments
Sebastian Raschka discusses recent papers on instruction finetuning and LoRA for large language models (LLMs). LoRA is less effective than full finetuning for learning new knowledge but better at retaining existing knowledge. New methods like MoRA aim to balance efficient finetuning with effective knowledge absorption.
JAX is for More Than Just Machine Learning
JAX is a Python library for high-performance numerical computation, not just for machine learning. It features automatic differentiation, hardware acceleration, and a NumPy-like interface. JAX is versatile and can be used in various applications beyond machine learning.
LLM Merging Competition: Building LLMs Efficiently through Merging
The LLM Merging Competition aims to efficiently build large language models by merging fine-tuned models. Participants will use publicly available expert models from Hugging Face that meet specific criteria. Key dates include submission openings in June 2024 and winners being announced in November 2024.
FineWeb: decanting the web for the finest text data at scale
Discover amazing ML apps made by the community
LLM Research Insights: Instruction Masking and New LoRA Finetuning Experiments
This article discusses three new research papers on finetuning large language models (LLMs). One paper finds that not masking instructions during finetuning improves performance. Another paper introduces MoRA, a high-rank updating method, which outperforms LoRA in memory-intensive tasks while maintaining efficiency.
What you need to master Prompt Engineering
Prompt engineering involves designing and optimizing prompts to improve responses from large language models (LLMs). Despite advancements in LLMs, crafting effective prompts remains essential, requiring clear and focused instructions. Experimentation is key, as different prompts may yield varying results depending on the specific LLM used.
AutoTrain: Train ANY Large Language Model with 1 Command
AutoTrain is an easy-to-use library that allows you to fine-tune large language models with just one command, without needing to write any code. It supports both supervised and preference-based fine-tuning on Linux and Mac, and can be used via command-line interface or a user-friendly interface. The trained models can be directly uploaded to Hugging Face for deployment.
Building a dual 4090 rig under $6000
The author built a dual-4090 deep-learning rig for under $6000, detailing the parts and their pricing. They shared insights on each component, including the challenges faced during the build. Concluding, the author suggested that Quadro cards might be a better choice for similar projects.
Here's my conversation with Roman Yampolskiy (@romanyam), AI safety researcher...
Roman Yampolskiy, an AI safety researcher, believes there's a 99.9999% chance AGI will destroy human civilization. The conversation covers various AI risks, including existential, suffering, and control. You can watch or listen to the full discussion on YouTube, Spotify, and other platforms.
OpenAI revives its robotic research team, plans to build dedicated AI
OpenAI restarts robotics research team to develop AI for robots, focusing on multimodal models. They aim to enhance robot capabilities in collaboration with external partners. Apple also partners with OpenAI for ChatGPT integration in iOS.
UPDATE: Sam Altman finally responds to Helen Toner's revelations from...
Sam Altman responded to Helen Toner's revelations during the UN's 'AI for Good' Summit. Helen Toner shared shocking details about Altman's firing from OpenAI on The TED AI Show. For more, listen to her interview on Apple Podcasts and Spotify.
Thanks for reading! If you’d like multiple of these resource updates each week, you support Society’s Backend for just $1/mo for your first year: