Welcome to the weekly Machine Learning for Software Engineers newsletter! Each week I curate an AI reading list specifically for software engineers. What does that mean? It means this weekly newsletter contains everything software engineers should know about AI:
A short, general lesson about AI from the past week.
Interesting and important things that happened this past week.
~5 resources to learn and become a better engineer.
A mini-update on the current job market w/some promising openings.
I try to make these dense so they’re as beneficial for you as possible. Subscribe to get these in your inbox each week.
If you read these each week, welcome back! You’ll notice I’ve included a jobs section this week. I plan on doing this each week to give y’all a general overview of available positions. I’ve got a few articles upcoming that I’m really excited about. Stay tuned for:
Why the era of the ‘coder’ is over and why that’s most important for anyone who doesn’t write code
A case study of Spotify [paid]
A complete overview of the AI job market for July 2025 [paid]
How to build a simple recommendation system from scratch
What makes ML experimentation so f*cking hard
Enjoy the weekly update!
I ran into an interesting post this past week I want to share with y’all:
Like all transformative technology (think: the internet, video games, video streaming, etc.), we’ll need to determine our relationship with AI. We’ve already seen a reliance on LLMs decline cognitive ability and I honestly fear for the downsides of AI reliance especially in young kids that are going through their schooling with it.
The most important thing I was taught in school was how to think. In my opinion, that is the purpose of an education. You absolutely should learn basic information needed for life, but you should also be taught to think for yourself and how to deduce more information based on the information given to you. If you haven’t learned this, your education has failed you.
This learned skill is like a muscle: The more you use it, the better it gets. If you stop using it, it atrophies. Just like working out to grow muscle, the process to develop this skill requires a struggle.
Technological advancement has necessitated physical workouts to keep our bodies in healthy shape as jobs have shifted to be more stagnant and less physical. We created gyms and other sources of exercise to overcome this stagnation, yet so many people fail to utilize them and we’ve seen the impact that has had on the health of our populace. I argue that an atrophy of the same level, but in our minds, would be much more detrimental to the populace.
So the question is: What’s the mental equivalent of a gym and how do we ensure it’s effective? AI isn’t going anywhere and there’s no doubt it’ll be economically transformative, but what should our relationship with it look like?
Let me know what you think in the comments.
If you find ML for SWEs helpful, consider becoming a paid subscriber (and get all the articles above!). It’s still just $3/mo, but will go up to $5/mo after 10 more subscriptions.
Important things that happened last week
Zuck is recruiting ruthlessly for his Superintelligence team at Meta (to be called the Meta Superintelligence Lab or MSL). Mark Zuckerberg is personally recruiting high-profile AI talent, much of which has come from OpenAI. He’s reportedly offering 7 figure+ salaries and generous signing bonuses (not $100M like many claim, but still high).
This initially prompted Sam Altman to say Meta isn’t creating a great culture by poaching talent and that none of their really impressive talent had left. Less than a week later, OpenAI released a statement saying they will be reviewing their employees’ compensation packages to ensure they keep talent.
Anthropic had Claude run a vending machine and it was wildly unprofitable. Anthropic let “Claudius” control the shop w/o human oversight. I suggest reading Anthropic’s report above for all of Claudius’s intricacies. The outcome: AI is great for certain tasks (research, web browsing, text generation, etc.) but not economically competent on its own (in Claude’s current state, at least).
Airtable has released an AI app-creating assistant. The Airtable CEO believes app creation is the new interface for which AI will interact with components around the web. Airtable’s new assistant allows users to create apps and interfaces they want and use Omni (the assistant) to interact with those interfaces and the data they contain. AI will fundamentally change the way information on the internet will be interacted with and this is a good look at how.
A court ruled in favor of Anthropic and Meta purchasing copyrighted material and using it to train their AI (specifically books, but it will likely extend to other mediums). The court ruled this as ‘fair use’. This will not only settle other lawsuits still in progress, but its huge for generative AI companies looking to acquire training data for their models AND has implications for authors and creators about how they’ll publish their creations.
Google DeepMind introduces Gemma 3n. This new open model brings powerful, multimodal AI to edge devices, meaning it can run within <3GB of memory. It’s the highest scoring model of its class in LMArena coming in right under Gemini 1.5 Pro. Google credits this advancement to Gemma’s mobile-first architecture which you can read more about at the link above.
PadChest-GR: A new benchmark for testing biomedical imaging models. We’ve already seen AI outperform radiologists at diagnosing disease by interpreting medical images, but AI is still seen as a copilot because the cost of a false negative in this domain too high to risk. This benchmark aims to accurately evaluate the performance of these models in the real world.
DeepMind releases AlphaGenome to predict how DNA variants impact gene regulation. AlphaGenome is available via API and aims to accelerate scientific understanding of disease and drive the discovery of new treatments. I’m not going to pretend I fully understand AlphaGenome’s applications, but this is an incredible application for AI (much better than most LLM applications we’ve seen so far).
Learn a language by chatting with a voice AI assistant. The best way to learn a language is by speaking it. Previously, this meant immersion, but now we have AI assistants that can speak the language and direct learning as the user needs. You can try it out yourself at the link above. It still needs tweaking, but it shows the potential of this application.
Anthropic allows for building and sharing interactive apps directly in Claude. Users can create applications in English and share them with others. Claude Artifacts acts as a personal app gallery with each app operating in its own sandboxed environment.
Apple is thinking of using Claude or ChatGPT to power Siri. This is a huge diversion from Apple’s usual ‘built in house’ mantra. They have a tight control over their hardware and software and using an external AI assistant relinquishes some of this control. This could be the end of an era for Apple and it shows how their AI blunder is having a very real impact.
Tesla delivered their first vehicle completely operator-free. This means there was no one in the car as it drove from Tesla’s factory to the recipient in Austin. I love this because it’s a concrete example of the benefit of self-driving technology. Imagine every car being delivered directly to the recipient. The time-savings of not have to be at a dealership all day will alone double the US GDP.
Humor aside, I don’t think most people realize the benefit of self-driving cars. There’s a lot of fear in this area because mistakes are costly, but the data shows just how much safer self-driving cars are than humans. Self-driving cars means lives saved and less time wasted commuting. You can see the video of Tesla’s delivery below.
If you want a more complete overview of AI this week, read ’s weekly roundup. His is one of few overviews I read each week:
Resources so you can become a better engineer
Nvidia's Blackwell GPUs achieve immense scale by prioritizing Streaming Multiprocessor density. The GB202 die is massive, featuring 92.2 billion transistors and 192 Streaming Multiprocessors. Its 1:16 SM to GPC ratio allows high compute density, though it can challenge feeding short-duration tasks. Blackwell also improves workload overlap compared to previous generations.
Blackwell's GB202 is a giant GPU die with 92.2 billion transistors and 192 SMs.
Its 1:16 SM-to-GPC ratio enables high compute density, but can limit short-duration workload throughput.
Blackwell improves performance by allowing different workload types to overlap on the same queue.
Life of an inference request (vLLM V1): How LLMs are served efficiently at scale
Optimize large language model inference at scale with vLLM V1. Ubicloud leverages vLLM to deploy and manage large language models across GPUs. It efficiently load balances incoming requests, ensuring scalable and robust service. The article details the path of an inference request through vLLM's OpenAI-compatible API server and core engine.
vLLM V1 architecture achieves state-of-the-art text generation performance.
Efficient LLM serving requires deploying multiple instances and smart load balancing.
Robust management includes health checks and seamless upgrades for high availability.
Understanding the internal request flow is crucial for scaling and optimizing LLM services.
Why "AI Hate" is Your Next Billion-Dollar Opportunity [Markets] by
Turn AI dissent into your next billion-dollar investment. The growing negative sentiment towards Artificial Intelligence, driven by real concerns and public distrust, creates a unique contrarian market opportunity. This "AI hate" acts as an essential antithesis to pervasive AI hype, leading to a synthesis that yields more robust and widely adoptable technologies.
AI "hate" offers asymmetric financial returns by betting against market hype.
Critical perspectives force the creation of more robust and user-aligned AI systems.
Embracing the "antithesis" of skepticism is crucial for meaningful technological progress.
The Google for Startups Gemini kit is here
The Google for Startups Gemini Kit empowers entrepreneurs to build faster using AI. This new, free suite provides essential tools, credits, and support for integrating AI into products. It streamlines the adoption process for startups at any stage.
Gain instant access to the Gemini API and Firebase Studio for development.
Unlock up to $350,000 in Google Cloud credits for AI usage.
Access comprehensive training, documentation, and expert workshops.
Join a dedicated community for shared learning and inspiration.
Watch the full video here:
Collaborative filtering deep dive
A guide/Jupyter notebook to get in-depth with collaborative filtering, an algorithm commonly used by recommendation engines (think video or song recommendations that would be used by companies like Netflix or Spotify). I’ll be writing more about this (and introducing it via code) in an upcoming article, so stay tuned!
Learn how to recommend products that would be useful for a specific user given a list of users, products, and user interactions with those products.
Use the MovieLens data to recommend movies a user may like.
Use the fastai deep learning library.
This week’s jobs
Here are some interesting opening this week. I also post these on my socials, so make sure to follow me on X and LinkedIn. I’ll develop this section as time goes on to include more and more relevant jobs.
If you want to learn more about the current job market, support ML for SWEs as a paid subscriber and you’ll get less frequent, but much more comprehensive AI job market overviews.
1. ML/AI Engineer CAD Infrastructure - Tenstorrent (Remote)
Overview: Develop ML-powered systems to automate post-silicon validation, debug, and root cause analysis for semiconductors. This role builds and maintains the infrastructure for large-scale silicon testing and creates predictive models to analyze behavior, optimize performance, and detect anomalies.
Key Skills/Requirements:
Master's degree or Ph.D. in Computer Science, Electrical Engineering, or a related field
5+ years of experience in post-silicon validation, silicon bring-up, or CAD infrastructure development
Strong background in machine learning and deep learning frameworks (PyTorch, TensorFlow)
Proficiency in Python and data analysis libraries (NumPy, Pandas, Scikit-learn)
Strong understanding of computer architecture and instruction set architectures (ISAs)
2. Machine Learning Engineer Graduate - TikTok (San Jose, CA)
Overview: Participate in the development of large-scale Ads systems and next-generation monetization platforms. Responsibilities include developing state-of-the-art applied machine learning projects, owning key targeting components, and working with product teams on product vision.
Key Skills/Requirements:
Go, C/C++, or Python programming
Deep learning and machine learning experience
Ads bidding & auction, ads quality control, and online advertising systems
3. Quantitative Researcher Machine Learning - Two Sigma (New York, United States)
Overview: Use a rigorous scientific method to develop sophisticated trading models and shape insights into how markets will behave. Apply machine learning to vast datasets to create and test complex investment ideas for trading a variety of global markets.
Key Skills/Requirements:
Degree in a technical or quantitative discipline (statistics, mathematics, physics, etc.)
Intermediate skills in at least one programming language (C, C++, Java, or Python)
Understanding of machine learning algorithms
Experience with applied machine learning to real-world datasets
4. Senior Engineer AI Research - Qualcomm (San Diego, CA)
Overview: Conduct fundamental and applied research in machine learning, creating new models and training methods in areas like generative AI, LLMs, and reinforcement learning. Responsibilities also include developing and optimizing software, tools, and compilers to deploy efficient machine learning solutions on resource-constrained devices across mobile, automotive, and IOT sectors.
Key Skills/Requirements:
PhD in AI, Computer Science, Engineering, or a related field (or MS with 4+ years experience)
Strong programming skills in Python and PyTorch
Experience with LLM, LVM, LMM models
First author publications at major AI conferences (e.g., NeurIPS, ICML, ICLR)
5. Early Career Machine Learning Engineer - NFX (San Francisco Bay Area)
Overview: Develop and deploy machine learning, NLP, and generative AI models that power the Piai™ claims-intelligence platform. You will turn raw legal and medical data into production-ready models that directly improve justice for personal-injury clients by helping law firms secure better outcomes.
Key Skills/Requirements:
Ph.D. or M.S. in Computer Science, Machine Learning, or a related field
Hands-on experience with NLP or generative-AI techniques (e.g., transformers, LLMs)
Proficiency in Python and ML/NLP libraries (e.g., PyTorch, TensorFlow, Hugging Face)
Familiarity with SQL
Thanks for reading!
Always be (machine) learning,
Logan
