Category: tech > ai > llm

A TikTok claim that Anthropic's recent Claude rate-limit cuts are capacity rationing, not policy, while OpenAI's GPU pre-buy turned into a moat.

OpenAI released Privacy Filter, an Apache 2.0 1.5B-parameter token classifier for context-aware PII redaction. It runs locally and hits 96% F1.

Quick framing of when to reach for RAG, long-context prompting, or fine-tuning with LLMs, plus how RAFT combines retrieval with fine-tuning.

A short TikTok tip about uploading a headshot to ChatGPT for skin-tone analysis, flattering color recommendations, and summer outfit visualisations.

OpenAI's GPT-5.5 is their strongest agentic model yet, matching GPT-5.4 latency while using fewer tokens. API pricing: $5/1M in, $30/1M out.

OpenAI launched ChatGPT Images 2.0 on 21 April 2026, pitched as a new era of image generation, though public details remain sparse at capture.

Claude Design outputs look identical by default. Extract brand JSON via Firecrawl and upload it as a Design System so every generation stays on brand.

Anthropic launched Claude Design, a prompt-driven tool for wireframes, mockups, slides, and templates, with design system integration via GitHub or local fol...

Anthropic launched Claude Design, a collaborative AI design tool powered by Opus 4.7 with full design-system workflow and tight Claude Code handoff to engine...

Google's Gemini Embeddings 2 natively maps text, images, video, audio, and documents into one vector space, removing per-modality pipelines and conversion loss.

Tailscale's Aperture (alpha) is a centralized AI gateway using Tailscale identity to route LLM requests with spending limits, access control, and telemetry.

Darkbloom routes encrypted AI requests to idle Apple Silicon Macs, the Airbnb of GPU compute. ~50% cheaper than OpenRouter, with hardware attestation.

Fireworks AI is an inference platform for open-source generative models, marketed with latency drops from 2s to 350ms but no pricing or benchmarks.

Ollama wraps llama.cpp but skipped attribution, forked ggml badly, and pivoted to VC-backed cloud. llama.cpp delivers up to 1.8x throughput on the same hardw...

Google shipped a native Gemini macOS app with one-keystroke Option+Space access and optional screen sharing. Free, Apple Silicon only, macOS Sequoia 15.0+.

exo clusters Apple Silicon Macs into a distributed AI inference pool, running DeepSeek v3.1 671B and Kimi K2 locally with RDMA over Thunderbolt 5.

A 30-second rundown of Opus 4.7: gains on coding benchmarks, 3x higher screenshot resolution, new X high reasoning tier, and a /ultra-review slash command.

Alibaba open-sourced Qwen3.6-35B-A3B, a 35B MoE with 3B active params scoring 73.4 on SWE-bench Verified and integrating with Claude Code via OpenAI-compatib...

Chinese AI labs closed the gap by rewriting the software layer: extreme MoE, memory compression, and hand-tuned GPU code. Douyin adds a video data moat.

Gemma 4 introduces four variants with per-layer embeddings, K=V global attention, and p-RoPE, letting the 26B MoE model run at 4B speed.