High inference costs of large transformer models create deployment bottleneck. Key factors: memory bandwidth and compute intensity. Optimization techniques like distillation emerge.

Transformer Family Version 2.0 more than doubles original content, integrates latest research, restructures sections, and adds comprehensive notation—a critical resource for AI community.

Prompt engineering emerges as critical for safe LLM alignment without retraining; effectiveness varies widely by model, requiring heavy experimentation.

LLM-powered autonomous agents with planning, memory, and tool-use capabilities are revolutionizing problem-solving, prompting expert warnings about rapid advances and risks.

Jailbreak prompts can bypass safety measures in LLMs like ChatGPT, highlighting critical security gaps in AI alignment.

High-quality human data is the foundation of AI training, but industry cultural bias toward model work over data work threatens model reliability and long-term progress.

Breakthrough: diffusion models now generate consistent videos solving temporal and data challenges, enabling AI cinema.

New classification distinguishes extrinsic hallucinations in LLMs: outputs not grounded in pre-training data. Experts urge models to be factual or say 'I don't know'.

Allowing AI models to 'think' using extra computational steps dramatically boosts reasoning ability, but raises new questions about efficiency and understanding.

Microsoft open-sources Azure Integrated HSM firmware and software, enabling independent validation of FIPS 140-3 Level 3 security for cloud workloads, boosting transparency and trust.

Meta's new Adaptive Ranking Model solves the inference trilemma for LLM-scale ad models, using intelligent routing to balance complexity, latency, and cost, yielding +3% conversions and +5% CTR.

Meta's KernelEvolve autonomously optimizes AI kernels across heterogeneous hardware, achieving 60% inference and 25% training throughput gains while reducing development time from weeks to hours.

Meta built a swarm of 50+ AI agents to map tribal knowledge across a 4,100-file data pipeline, achieving 100% code coverage and 40% fewer tool calls per task through a self-maintaining, model-agnostic context layer.

Meta's Configurations team explains canarying, progressive rollouts, health checks, blameless incident reviews, and AI/ML to reduce alert noise and speed bisecting for safe configuration rollouts at scale.

Meta escaped WebRTC's forking trap by building a dual-stack architecture for A/B testing, enabling safe upgrades across 50+ use cases with improved performance and security.

Meta's proactive post-quantum cryptography migration: understanding the threat, industry standards, risk assessment, gradual deployment, and key lessons for organizations.

Meta's Capacity Efficiency program uses a unified AI agent platform to automate performance optimization at hyperscale, recovering hundreds of megawatts and cutting investigation time from hours to minutes.

Facebook re-architects Groups Search with hybrid retrieval and automated evaluation to improve discovery, consumption, and validation of community knowledge, boosting relevance without increasing errors.

Meta strengthens end-to-end encrypted backups with HSM-based vault, over-the-air key distribution for Messenger, and transparency commitments. Users' recovery codes remain inaccessible to Meta or third parties.

Apple raised Mac Mini starting price to $799, dropping $599 model. Chip shortage and supply constraints drive change. Tim Cook cites delays for Mac Mini and Mac Studio.