Amazon Science
·
2 months ago
Amazon has built a comprehensive responsible AI pipeline across four stages of model development—pretraining, post-training, evaluation, and third-party monitoring—incorporating over 70 internal and external RAI tools and funding or publishing more than 500 research papers. The company teaches RAI principles during pretraining using curated datasets, optimizes policy adherence through reinforcement learning from human feedback, creates model-breaking datasets for evaluation, and partners with third-party experts to assess frontier risks like CBRN and cyberattack research. Amazon's approach aims to anticipate risks, teach models to handle ambiguity, and build systems that adapt to regulatory and social changes across different geographies and applications.
Import AI
·
2 months ago
Researchers and AI developers report that AI systems are demonstrating capabilities to automate components of AI research and development, with progress measurable across multiple benchmarks from code generation to scientific paper reproduction. AI systems have improved from solving 2% of real-world software engineering problems in late 2023 to 93.9% by 2026, while their ability to work independently on complex tasks has extended from 30 seconds in 2022 to 12 hours by 2026. If these trends continue, frontier AI labs may begin delegating larger portions of AI R&D work to autonomous systems, potentially accelerating the cycle of AI model development itself.
Last Week in AI
·
2 months ago
A podcast episode covering major AI developments in late April 2026 discussed OpenAI's GPT-5.5 release with improved coding capabilities, xAI's Grok Voice Think Fast 1.0 achieving 67.3% on voice benchmarks, and DeepSeek's open-sourced V4 model featuring 1M-token context windows. Key business moves included Google planning a $40 billion investment in Anthropic with 5GW compute commitment, Meta's AWS Graviton chip deal, and China blocking Meta's Manus acquisition. Research presented covered AI safety concerns including model sabotage of safety research, document corruption from LLM delegation, and neural network disruption via bit-flip attacks.
