Amazon Science
·
3 months ago
Amazon researchers addressed three problems in LLM-based text-to-speech systems: accent leakage in multilingual voice cloning, limited expressiveness, and reliability issues like hallucinations and cutoffs. They used locale-specific fine-tuning with LoRA, classifier-free guidance for prosody, and chain-of-thought reasoning to predict phonemes and duration before generation, reducing critical errors to less than one second per hour on long-form text. These techniques improved speech quality scores by 5% to 20% across nine language locales compared to their previous model.
Last Week in AI
·
3 months ago
● 2 sources
This podcast episode covered major AI developments from March 2026 including OpenAI releasing GPT-5.4 mini and nano models with 400k-token context windows at higher per-token costs, Mistral open-sourcing its Small 4 model family with 119B total parameters, and multiple companies advancing agent systems including Meta's Manus launching a Mac agent and Nvidia announcing NeMo for sandboxed agent runtime. OpenAI shifted strategic focus toward enterprise productivity amid competition, Microsoft reorganized its AI division, Meta delayed its next model rollout, and new safety research addressed steganography detection, chain-of-thought faithfulness, fine-tuning defenses, and cyber-attack evaluations. The episode included technical discussions of Attention Residuals and Mamba-3 sequence modeling research alongside business updates about hardware forecasts and ByteDance accessing top Nvidia chips.
Hugging Face Blog
·
3 months ago
Falcon Perception is a 0.6-billion-parameter early-fusion Transformer model that handles open-vocabulary image segmentation and object grounding from natural language prompts using a single unified architecture rather than separate modular pipelines. The model achieves 68.0 Macro-F1 on the SA-Co benchmark compared to 62.3 for SAM 3, with particularly large gains on attribute-heavy (+8.2 points), food and drink (+12.2 points), and spatial understanding tasks (+21.9 points on PBench's spatial tier). The unified design allows the model to excel at compositional prompts requiring text reading, spatial reasoning, and dense scene understanding where modular approaches typically fail.
OpenAI Blog
·
3 months ago
Gradient Labs has deployed AI agents powered by GPT models to handle banking support tasks for customer accounts. The system uses GPT-4.1 and smaller variants (GPT-5.4 mini and nano) to manage automation with low latency performance. Banks can now route standard customer inquiries to AI rather than human support staff.
Together AI
·
3 months ago
Together AI's kernels team, led by Dan Fu and Tri Dao, develops GPU optimization software that bridges the gap between AI models and hardware efficiency. The team achieved a 3.6x speedup for a real-time voice agent company, reducing latency from 281ms to 77ms on Llama-3.2-1B, and created ThunderKittens, a library that reduced CUDA code from 1,000+ lines to 100-200 lines for adapting kernels to new NVIDIA Blackwell GPUs. This kernel optimization work directly impacts production AI systems by determining inference costs, training time, and whether AI applications feel responsive to end users.
Hugging Face Blog
·
3 months ago
Gradio released gradio.Server, which allows developers to build custom frontends using React, Svelte, or vanilla HTML/JS while retaining Gradio's backend features like queuing, API infrastructure, and ZeroGPU support. The Text Behind Image demo application demonstrates this with approximately 50 lines of Python backend code and a ~1300-line vanilla HTML frontend that manages image layering and text rendering. This enables developers to choose between Gradio's built-in UI components or entirely custom frontends while maintaining access to Hugging Face Spaces hosting, concurrency management, and the gradio_client SDK.
