The article examines Meta's massive AI spending and high‑salary talent raids, compares its team‑building approach with Anthropic, DeepMind and DeepSeek, and argues that hiring alone cannot create a sustainable, top‑tier AI research organization.
The article examines how Meta’s CEO Mark Zuckerberg’s aggressive talent‑buying and commercial focus have sidelined Turing‑award winner Yang Li‑kun, detailing the restructuring of Meta’s AI labs, the clash over research directions, and the broader dilemma of balancing academic innovation with business imperatives in the AI industry.
In a candid interview, Mark Zuckerberg outlines Meta’s ambition to deliver a personal superintelligence through AI‑powered glasses, predicts a two‑to‑three‑year timeline for superintelligent models, and details the company’s multibillion‑dollar investment in gigawatt‑scale data centers, talent recruitment, and internal AI enhancements while acknowledging monetisation challenges.
Meta AI Vice President Ahmad Al‑Dahle addressed recent criticisms of the newly released Llama 4 model, denying claims of test‑set cheating, explaining quality variations as post‑release optimization, and acknowledging internal concerns that led to staff resignations and calls for transparency.
Meta's newly released Llama 4 quickly became a controversy as internal leaks reveal training‑data cheating, benchmark over‑optimization, and disappointing code‑generation performance that fails to match even older models, prompting resignations and widespread criticism from the AI community.
Meta’s newly released Llama 4 models—Maverick with 4 020 billion total parameters and Scout with 1 090 billion—feature a 128‑expert MoE, 10 million‑token context, native multimodal fusion, and FP8 training, delivering benchmark‑leading performance that outpaces GPT‑4o, Gemini 2.0 Flash and DeepSeek v3, while being openly available on Hugging Face and GitHub.
Meta’s latest open‑source Llama 3.1 model, available in 8B, 70B, and 405B sizes, is evaluated against top competitors and can be easily run locally on the 8B version using Ollama with a simple step‑by‑step guide.
Meta announced the open‑source Llama 3 series (8B and 70B parameters), detailing its decoder‑only Transformer architecture, 15 T‑token multilingual training data, superior benchmark scores over competitors, a limited 8K context window, and upcoming cloud and web‑based deployments.
Code Llama, Meta’s specialized extension of Llama 2, is a large language model fine‑tuned on code data that can generate, complete, and debug software across multiple languages, supports up to 100 k tokens, and is freely available for research and commercial use.
This article translates and analyzes Nathan Lambert’s commentary on Meta’s Llama 2 paper, detailing the model’s architecture, training data, RLHF pipeline, reward models, evaluation methods, safety improvements, licensing terms, and the broader implications for open‑source large language models.
