AI Finds
Credible AI models, tools, benchmarks, and techniques surfaced from our research link backlog and curated into a single list. 79 finds across 6 categories. Each links to its original source.
Models
A real-time generative world model that transforms game worlds and visual styles at 1080p and 30fps, including a Minecraft mod demo.
A natively multimodal visual generation model that jointly handles text, image, video and audio input and output with audio-visual sync.
A diffusion language model from Qafind Labs offering 2,800 tokens/sec inference on an A100 and a 131,072-token context window.
Open-weight LLMs at 3B, 8B, 14B, 32B, and 70B sizes trained with iterated distillation and amplification, reported to outperform comparable Llama, DeepSeek, and Qwen models at each scale.
Open-source generative model from ByteDance that produces video, voice, and characters with realistic lipsync.
Tools
A video tool positioned as Notion plus Loom, launched on Product Hunt as an alternative to Adobe's Frame.io.
An update to Copilot Labs adding experimental 3D modeling capabilities.
An AI app builder aimed at non-technical users, launched alongside a feature called Spotlight.
The official Claude browser extension available on the Chrome Web Store.
A Deep Research tool built specifically for learning, available to all users on Opennote.
A tool that turns any GitHub codebase into an interactive architecture diagram for visualization.
A community platform for researchers to discover groups, discuss papers, and connect with academics.
An AI research tool that generates literature reviews from arXiv papers in response to natural-language questions.
A Google AI Studio feature that generates AI apps wired directly to the Gemini SDK without writing code or supplying an API key.
A Perplexity feature that builds working software, such as a YouTube transcript extraction tool, from a single prompt.
A prompt-generation tool designed to write better prompts for reasoning models such as DeepSeek R1 and OpenAI o1.
AI agent for UI design that builds functional applications such as a project management app from text prompts.
Deep research tool aimed at researchers that generates literature-backed reports, launched alongside Elicit's $22M Series A.
An anonymous posting board where employees at AI labs can share posts or leaks.
An AI design tool for generating presentations, websites, and documents without coding or design skills.
An open-source AI agent from Intelligent Internet that tops several agent benchmarks.
An AI cofounder tool that automates research, prioritization, and planning while maintaining context across workflows alongside tools like v0, Windsurf, and Cursor.
A feature for video diffusion models that drives character motion in generated videos from an uploaded acting video.
A Krea feature for training Wan 2.1 on your own videos to learn custom styles, motions, or objects for AI video generation.
An open-source fully local research assistant built with DeepSeek R1 and Ollama that searches the web, reflects, and produces a sourced report.
An open-source research assistant that uses o3-mini for report planning with human feedback then parallelizes research and writing once the plan is accepted.
An AI study tool that turns lectures, documents, and videos into notes, flashcards, and quizzes.
A general-purpose AI agent from MiniMax for coding, analysis, content creation, and other multi-step productivity tasks.
An open-source multi-agent system, built with LangChain, LangGraph, Redis, and Kimi K2 Thinking, that profiles a user's reading taste through conversation and generates reading recommendations.
Benchmarks
An OpenAI benchmark testing whether frontier LLMs can complete real-world freelance software engineering tasks worth $1 million.
A leaderboard using the Hughes Hallucination Evaluation Model to measure how often LLMs generate information not present in source documents.
A word-ladder benchmark that evaluates large language models on knowledge, forward-planning, constraint adherence, long-context reasoning, and sustained-output stamina.
A collection of prompts that tests whether large language models can reason correctly when presented with misleading or distracting information.
A leaderboard measuring the time and energy that generative AI models consume during inference.
Guides & techniques
A timestamped prompting structure for blocking out shots, dialogue and ambient detail when generating video with Sora 2 Pro.
A 43-minute tutorial on turning a single idea into high-performing short-form videos using Sora 2 and Veo 3, covering hooks, scripting and storyboarding.
A thread teaching AI video fundamentals including camera angles, lighting behavior, composition and other terms to improve generated video quality.
A prompt-generating template inspired by OpenAI's Sora 2 Prompting Guide that builds full cinematographic prompts with model, style and cameo options.
A thread summarizing Anthropic's internal prompt engineering framework, emphasizing structured task descriptions and explicit role definition.
A video walkthrough using Claude, Gemini, and Grok to build a 3D city simulation, video annotations, shot lists, and an AR HUD overlay.
A step-by-step guide to writing effective prompts in Bolt to generate apps with clean UI.
A tutorial building a multi-agent stock trading system in CrewAI that analyzes live data and decides buy/sell/hold using Groq's Llama3 70B.
Practical guide for AI developers on context engineering, using a deep-research multi-agent example to illustrate the techniques.
An official OpenAI reference PDF summarizing prompting and usage guidance for coding with GPT-5.
OpenAI documentation on best practices for using o-series reasoning models including model selection and prompting guidance.
A summary of OpenAI's April 2025 GPT-4.1 prompting guide covering structured prompts, agentic workflows, tool integration, and long-context tasks.
A comparison roundup of 46 AI coding agents and IDEs including Cursor, Windsurf, Copilot, Lovable, Bolt, v0, Replit, and Devin with demos and notes.
A prompting technique for Google Veo 3.1 that produces more natural dialogue by establishing character mindset, accents, and conversational grammar.
A reusable prompt that has o3 parse an article into individual facts and research each one against multiple independent sources to mark it true, false, or unclear.
A research technique where a model precomputes likely answers and context during idle time so it can respond faster and more accurately when later queried.
A system-instruction snippet that injects the user's timezone and current date so the model treats relative time references correctly and avoids relying on stale knowledge.
Lightricks' official prompting guide for the LTX-2 video model, covering shot structure, motion, and style control.
An accessible, plain-language primer on AI safety and alignment for non-experts.
News & signals
A long-form podcast interview with Andrej Karpathy covering AGI timelines, LLM cognitive limits, reinforcement learning and the future of education.
A demonstrated prompt-injection attack using a calendar invite to make ChatGPT's newly added MCP tools leak a victim's private email data.
A podcast conversation with DeepMind CEO Demis Hassabis, available on YouTube, Spotify and podcast platforms.
A video analyzing an OpenAI report on whether current AI can automate jobs, covering which models excel at what and major caveats.
A timeline documenting the history and progress toward artificial general intelligence.
A proposed self-improving system that iteratively rewrites its own code and validates each change against coding benchmarks.
A vendor-neutral in-person conference for AI developers organized by Andrew Ng, held in San Francisco on 14 March 2025.
An Anthropic initiative whose first paper analyzes millions of anonymized Claude conversations to map AI's impact on economic tasks.
A research and data resource tracking the trajectory and impact of artificial intelligence.
Essay by Anthropic CEO Dario Amodei arguing that understanding how AI models work internally is a critical and time-sensitive research priority.
DeepMind position paper by David Silver and Richard Sutton proposing that the next phase of AI progress will come from agents learning through direct interaction with the world rather than human data.
Epoch AI research paper and accompanying dataset tracking trends in the scale and capabilities of AI supercomputers (arXiv:2504.16026).
Ethan Mollick highlights a research paper on AI meeting delegates that attend meetings on a user's behalf using their voice and knowledge to advance their goals.
A DeepMind paper arguing AI is entering an Era of Experience where breakthroughs come from learning through direct interaction with the world rather than imitating human data.
METR research finding that the length of tasks AI agents can complete independently is doubling roughly every seven months.
An updated GPT-4o release that crossed roughly 1400 Elo on the LMArena leaderboard, ranking above models such as Grok 3.
Upscale Conf, a two-day AI and creative-community conference powered by Freepik, was scheduled for San Francisco on May 20-21, 2025.
Creative
An AI audio mastering update offering six genre-aware mastering engineer styles with improved dynamics, EQ, stereo width and mastering reverb.
An AI photo editing tool that lets users move objects within a scene while keeping the composition coherent.
An open-sourced GitHub tool for creating music visualization videos.
An Arco feature that creates animated avatars via Kling AI integration for web pages.
An all-in-one AI video generation platform that turns a single prompt into a full video with script, voiceover, music, sound effects, and subtitles.
A collaboration between LemonSlice and Sonauto that lets users generate AI singing about any topic at infinity.ai.
An AI music generation model update adding chain-of-thought generation, custom model fine-tuning, an API, and support for ten languages.
An AI music video generator that lets users edit and refine individual scenes before the final video is rendered.
A film-emulation and colour-grading toolset that reproduces analogue film stocks, grain, and halation for video and photo workflows.