AI pulse last 7 days
Daily AI pulse from YouTube, blogs, Reddit, HN. Ruthlessly filtered.
Sources (41)▶
- criticalAndrej Karpathy
Były dyrektor AI w Tesli, OpenAI cofounder. Każde video to gold.
- criticalAnthropic
Oficjalny kanał Anthropic. Każdy release Claude'a.
- criticalComfyUI Blog
Release log dla integracji ComfyUI — Luma Uni-1, GPT Image 2, ACE-Step music gen, Seedance. Pokrywa video+image+music+workflow.
- criticalOpenAI Blog
Oficjalny blog OpenAI. Wszystkie release.
- criticalSimon Willison's Weblog
Najlepszy 'thinker' AI. Codzienne posty, deep insights, niska hype rate.
- highAI Explained
Głęboka analiza papers i benchmarków, niska hype rate.
- highAI Jason
Praktyczne tutoriale Claude Code, MCP, workflow vibe codingu.
- highBen's Bites
Daily AI digest, creator-friendly tone. Codex, model releases, agentic AI.
- highCole Medin
Vibe coding + agentic workflows + Claude Code MCP integrations.
- highFal AI Blog
Fal hostuje większość nowych AI image/video modeli — ich blog to wczesne sygnały premier.
- highHN: 3D & Gaussian Splatting
HN signal dla 3D generative — Gaussian Splatting, NeRF, image-to-3D. Próg 20 bo niszowa kategoria (top historic 182pts).
- highHN: AI agents / MCP
HN posty o agentach, MCP, vibe codingu z min 100 pkt.
- highHN: Claude / Anthropic
HN posty z 'Claude' lub 'Anthropic' z min 100 pkt.
- highHugging Face Blog
Releases dla image, video, audio, 3D modeli. Część tech-heavy — Gemini relevance odfiltruje noise. Downgraded z critical: za duży volume na 'must-read' status.
- highIndyDevDan
Claude Code power user, prompty, hooki.
- highInterconnects (Nathan Lambert)
AI policy + research analysis. Niska hype rate, opinionated.
- highLatent Space
Podcast + blog Swyx — wywiady z founderami i deep dives engineeringowe.
- highMatt Wolfe
Comprehensive AI tools weekly digest. ~700K subs.
- highMatthew Berman
AI news, model release reviews, agent demos. Wysoki output.
- highr/aivideo
Community AI video — Sora, Veo, Runway, Kling, LTX. Co naprawdę zaskakuje twórców.
- highr/ClaudeAI
Społeczność Claude'a — power users, tipy, problemy.
- highr/LocalLLaMA
Open-source LLMs, lokalne uruchamianie, benchmarks bez hype.
- highr/StableDiffusion
Największa community open-source image gen (700k+ users). Premiery modeli, LoRA, ComfyUI workflows.
- highRiley Brown
Vibe coding, AI builder workflows, Cursor + Claude tutorials.
- highThe Decoder
Niemiecki AI news outlet po angielsku, dobre breaking news.
- highTheo - t3.gg
TypeScript + AI dev workflows. Hot takes, narrative-driven.
- highYannic Kilcher
Paper reviews i deep dives w research AI.
- lowAI Weirdness
Janelle Shane — playful AI experiments, image gen quirks. Niski volume, unikalna perspektywa.
- mediumbycloud
AI papers digestible — między 2MP a Yannic Kilcher.
- mediumCreative Bloq
Design industry — gdzie AI ingeruje w klasyczne dyscypliny graficzne.
- mediumFireship
100-sec format, often AI/LLM + tech news.
- mediumfxguide
VFX i film industry — coraz więcej AI w pipeline. Profesjonalna perspektywa.
- mediumGreg Isenberg
Solo founder vibe — buduje produkty z AI, podcasty z indie hackers.
- mediumr/ChatGPTCoding
Vibe coding tipy, IDE setupy, prompty. Mix wszystkich modeli.
- mediumr/comfyui
ComfyUI workflows — custom nodes, JSON workflows, optymalizacje.
- mediumr/midjourney
Midjourney community — premiery v7+, style references, prompt patterns.
- mediumr/runwayml
Runway-specific community — premiery features, prompt patterns, comparisons z konkurencją.
- mediumr/SunoAI
Suno music gen community — nowe wersje modelu, lyric prompting techniques. Audio AI ma slaby RSS ecosystem.
- mediumTina Huang
AI workflows for data science, practical applications.
- mediumTwo Minute Papers
Krótkie streszczenia papers AI, świetne dla szybkiego scan'a.
- mediumWes Roth
AI news z bardziej clickbaitowym tonem — filtr Gemini odsiewa hype.
Vibe coding and agentic engineering are getting closer than I'd like
As AI agents become more reliable, the focus of software quality is shifting from 'clean code' to 'proven real-world usage' and human-led architectural oversight.
Simon Willison explores the blurring lines between 'vibe coding' (non-expert, result-oriented) and 'agentic engineering' (professional, process-oriented). He admits that as tools like Claude Code improve, even experienced engineers are tempted to skip line-by-line reviews, treating agents as 'black box' internal teams. This shift challenges traditional software evaluation; since AI can generate perfect-looking READMEs and tests in minutes, real-world usage becomes the only true metric of quality. Willison also notes that while productivity has jumped from 200 to 2,000 lines a day, the inherent complexity of software remains a barrier that still requires human expertise to navigate safely.
Simon Willison's Weblog·opinion·05/06/2026, 02:24 PMdatasette-llm 0.1a7
You can now set global default parameters like temperature for specific LLMs within Datasette, ensuring consistency across data enrichment tasks.
Simon Willison released version 0.1a7 of the datasette-llm plugin, introducing a new configuration mechanism for model defaults. This update allows users to define specific settings, such as temperature or system prompts, that apply globally to a chosen model within the Datasette ecosystem. For example, you can now ensure all automated data enrichment tasks use a temperature of 0.5 by default without manual overrides. This is a significant step in making LLM-powered data workflows more predictable and easier to manage. The plugin serves as a bridge between structured SQLite databases and generative AI capabilities, streamlining how developers interact with models.
Simon Willison's Weblog·tooling·05/05/2026, 01:56 AMllm-echo 0.5a0
Test how your applications handle model 'thinking' blocks for free using the new simulated reasoning feature in the llm-echo plugin.
Simon Willison has released version 0.5a0 of llm-echo, a utility plugin for his LLM CLI tool designed for automated testing. The plugin provides a 'fake' model that echoes inputs instead of calling an actual API, saving costs during development. This update introduces a new '-o thinking 1' option, which simulates the reasoning blocks produced by modern models like DeepSeek-R1 or OpenAI's o1. By outputting a fake reasoning block to standard error, it allows developers to verify that their parsers and workflows correctly handle multi-stage model outputs without incurring any API fees.
Simon Willison's Weblog·tooling·05/05/2026, 01:31 AM
Relevance auto-scored by LLM (0–10). List shows top 30 from the last 7 days.