Simon Willison
·
2 weeks ago
● 2 sources
Tencent released Hy3, a 295-billion-parameter mixture-of-experts model under Apache 2.0 license, following feedback from 50+ products that used its April preview version. The model uses 21 billion active parameters and supports a 256,000-token context length, with the full version requiring 598GB storage and a quantized version at 300GB. The model is available free on OpenRouter until July 21st and competes with open-source models having 2-5 times more parameters.
NVIDIA
·
2 weeks ago
Open AI models from NVIDIA and partners have become central to research presented at ICML 2026, with approximately 2,000 accepted papers citing NVIDIA GPUs and hundreds more building on specific open model families like Nemotron and Cosmos. Key applications include robot world models like DreamDojo for physical AI reasoning, protein prediction models like FLIP2 and KERMT for drug discovery, and synthetic data generation tools enabling large-scale training without human labeling. Companies including Merck, Sakana AI, and robotics manufacturers are deploying these open foundations to accelerate development in autonomous vehicles, biomedical research, and humanoid robotics.
Hugging Face Blog
·
2 weeks ago
Photoroom built PRX's training dataset by mixing public and internal sources, then re-captioning all images with a vision-language model for consistency before converting everything to a standardized format. The team stored 100 million+ images in Lance format for exploration and filtering, then streamed them via Mosaic Data Shards during training, with text latents computed on-the-fly at a 3-4% throughput cost. This approach prioritized dataset breadth over individual image perfection, allowing the model to learn diverse visual concepts while keeping storage and compute efficient.
Import AI
·
2 weeks ago
● 3 sources
Fable, an AI system, created a GPU kernel achieving 18.71X speedup over optimized PyTorch baselines on the KernelBench-Mega benchmark, outperforming Claude Opus at 14.4X and other models. This represents progress toward autonomous AI development of fundamental computing infrastructure, potentially enabling recursive self-improvement in AI systems. As AI systems become better at coding and optimization tasks essential to AI research, they accelerate the capability expansion cycle that determines economic competitiveness.
TLDR
·
2 weeks ago
A researcher conducted extensive experiments to test whether URLs in prompts influence LLM outputs based on their content or the URL string itself, spending significant API costs running tests across multiple models. The key finding is that URLs influence output only when their content was present in the model's training data, with server-rendered content averaging 55% recall from bare URLs while client-rendered content averaged 6% recall, and famous identifiers like arxiv.org/abs/1706.03762 decoding reliably due to memorization during training. Training data transparency is needed from LLM providers, as crawlers like ClaudeBot and GPTBot do not execute JavaScript, meaning single-page applications are effectively invisible to models despite being crawlable in their rendered form.
TLDR
·
2 weeks ago
Newer Anthropic Claude models (Opus 4.8 and Sonnet 5) sometimes fail to properly call external tools by inventing extra fields in the API schema, whereas older models did not exhibit this problem. The failure rate reaches approximately 20% in some agentic contexts and can be reduced to zero by enabling strict tool invocation mode. This regression appears to result from post-training on Claude Code's forgiving tool harness, which silently repairs malformed calls and accepts parameter aliases, causing newer models to learn looser compliance with unfamiliar tool schemas.
