AI Briefing#llama
The latest AI and developer news about #llama, with the original source and a short summary.
The latest AI and developer news about #llama, with the original source and a short summary.
Deposition testimony reveals a former Meta researcher downloaded the pirated Books3 dataset, contradicting Meta's earlier claim that it did not obtain the data via BitTorrent.
Meta reported 97 mentions of Muse and zero of Llama in its Q3 2026 earnings call, alongside a 2026 capex guide of $135B to $145B.
The experimental proof of concept burns the model into silicon, achieving speeds 213x faster than an RTX 5070.
Taalas and Groq tackle the memory bottleneck in AI inference by fixing model weights onto chips.
Analyzed the independent development of LLMs using public configurations, tokenizers, and weights.
The Petals project uses a BitTorrent-style approach to let consumer GPUs run huge LLMs like Llama 3.1 405B in a distributed manner.
LLMs can significantly improve their code generation ability using only their own outputs, without needing a verifier or separate training.
Amazon extended the Chinchilla scaling law to present a framework that speeds up LLMs without losing accuracy.
Meta has been hit with a class action lawsuit from publishers and authors alleging it used copyrighted books and academic materials without authorization to train Llama.
Meta unveiled Muse Spark while continuing to offer the existing Llama.
Meta's new AI model Muse Spark is drawing early attention, but the monetization strategy is what matters.
Adversarial user prompts consistently lowered IFEval performance across 14 LLMs.
Rewriting a 13KB KB alone raised the Llama 3.1 8B extraction success rate from 60% to 100%.
Frontier open models require sustained funding, so they will likely end up moving toward a consortium.
Meta has announced a new AI model strategy that releases only some components to protect its core competitiveness.
Meta is pursuing a hybrid strategy of releasing some new AI models as open source to target the consumer market.
A self-distillation (SSD) technique that uses only the model's own outputs, without a separate teacher model or reinforcement learning, significantly improved LLMs' code generation performance.
AWS significantly boosted multi-turn agent performance using online RL with a small amount of data.
With SDK 2025.3, RNGD achieves 3x throughput and a 35% improvement in TTFT on Llama 3.3 70B.
Hugging Face released the StackLLaMA model and a guide for training LLaMA using RLHF.