"You can usually run only one prompt at a time" There's been some very recent work on MLX that addresses that limitation
MACHINE LEARNING
-
Flexible parsing unlocks hidden capabilities in weaker language models
By
–
Flexible parsing revealed hidden capabilities in weaker models LLM-based parsers captured valid answers from free-form reasoning, improving reported accuracy for models that struggled with rigid constraints. It’s a trade-off: exactness versus resilience.
-
ULMFiT Discriminative Fine Tuning in fastai v1
By
–
In ULMFiT we called this discriminative fine tuning. It's been the default fine tuning method in fastai since v1. 😀
-
Medical AI Benchmarks: Beyond Memorization to Real-World Performance
By
–
Another paper pointing out the inadequacy of older public benchmarks for determining whether AI is actually good for tasks like medicine. Models are clearly memorizing or using heuristics for some answers. A new wave of benchmarks based on real-world data will help, more needed.
-
Open Science and Collaborative Research in Neural Network Optimization
By
–
Looking through those little hidden gem stories in the footnote, you will find it so inspiring that researchers with interests on the same topic are able to work together to advance a field despite their roles and locations. This is the power of open science and community. Thinking Machines (@thinkymachines) Efficient training of neural networks is difficult. Our second Connectionism post introduces Modular Manifolds, a theoretical step toward more stable and performant training by co-designing neural net optimizers with manifold constraints on weight matrices. thinkingmachines.ai/blog/mod… We explore a fundamental understanding of the geometry of neural network optimization. — https://nitter.net/thinkymachines/status/1971623409873244462#m
→ View original post on X — @lilianweng, 2025-09-26 19:03 UTC
-
AI Progress, Evals, and Reinforcement Learning in Job Market
By
–
quite a privilege to be a guest on Odd Lots! talked w joe & tracy about pace of progress in AI, broken evals, the crazy job market, and the difference between supervised and reinforcement learning
-
Machine Consciousness: Testing Organism-Based Hypothesis Through Computing
By
–
The machine consciousness hypothesis has two parts: a conjecture about consciousness in organisms, and its validation by translation into a testable computer model.
-
OpenAI Google DeepMind ICPC models code execution tools clarification
By
–
Anyone know if there's a solid answer from OpenAI or Google DeepMind yet on whether the models they used in their ICPC programming competition wins had access to tools like the ability to run generated code through a compiler?
-
Constrained Decoding vs Out-of-Distribution: Grammar-Based Selection
By
–
the hope is that “constrained distribution” is not quite the same as true OOD right like if i constrain decoding to valid json grammar, json is still within distribution, im just choosing a “top k” subject to my constraints
-
Compression data reveals patterns uncompressed information structure
By
–
It has to do with the way compressed (something) reveals uncompressed (something else)