This small 30B model just outperformed systems over 20x larger including DeepSeek-V3.1-671B and OpenAI’s o4-mini. The secret? It doesn’t just store memory it folds it. A self-managing memory system that keeps its thoughts organized while solving 500-step problems.
MACHINE LEARNING
-
Alibaba’s AgentFold introduces a human-style memory system for web agents
By
–
Alibaba just dropped a monster paper in agent research. It’s called AgentFold, and it basically gives web agents a human-style memory system that manages itself. Here’s why it’s wild Current agents either: ∙ keep everything (context bloat, chaos)
∙ or summarize too -
PyTorch Version of Hands-On Machine Learning Book Released
By
–
PyTorch version of Hands-On Machine Learning Book by Aurélien Géron is just out! This was one of my favorite books when I started learning machine learning and it's nice to see it using PyTorch. Data science courses/MOOCs back then used to teach you numpy, pandas, matplotlib,
-
Neural Networks Tolerate Mantissa Bit Reduction for Bandwidth Savings
By
–
Rough thinking: neural nets are very tolerant of noise, so chopping off mantissa bits is akin to small amounts of noise. So I'd rather lose nearly all mantissa bits before losing exponent bits. So tried dropping 16 mantissa bits & model trained just fine w/ half the bandwidth.
-
The Dependence of LLMs on Wikipedia Data
By
–
LLMs would be really dumb without the existence of Wikipedia.
-
Frontier Models vs GPT Codex: Speed to Success Comparison
By
–
Even frontier models fail all the time. The difference is we fail in 15 sec and with one more prompt you get the right answer. Time to success = 30 sec. On GPT Codex it takes 22min just to find out it failed.
-
Language Models Struggle With High School Math Fundamentals
By
–
Language models perform poorly on high-school math? You don't want to hear this, but the problems started in grade-school. The moment we (collectively) found acceptable that mid-tier models could score only 75%-85% on a GSM test set of 1.32k straightforward problems…
-
OpenAI Launches Aardvark Security Research Agent
By
–
OpenAI is launching a new Security Research Agent, "Aardvark", that can identify and resolve vulnerabilities in code. Aardvark is currently available in private beta and powered by GPT-5 and Codex.
-
Stanford Method Detects Copied Fine-Tuned AI Models
By
–
Someone stole your model & u can’t prove it? This Stanford paper just showed that you can find out if a model is a copy or finetuned based on your model with just its generated text So if someone yoinks DeepSeek-v3.2 and finetunes it, it’ll leave statistical traces!
-
TU Berlin’s Legacy of Success in Machine Learning
By
–
We do TU Berlin is one of the most successful schools in the history of ML, Klaus, Alex, Marc, Bernhard … you should be very proud I hope.