Topic
Everything filed under Machine Learning, newest first.
RSS · JSON · All topics
On Chomsky and the Two Cultures of Statistical Learning
Peter Norvig responds to Noam Chomsky's 2011 criticism of purely statistical approaches to language, explaining what statistical models are, documenting their success in language tasks, and arguing that probabilistic modelling is a legitimate scientific way to understand language rather than mere mimicry.
37 min · 8,521 words
What objective is reasoning solving?
Alex Wang frames LLM reasoning as pathfinding in token space, comparing best-of-k sampling with training models to generate better samples, and uses that lens to discuss amortization, unfaithful chains of thought, and reward hacking from both verifier and sampler.
10 min · 2,275 words
The Surprising Effectiveness of Deleting Weights in LLM Post-Training
NeurIPS 2026 paper introducing Deletion Tuning: zeroing fewer than 0.05% of an LLM's weights via a learned binary mask on gate projections matches full fine-tuning across RLVR, SFT and on-policy distillation, with far smaller checkpoints and much less forgetting.
1 min · 208 words
John Langford explains why he is leaving Microsoft Research as its AI Frontiers group, formed after ChatGPT to organise research at scale, is wound down: what the group achieved, how reorganisations and leadership decisions undercut it, and what Microsoft would need to become attractive for research again.
4 min · 922 words
Hugging Face walkthrough of Phonon TTS: high-quality on-device speech models that top benchmarks, with details on architecture, latency, and open release.
6 min · 1,387 words
Faster Retrieval for Better, More Efficient RAG
Apple ML Research frames maximum inner product search as a support function and introduces learned routing (SupportNet) that plugs into existing vector indices to speed up retrieval for RAG and recommendation systems.
11 min · 2,612 words
The AI that builds your AI. Now for agents.
Oumi’s November update extends FAiE (Frontier AI Engineer) from models to full agents: build/evaluate/train agents, Bring Your Own Harness, DPO plus GRPO private preview, multi-stage training, and clearer cost visibility—so teams can automate agent improvement while owning the weights.
3 min · 678 words
Gemini Omni 1.2 Flash: Fast and cost-effective with longer clips
Google DeepMind updates Gemini Omni with 1.2 Flash: native 15-second video generation and editing, stronger consistency, faster generation, and about 30% lower pricing.
6 min · 1,339 words
The Ultimate MoE Training HandbookHow to train a Mixture-of-Experts model at scale, setting by setting, with the runs that back each choice
Hugging Face publishes a free, hands-on MoE training handbook distilled from hundreds of thousands of GPU hours: expert sizing, routers, load balancing, all-to-all communication, upcycling, and serving.
185 min · 42,497 words
An open-weight world action model to accelerate industrial robot deployments.
14 min · 3,295 words
Does a Weight Edit Teach Qwen3.5-0.8B Math, or Stop It Looping?
An agent edited 131,072 weights of Qwen3.5-0.8B and MATH-500 rose from 7.9% to 48.1% with thinking on. The edit is rank one in two layers, and its neurons fire 4.5 times as hard on repeated text. With thinking off, the card's default, the model scores 43.6% unedited and 45.2% edited.
9 min · 2,071 words
Shaping and Monitoring Models Through Their Internal Representations
We currently reward AI models for succeeding in training, but don’t specify how they should get there – leading to misgeneralization and misalignment. Our latest paper introduces mechanistic shaping, a way to teach models to be “right for the right reasons”, and studies its effects on monitorability.
16 min · 3,622 words
PLaMo 3 Translate: releasing world-class translation LLM weights
Preferred Networks releases PLaMo 3 Translate model weights (0.8B–31B) on Hugging Face and explains how the translation-specialized LLMs were trained and evaluated for Japanese and multilingual translation.
9 min · 2,000 words
Thinking Machines Lab introduces Inkling 2 Small, a Mixture-of-Experts model with 594B total / 17B active parameters, image inputs, variable effort, and large gains on agentic coding and related benchmarks.
21 min · 4,889 words
Pangram announces Pangram Image 1.0 and Image Flash 1.0 for detecting AI-generated images and video, with heatmap localization, robustness improvements, and notes from the research-preview period.
4 min · 832 words
LLM Judges for Identity Resolution at 4x Lower Cost
We wanted to test whether LLMs and decision models, specifically Jev and Luna, could solve a hard problem that we encounter regularly (grading identity resolution results). We found that while both models performed okay out of the box (88.8–97.3% accuracy), implementing a hybrid model that combines Jev + Opus achieves frontier accuracy at > 4x lower costs. - Luna performs better than Jev out of the box, but the gap disappears on harder cases. - Jev's confidence is better calibrated. Most of its mistakes come with low confidence its right. Most of Luna's mistakes come with…
11 min · 2,468 words
From Static to Polymorphic Inference: A new era for production open-source inference systems
Models, traffic and chips are constantly changing. Polymorphic Inference adapts to each, and Together Compiler uses AI agents to keep production inference tuned.
7 min · 1,647 words
Training a Text-to-Image Model for $21K
Linum shares field notes from a roughly $21K proof-of-concept run training a text-to-image model with its Pyramid-JiT approach, covering the training regime, results, model limitations and Apache 2.0 weights.
7 min · 1,632 words
Engram shows that a short reinforcement-learning phase on top of fixed text notes improves agent accuracy by about 7–20 points across workspaces, often cuts inference tokens, and remains useful even after millions of tokens of note edits because models learn note-use strategies rather than memorizing contents.
16 min · 3,790 words
Qwen 3.8 27B on my laptop can interpret model weights, too
Simon Willison tries an example from the 'Sparks of Weight-Space Programming' paper and finds Qwen 3.8 27B running locally on his laptop can also interpret raw neural network weights written out as text, while arguing the wider commentary that source code is over goes too far.
5 min · 1,241 words