220 stories
Yesterday I shared our open-source Clash Royale simulator and its recurrent PPO agent here. A training loop is easier to understand when you can watch it, so we put a small interactive version online:
Sharing our recent work, now accepted at NeurIPS: Functional Gradient Descent with Adaptive Representations. Functional GD algorithms generally outperform neural nets, but are hard to accurately implement.
I benchmarked Qwen3-VL 8B Instruct (Q4_K_M, Ollama, M5 24GB, ~30s/doc) against Claude Opus 5.5, Sonnet 5 and GPT-5.6 Terra on: - receipts: CORD (Indonesia) and SROIE (Malaysia), 30 each - 20 scanned 1980s-90s invoices , answer keys human-verified - 32 real IRS forms, 4 damage levels (generated this…
Hi, I work as a Data Engineer at a manufacturing company where we build engines. A significant part of my work involves ML/DL-related tasks, and I’d like to turn one of my projects into a research publication.
Hi, same as the title, I am currently trying to begin with some research on clustering using LLMs. So my requirement is as follows: I will be given some 100 document files, the end goal is to have clusters in such a way that documents with similar procedures or content should be clubbed under…
AI Engineering from Scratch is an MIT-licensed curriculum: 523 lessons across 20 phases, from linear algebra and backprop to transformers, LLMs, agents, and production serving. The code is stdlib-first, so you see every step instead of calling a library.
Saw this job posting recently from Deepmind where the minimum requirements are a Bachelor's and 2 years of experience: https://www.google.com/about/careers/applications/jobs/results/120898404719960774-optimization-research-engineer-deepmind
OpenTrainDNN is an open-source, client-side web application designed to render the step-by-step training mechanics of deep neural networks in real-time.
Help: Project l'm building a shelf audit tool. A photo goes through YOLO, which crops each product, and then I embed the crop and search a small gallery of reference photos to get the SKU. New products should be addable by just dropping in photos, no detector retrain. Detection is basically fine.
I was reading a paper that surveyed the field of neural architecture search, where it said within 5 years, around 3000+ new models were proposed. The amount of compute and resources spent on this is absolutely astronomical.
The opponent plans by simulation: every second it scores each candidate play by running the match 10 seconds ahead in the engine. Our PPO agent learned to park its Cannon behind its own King. Losing a building in a fight cost reward, and letting it decay cost nothing, so it found the loophole.
Recently I have been working on image attachment integration for my VEX agent runtime, and I came up with an idea for a significant optimization step. Typically images attached to LLM context as raw bytes (base64 encoded string or publicly accessible URL), and the VLM processes image data by…
We published Intern-Decision multi-modal model family (0.8B, 2B, 4B), and the 4B model outperformed Jev 1.13.0. https://huggingface.co/collections/internlm/intern-decision https://github.com/InternLM/Intern-Decision Performance Testing on RTX 4090, we find local deployment is fast enough to enable…
In this project I wanted to see if any interesting emergent behaviors would appear if we trained two agents to play a streetfighter-like game using RL. Maybe obvious in retrospect, but the agents are really good at reward hacking.
A radiologist can mark a tumor on a CT scan by drawing a straight line across it. Creating a complete 3D segmentation requires outlining the tumor across the slices where it appears, which is more tim
Please post your personal projects, startups, product placements, collaboration needs, blogs etc.
Hiring: [Location], Salary:[], [Remote | Relocation], [Full Time | Contract | Part Time] and [Brief overview, what you're looking for]
Every AI agent tool call is just text, and the program reading it is unforgiving — one bracket out of place and the whole response fails to parse. Prompting your way around that gets you "most of the time," which isn't the same thing as "every time." This deep dive explains constrained decoding:…
Showing 216 of 220 stories
Load more storiesGet the top AI news and tools straight to your inbox every week.
Weekly · AI news
Join 15,000+ AI enthusiasts
