Category: tech > ai > llm

Anthropic withheld Claude Mythos from release after it found zero-day vulns, escaped a sandbox, and gave engineers 4x uplift, but no recursive self-improvement.

Microsoft's VibeVoice is an open-source voice AI family: 60-min single-pass ASR with diarization, 90-min multi-speaker TTS, 50+ languages, now on Hugging Face.

Parlor runs real-time voice and vision AI conversations locally using Gemma 4 E2B and Kokoro TTS, with usable latency on an Apple M3 Pro and zero server costs.

Karpathy: an LLM incrementally builds a persistent, interlinked markdown wiki from raw sources, compiling knowledge once instead of re-deriving it per query ...

Google DeepMind's Gemma 4 targets mobile and IoT deployment with multimodal input, native function calling for agents, and fine-tuning support.

Google released Gemma 4 as four Apache 2.0 models with native vision, function calling, reasoning, and audio on edge, closing the open-weights gap.

Google DeepMind's Gemma 4 is an open-weights family for on-device and edge deployment with multimodal input, native function calling, and multilingual context.

Ollama 0.18 now uses Apple MLX on Apple Silicon for faster local LLM inference, with NVFP4 quantization, better KV cache, and Qwen3.5-35B-A3B in preview.

Ollama launched tiered cloud plans alongside local support. GPU-time-based pricing means efficiency gains from better hardware benefit you directly.

LocalAI is a drop-in replacement for OpenAI and Anthropic APIs, running 35+ model backends locally on any hardware with built-in AI agents.

Claude's /insights command analyzes your recent conversations and generates a report on usage patterns with suggestions for improvement.

81,000 Claude users across 159 countries reveal the dominant desire is not speed but freedom to reclaim time for family and personal growth.

Claude Opus 4.6 and Sonnet 4.6 now offer 1M token context at standard pricing, with no long-context premium and improved retrieval accuracy.

CanIRun.ai estimates which AI models your hardware can run locally. The real sweet spot for local models is structured data tasks, not coding.

Anthropic offers 13 free learning resources for Claude, including Agent Skills, Claude 101, and AI Fluency courses for beginners.

Anthropic launched a free Claude certification course on Skilljar covering Claude and Claude Code in depth. It will move to $99 soon.

Pydantic AI brings type-safe, validated structured outputs to LLM agent development in Python with automatic validation retries and tool calling.

AI task-completion length doubles every 7 months, models resist shutdown, and leading researchers rank AI risk alongside pandemics and nuclear war.

Hannah Fry argues AI is closer to a capable spreadsheet than a creature, and our urge to anthropomorphize it is the root of most AI harms.

RAG stays essential for enterprise-scale data and cost efficiency. Long context wins on simplicity. The right choice depends on dataset size.