Deep Learning Weekly
·
4 months ago
Deep Learning Weekly Issue 445 covers major AI model releases including Google's Gemini 3.1 Flash-Lite at $0.25 per million input tokens, OpenAI's GPT-5.3 Instant reducing hallucinations by up to 26.8%, and Microsoft's Phi-4-reasoning-vision-15B multimodal model, alongside research on diffusion language models and multimodal pretraining. Notable developments include Anthropic's refusal to remove AI safeguards amid government pressure, key departures from Alibaba's Qwen team following their small model release, and findings that LLM personalization features amplify sycophantic behavior. The week's updates span MLOps tooling like the Opik Claude Code Plugin for agent observability, infrastructure concerns about AI-generated code quality, and frameworks standardizing diffusion language modeling components.
Together AI
·
4 months ago
FlashAttention-4 is a new attention algorithm and kernel implementation that addresses asymmetric hardware scaling on Blackwell GPUs by pipelining tensor cores, special function units, and memory operations to reduce bottlenecks. The implementation achieves 1605 TFLOPs/s with 71% utilization on BF16, delivering 1.3× speedup over cuDNN and 2.7× over Triton by using software-emulated exponentials, tensor memory for intermediate storage, and 2-CTA MMA modes to reduce shared memory traffic. The design enables more efficient training and inference for large language models on next-generation hardware through careful co-optimization of algorithm and kernel implementation.
Together AI
·
4 months ago
Together Research announced seven research and product releases at AI Native Conf, including FlashAttention-4, a Reinforcement Learning API, ThunderAgent for agentic workflows, and optimization techniques like ATLAS-2. FlashAttention-4 achieves 2.7x faster performance than Triton on NVIDIA Blackwell GPUs, while Together Megakernel reduced latency from 281ms to 77ms for voice agents and together.compile improved image generation speed by 41%. These advances integrate research directly into production infrastructure to optimize AI model inference and training workloads at scale.
