AI
Artificial intelligence: LLMs, agents, diffusion, vision, NLP and the latest from top labs.
593 published articles
Gemini triple model launch lands July 21
Google ships Gemini models: cheaper Flash, faster Lite, and a vetted cyber variant
Google's three-model Gemini drop targets production AI with better token efficiency and lower latency. 3.6 Flash cuts token usage by 17% versus its predecessor. 3.5 Flash-Lite runs at 350 output tokens per second. A cyber-focused variant ships exclusively to vetted partners.
2026-07-22
AI Models
The 33B model that just beat 137B models on coding benchmarks without changing hardware
Poolside's Laguna XS 2.1 improves SWE-bench Multilingual by 5.4 points to 63.1% while keeping the same 33B-total-3B-activated MoE architecture. The release includes quantized checkpoints, speculative decoding draft models, and an OpenMDW-1.1 license, making local AI coding more practical.
2026-07-21
Cyber Defense
Sakana's new cyber agent matches frontier models but tells you not to trust it alone
Sakana AI released Fugu-Cyber, a multi-agent orchestration model matching frontier cyber models on benchmarks. The company argues raw model access alone cannot fix enterprise security without human expertise and verification workflows.
2026-07-21
AI Tutorial
How to Write Effective Prompts for Claude: A Practical Guide for Professionals
A poorly structured prompt can cut Claude's answer relevance by half. This guide walks through context-instructions-format structure, few-shot examples, task-splitting, and reusable prompt patterns for professional work.
2026-07-21
LLMs & Inference
Speculative decoding is the nearest free lunch in LLM inference
Speculative decoding accelerates autoregressive generation by proposing candidate tokens with a cheap drafting mechanism and verifying them in a single target-model forward pass, without changing the output. This analysis breaks down the math, the main methods (draft models, EAGLE-3, DFLASH, Multi-Token Prediction, n-gram), and what acceptance rates actually mean for real-world latency.
2026-07-21
Model watch
The Gemini 3.6 flash leak has a far more interesting question: where's the pro?
A leak suggests Google is testing a 'Gemini 3.6 flash' model while the expected 3.5 pro remains absent. The version jump could mean a reshuffle of the Gemini roadmap, or a mislabeled internal build.
2026-07-21
Formal agent orchestration with stack-based execution
The 96% context cut that finally makes AI agent tool hierarchies work
A hierarchical, skill-based architecture for AI agents solves the bottleneck of flat tool registries. By organizing capabilities as a tree and using a stack-based execution loop, it scales sub-linearly with tool count while providing execution isolation and auditability for regulated domains like digital payments.
2026-07-21
Automated persistence
Grok now does your chores on a schedule, and email triggers are the real news
Grok now runs scheduled and email-triggered tasks without supervision. Users describe a task once, set a trigger, and Grok executes full conversations. Email triggers are a SuperGrok feature.
2026-07-21
AI IN HEALTH
Google's health AI bet is about reach, not just accuracy
Google Research details healthcare AI progress: Med-Gemini, AMIE diagnostic conversations, and global screening partnerships. The emphasis shifts from model performance to real-world access, with plans for millions of free screenings in India, Thailand, and Africa. But the direction is unmistakable: reach first, benchmarks second.
2026-07-21
AI regulation
Google bought itself a year of Android AI dominance, and Apple still can't touch it
Google secured a yearlong grace period to bring Android in line with the EU's DMA, giving Gemini more time to deepen its integration before rivals like OpenAI and Anthropic get comparable access. Apple, by contrast, was denied the same runway and still has no timeline for Siri AI in Europe.
2026-07-21
The cost engineering behind the model swap
Microsoft's Kimi K3 Copilot trial targets $600M in inference savings
Microsoft is preparing to test Moonshot AI's Kimi K3 inside Copilot in a bid to cut AI inference costs by up to $600 million. The company is shopping for cheaper alternatives to OpenAI and Anthropic models, but the Chinese open-weight model still needs to pass quality and latency checks.
2026-07-21
Moonshot AI's scheduler for automated agent tasks
Your AI tasks now schedule themselves. Kimi Work ships a cron engine.
Kimi Work adds a built-in cron engine to the Kimi AI platform, enabling scheduled execution of LLM calls, data processing, and agent workflows. It targets users who need 24/7 automation without local infrastructure.
2026-07-21