LLMS

Multiverse Computing open-sources two changes that make knowledge distillation cheap enough to run at scale: caching the teacher's top-100 logits once so it never sits in memory during training, and a fused chunked KL loss that never materializes the vocab-by-sequence matrix. A GPT-OSS 20B distillation at 32K context drops from four GPU nodes to one, with step time down 5x.

AI2 released TutorMoments, a replay-based benchmark that drops seven LLMs into real math tutoring transcripts and scores whether they scaffold when help is needed or push for rigor when the student can do more. The default finding: models over-help, and spelling out the trade-off in the prompt lifts every score but does not close the gap to a consistent human call.

A placebo-controlled study on MBPP+ finds that when small code models fail, resampling from scratch beats repair loops that feed the failed code back - at 2.5-5.5x fewer tokens. The failed attempt is the anchor.

A new 600-session benchmark shows coding assistants that read a user's resolved session history resolve ambiguous requests with far fewer clarifying questions - Claude Opus 4.8's first-turn success jumps from 24.3% to 60.3% when history is available.

Inkling-Small is a 276B-parameter MoE with 12B active per token, Apache 2.0, and open weights. It beats the 975B Inkling on SWEBench Verified (80.2), HLE (31.6), and tool use at a quarter of the size and a third of the output price.

Effort levels and model choice both cost more for more capability, but they are not interchangeable. Here is when to move the effort dial and when to switch models instead.

DeepSeek suspended its $74B valuation fundraising round after a leaked transcript of founder Liang Wenfeng's investor meeting laid bare the compute gap between Chinese and US AI labs - revealing he needed 200,000 Huawei 950 chips but received only 16,000.

Mozilla's inaugural report reveals open models now match closed AI on capability, but only 51% reach production. The harness layer and permission model gaps explain why.

A developer built an 85% accurate LLM text detector using TF-IDF and linear SVM - no neural networks required. Here is how it works and what HN thinks about AI detection.

A new American open-weights frontier model with multimodal capabilities, 1M token context, and competitive benchmarks. Here's what the HN community thinks.

A Hacker News discussion blows up over LLM vocabulary quirks, with developers sharing hooks, filters, and coping mechanisms for repetitive Claude-isms.

George Hotz publishes a post distinguishing genuine AI progress from manipulative hype narratives. HN's 126-comment thread debates whether he's right about doom-mongering and AGI inevitability.

A solo developer built a 1,300-line C inference engine that runs the 744B GLM 5.2 model on consumer hardware by streaming routed experts from disk. Here's how it works.

A new study from Dartmouth measures the impact of an AI tutoring platform on introductory statistics performance. Full engagement with the system correlated with significant exam score improvements, though selection bias remains a key limitation.

Anthropic's new research reveals LLMs have an internal 'workspace' for silent reasoning - and it could change how we build safer AI.

Comparing LLMs by token pricing alone can lead you to choose worse, more expensive models. Cost per task tells the real story.

Filippo Valsorda argues that LLMs have ended the era of treating security researchers with kid gloves. When anyone can discover vulnerabilities with an AI, the old coordinated disclosure model breaks down.

Unsloth's dynamic quantization makes GLM-5.2 runnable on a 256GB Mac or a 24GB GPU with CPU offloading. Here is the hardware math, the quantization tradeoffs, and what the HN community learned from actually running it.

A new paper shows a 3B parameter model hitting 94.3 on AIME26 and 96.1% on LeetCode contests - matching or exceeding models 100x its size. The catch: it traded general knowledge for pure reasoning ability.

Switzerland's fully open foundation model promises transparent training data and EU compliance. The HN crowd has questions about actual performance.