AI
Artificial intelligence: LLMs, agents, diffusion, vision, NLP and the latest from top labs.
593 published articles
AI Coding Tools
Claude Code can now route to a whole school of models
Sakana AI's Fugu multi-agent system gains a Claude Code interface, allowing developers to orchestrate a pool of specialized models from within Anthropic's closed-source coding environment. The integration highlights a broader shift: model routing is becoming a platform feature, not a user concern.
2026-07-26
AI Research
The gradient wall that blocked neural networks from reasoning just got a soft demolition
SoftReason is a fully differentiable neuro-soft-symbolic architecture for deductive reasoning. It removes the gradient barrier between perception and deduction, enabling end-to-end training over high-dimensional perceptual data and knowledge graph triples. On KVQA, it shows how soft reasoning can unify grounding, evidence injection, and closure.
2026-07-26
Data Infrastructure
Your AI agents are starving for clean data. Alibaba just automated the kitchen.
Alibaba Cloud launches DataBridge Agent to automate multi-source data collection, parsing and structuring for AI workflows. The tool tackles the growing disconnect between powerful models and messy enterprise data, acting as the chef your AI agents need.
2026-07-26
European autonomy
Poland just bought into ElevenLabs. The message is not about the check size.
ElevenLabs has taken state capital from Poland. The investment, routed through Vinci and paired with a new research lab in Warsaw, is a bet on keeping European AI founder talent from drifting to the US. The check size matters less than the signal.
2026-07-26
AI Research
The AI that tracks what students skip, not just what they click
MR-ConceptGCN combines personal knowledge graphs with multi-relational GCNs and SBERT embeddings to preserve the difference between a student's correct answers and skipped content. In a 31-person study, it improved recommendation accuracy, usefulness, diversity, and satisfaction over baseline approaches.
2026-07-26
Adversarial AI and Data Poisoning
The comment section is now a viable weapon against the models it trains
University of Washington researchers demonstrate that poisoning pretraining data via automated comment spam is a realistic threat. Their HalfLife analysis shows 0.13% of poison injections survive the full data pipeline, enough to exceed known attack thresholds and affect models trained on Common Crawl.
2026-07-26
Artificial Intelligence
480ms latency, Whisper-level accuracy, an open recipe, Nvidia's Canary redefines real-time ASR
Nvidia's Nemo Canary matches Whisper-level accuracy on LibriSpeech and Common Voice while streaming at 480ms. The open source release includes the full training recipe, making it a rare reproducible contender in real-time ASR.
2026-07-26
Analysis
The 0.8B model that just broke document parsing's pipeline monopoly
OvisOCR2 (0.8B parameters) from Tencent scored 96.58 on OmniDocBench, the first end-to-end model to top the leaderboard. The model uses reinforcement learning and distillation to outperform larger pipeline systems, challenging the necessity of multi-stage architectures.
2026-07-26
Model efficiency
A 4-billion-parameter model just did what 30-billion systems couldn't: fit on one GPU
Microsoft's Mage-Flow is a compact 4-billion-parameter image generation and editing model that matches larger systems like Qwen-Image and FLUX.2 while running on a single A100 GPU at interactive speeds. Its key innovation is a lightweight tokenizer that cuts encoding costs by 12x and challenges the assumption that bigger models are always better.
2026-07-26
AI & Science
Anthropic Launches Claude Science, an AI Workbench for Scientific Research
Anthropic launches Claude Science, a research environment that integrates fragmented scientific tools, manages compute, and produces fully auditable outputs. Available in beta for Pro, Max, Team, and Enterprise users.
2026-07-26
Artificial Intelligence
Better AI detectors might make people use AI more, not less
Imperfect LLM detectors can distort user incentives, leading to more AI use and lower quality outputs. The paper's findings challenge the naive assumption that detection tools cleanly reduce machine-generated content.
2026-07-26
Speech synthesis
Alibaba's new TTS model speaks 16 languages, laughs on command, and might actually ship
Alibaba's Qwen-Audio-3.0-TTS is a production-oriented speech synthesis system that combines a low-frame-rate tokenizer with progressive training. It supports 16 languages, 20 Chinese dialect regions, and natural-language instructions for emotion, pace, and speaking style.
2026-07-26