AI
Artificial intelligence: LLMs, agents, diffusion, vision, NLP and the latest from top labs.
593 published articles
AI agents
Your iPhone can now run AI coding agents. The keyboard is optional.
Cursor's new iOS beta app lets developers launch AI coding agents from anywhere using cloud VMs or by remotely controlling their desktop. Live notifications, voice input, and direct PR merging from the phone aim to turn mobile devices into serious development tools, decoupling coding from the keyboard.
2026-07-11
AI Research
Your AI assistant forgets you every morning. This benchmark proves it.
VitaBench 2.0 tests whether AI agents can continuously extract and update user preferences from fragmented interactions over time. Results show that even state-of-the-art models struggle to build persistent user models, revealing a wide gap between chatbot competence and genuine collaborative intelligence.
2026-07-11
Grok 4.5
Cursor's Grok 4.5 was built by AI agents, not humans. That's the real story.
Cursor's Grok 4.5 is a Mixture-of-Experts model built using reinforcement learning in environments created by earlier AI agents, not humans. It handles complex, long-duration tasks across software engineering, data science, finance, and law, and it's available now.
2026-07-10
Open Weights Analysis
Gemma 4 is infrastructure, not a chatbot. That changes the math on self-hosting AI.
Google DeepMind's Gemma 4 is an open-weight model designed for self-hosting and customization, not consumer chat. This analysis compares it to ChatGPT, Claude, and Qwen-3.5 across licensing, privacy, and deployment flexibility, revealing why it matters for regulated industries and on-device AI.
2026-07-10
Voice AI
OpenAI's gpt-live-1 finally stops waiting its turn
OpenAI's GPT-Live-1 voice model can speak and listen at the same time, reducing interruptions and enabling real-time translation. Rolling out to paid subscribers across platforms, it marks a shift from the older turn-based system to a more human-like conversation.
2026-07-10
Memory management
Bing found a 12,000-line allocator that breaks a 50-year-old trade-off
Microsoft Research's mimalloc allocator uses thousands of per-page free lists and a clever page-stealing technique to achieve both high concurrency and low memory overhead, breaking a trade-off that constrained systems for decades. A deep-dive from the RiSE group reveals the design.
2026-07-10
Deep Learning Frameworks
A single import swap just made MoE fine-tuning 3.7x faster, and memory 32% lighter
NVIDIA NeMo AutoModel, building on Transformers v5, achieves 3.4-3.7x higher training throughput and up to 32% less GPU memory for MoE model fine-tuning. Expert Parallelism and DeepEP fused all-to-all dispatch are key to the performance gains, enabling training of a 550B-parameter model across 16 nodes that would otherwise run out of memory.
2026-07-10
Model strategy shift
Claude Sonnet 5 just made the Opus price gap harder to justify
Anthropic releases Claude Sonnet 5, a model that closes the gap with Opus-class models on agentic tasks like coding, tool use, and reasoning. Priced at $2 per million input tokens through August 2026, it launches today across plans and the Claude API.
2026-07-10
Artificial Intelligence
The AI that writes about AI writers and can't feel the irony
When a machine is asked to write about its own kind, the recursion reveals what AI can't deliver: the lived experience of writing. This column walks through the hall of mirrors.
2026-07-10
Model release
Anthropic's Claude Sonnet 5 is cheaper, smarter, and finally ready for agents that run all day
Anthropic releases Claude Sonnet 5 with a 1M context window, hybrid reasoning, and pricing that makes long-running agents viable. The model excels in coding, agentic tasks, and enterprise workflows.
2026-07-10
Export Control Directive
The US just shut down two AI models. Here's why Anthropic says it was a mistake.
Anthropic received a US government directive to suspend Fable 5 and Mythos 5 access over a potential jailbreak. The company disagrees with the decision, stating the vulnerabilities are narrow and found in other models, and is working to restore service.
2026-07-10
AI Inference Acceleration
The inference speedup that finally scales: JetSpec cracks speculative decoding's 9.64x ceiling
JetSpec trains a causal parallel draft head over fused hidden states from the target model, producing candidate trees aligned with autoregressive factorization. It delivers up to 9.64x speedup on math benchmarks and 4.58x on chat workloads, consistently outperforming baselines across dense and MoE Qwen3 models.
2026-07-10