Andriy Burkov's Linkedin Analytics

This week's edition of my AI newsletter has just been dispatched to subscribers. Featured in this issue: AlphaEvolve: a Gemini-powered coding agent for designing advanced algorithms; who's to blame when AI agents screw up?; a foundation model for the Earth system; teaching AI to communicate sounds like humans do; the use of artificial intelligence in courts: a warning; insights into DeepSeek-V3: Scaling challenges and reflections on hardware for AI architectures; and more. Subscribe to ensure receipt of future editions directly in your email inbox. #artificialintelligence #ai #machinelearning #ml #deeplearning #analytics

Someone asked if this post was satire. It's not satire. The need for junior developers and developers in general is currently reduced because of the productivity boost provided by LLMs, and it's hard to predict whether it will create more jobs or fewer jobs in the long term. I'm able to build very well-functioning MVPs alone in a matter of a week, often without knowing what I do. In the past, to achieve this, I would need to spend weeks if not months to learn all technologies involved in building this MVP and then I would need months for implementing it. Senior developers benefit the most: they can easily see when the LLM suggests something wrong or poorly designed and correct it. If an LLM is stuck, they can unblock the project by hand and continue with the LLM. Again, what would need a team, now takes one senior developer. I anticipate comments like, "But you build crap with shitty code." Yes. But in the past, I did nothing. Now, I create MVPs which, if they become interesting to users, I can use to find investors, hire a senior developer, and refactor as long as the codebase is still small.

This week's edition of my AI newsletter has just been dispatched to subscribers. Featured in this issue: it’s too expensive to fight every AI copyright battle; the hidden bloat in machine learning systems; do LLMs estimate uncertainty well?; a deep dive into memorization in deep learning; RL backlog; an open source alternative to Weights and Biases; mining the potential of multi-modality vision language models; and more. Subscribe to ensure receipt of future editions directly in your email inbox. #artificialintelligence #ai #machinelearning #ml #deeplearning #analytics

Hypothesis: Observational sciences like astronomy, paleontology, geology, history, archaeology, evolutionary biology, psychology, sociology, anthropology, and economics, where results are usually impossible to verify or reproduce, are being bloated with LLM-generated publications. This will result in an avalanche of bogus Master's and PhD degrees in the next 3-5 years.

I'm happy to share that I'm joining MindsDB as an Advisor. MindsDB is focused on fascinating and, in my view, crucial aspect of applied AI: seamlessly integrating machine learning capabilities directly into the data layer. They empower developers to leverage existing databases as predictive engines, using the familiar language of SQL. This dramatically lowers the barrier to building AI-powered applications, abstracting away much of the traditional MLOps complexity. As you most likely know if you read this post, making advanced technology accessible and understandable has always been a core theme in my work. The approach Jorge and the team have taken – bringing AI to where the data lives and to the tools developers already master – resonates deeply with me. I see significant potential in the product and the team's vision to democratize AI adoption within the broader software engineering community. I will mainly contribute by: - Helping in defining product features, making sure MindsDB continues to solve real-world developer problems with elegance and power. - Hosting webinars designed to clearly demonstrate the simplicity and versatility of MindsDB for software engineers. We'll explore practical applications and, of course, I'll be diving into how it all works "behind the scenes" to provide that deeper understanding. I believe MindsDB is on a compelling trajectory, and I look forward to being part of their journey and engaging with the community. The first webinar is coming soon. Stay tuned!

Vibe fixing a bug:

In my new book on reinforcement learning (The Hundred-Page Reinforcement Learning Book), the reader will train a SpaceX booster to land on a moving naval platform under windy conditions:

I’m happy to share that I’m starting a new position as Advisor at MindsDB!

If you're struggling to find a job as a software engineer, you aren't alone. The productivity boost that senior software engineers gain from using LM-assisted coding will push this graph further down. The hardest part will be getting a junior position.