Yo, space is old hat. We’re naming AI models after elements now blog.google/innovation-a...
Al Merose
Researcher with public evidence across AI research.
- AI signals
- 4 past 30d
- Sources
- 1 distinct domains
- Discussões
- 9 past 30d
- Latest signal
- 10d ago
Articles & links
Sneak peek: I implemented a 3 layer MLP that predicts the fashion MNIST dataset using SQL (with my own autograd impl inside the database). Apparently einsum notation is better expressed as a JOIN. github.com/xqlsystems/x...
- The benchmark trains a 784→196→32→10 MLP on Fashion-MNIST with forward pass, backpropagation, and SGD all expressed as SQL queries.
- A SKIP_ZERO_PIXELS flag drops background pixels via a WHERE clause and delivers roughly a 1.8x speedup on Fashion-MNIST.
- Hyperparameters are demo-scale: 500 training samples, 200 test samples, learning rate 0.5, 60 steps, chunk size 250.
Maybe a troll, but also maybe a union one day: concernedaistaff.org
New result: Automatic derivation of FlashAttention from vanilla attention in ~25 lines of SQL via three portable extensions. github.com/xqlsystems/x...
Not yet merged, but I added autograd to my array database experiment. I’m building a new NN framework primarily in SQL. github.com/xqlsystems/x...
Now that models have gotten smarter, I’m experimenting with putting in ideas about ocean physics in my AGENTS .md, specifically what invariants in the scientific domain to look for to find modeling bugs. github.com/m2lines/Samu...
With my friend @realcromwellian.bsky.social’s help (and his Mac Studio), we got a 126M param NanoGPT model training in ~165s — primarily written in SQL. github.com/xqlsystems/x...
In other words, you’re free to write the mathematical definition of attention in SQL, and you get the high performance just by adding on my smart query planner: github.com/xqlsystems/e...
I would use this tree map from USGS as a ground truth to train a small regression model on top of Alpha Earth Embeddings to get a close bound. data.fs.usda.gov/geodata/rast... deepmind.google/blog/alphaea... Then, I would add data to see if I could make the bound tighter.
Oy vey. Better to avoid that direction. I’m constraining myself to a $20/mo plan for my side projects. Instead, I’m trying this with my sub agents: github.com/xqlsystems/e...
All of our code is open source, and contributions are welcome: github.com/m2lines/Samu... We're working on fitting our model training into a (free-tier) colab notebook, so anyone can help advance climate science: github.com/m2lines/Samu...
All of our code is open source, and contributions are welcome: github.com/m2lines/Samu... We're working on fitting our model training into a (free-tier) colab notebook, so anyone can help advance climate science: github.com/m2lines/Samu...
Recent commentary
For file type, there is at least one type of social media built around it: images > instagram, video > youtube, vertical video > tiktok, audio > clubhouse, safetensors > huggingface, executable files > github
I’m like 6 months to a year behind using all sorts of agentic coding patterns (I just asked Claude to delegate impl to subagents, etc.) and honestly (surprisingly?), I think I’m better off for it.
Agentic engineering feels like “problem engineering” to me. Coding becomes defining goals and constraints for agents to optimize towards. Like all coding, this requires taste and discernment.
Rust is ideal for agentic programming because large compile times give you an opportunity to think and reflect
Open source project idea: give that old phone in your drawer new life by installing Prism ML’s LLM on it to make it a stateful agent.
2004: real graphics programming is hard, it can’t just be simple C programs. 2011: real machine learning is done in C or C++, Python is just a toy or pedagogical language. Now: SQL for machine learning is cursed and will never work.
I don’t understand why there is so much agentic tooling. I feel very productive with just Claude equipped with the gh CLI, using standard GH features like issues and comments. While I use prompt patterns, I feel most cases require sufficient nuance that I don’t want to capture something in a skill.
I’m happy to be working on open source science and language models because the biggest safety risks of AI, imo, stem from concentrations of power — in corporations, governments, and their intersection.
If there is one contribution that I hope to make to computer science, it’s that relational algebra is as important as linear algebra or calculus is to machine learning.
Kind of a crazy result: for single step prediction, compression is better than persistence as an ocean physics forecast, and in some cases better than my ~100M parameter neural network, Samudra. (Ofc, this 4000 param model blows up on short autoregressive rollouts.)
In Al Merose's orbit
Center = Al Merose. Left = members they follow (green edges). Right = members who follow them (blue edges). Top = mutual follows (orange edges, slightly larger). Drag any node to reposition; click to open that profile.
Are you Al Merose? Show it.
Add the Who’s Who of AI badge to your site or bio. It links back to this profile.
Markdown: [](https://aiweekly.co/whos-who/person/al-merose-com)