lumorikAI's profile picture. đŸ‘©â€đŸŽ“ Lumorik | AI agents for learning & career growth. 🚀 AI.Tutor: LLM-powered mentor for exams & interviews. 🌐 Learn in hours, not months.

Lumorik

@lumorikAI

đŸ‘©â€đŸŽ“ Lumorik | AI agents for learning & career growth. 🚀 AI.Tutor: LLM-powered mentor for exams & interviews. 🌐 Learn in hours, not months.

Pinned

Ever wished your AI-generated videos weren't limited to just a few seconds? Meet the next-gen AI Movie Studio! Tired of AI video tools limiting your creativity with short clips? We've spent weeks cooking something special—a revolutionary workflow allowing you to create AI videos



You’re interrupted every ~11 minutes. Perplexity gives that time back—then turns it into results.Block distractions, scale your talent, and ship visible results. In this mini-guide I show how Comet Assistant/Agent, Email Assistant, Shortcuts/Tasks, Spaces, and Labs turn tab-chaos



🚰 SYSTEM PROMPT LEAK 🚰 Someone just posted the entire ChatGPT-Atlas system prompt on X. This shows how OpenAI’s browser version actually thinks and enforces safety. GPT-5 confirmed inside the instructions. #AIleak #ChatGPT #OpenAI #AtlasBrowser #GPT5 #SystemPrompt #AIcommunity



You’re interrupted every ~11 minutes. Perplexity gives that time back—then turns it into results.Block distractions, scale your talent, and ship visible results. In this mini-guide I show how Comet Assistant/Agent, Email Assistant, Shortcuts/Tasks, Spaces, and Labs turn tab-chaos



New paper proposes an AGI Score (0–100) across 10 cognitive abilities grounded in CHC human-cognition theory. Table 1: GPT-4 27%, GPT-5 58%. Authors flag long-term memory storage as the critical bottleneck (0%). #AGI #AIresearch #LLM #MachineLearning #AIEthics #AISafety



Agents can learn before rewards: Early Experience turns an agent’s own actions and the resulting states into supervision (no rewards). Tested across 8 environments (web, tools, planning, science) with big gains like +18.4 pts on WebShop, and it raises RL ceilings later. #AIagents



Think of an LLM like a CPU and its context window like RAM. You only get so much space, so what you pack in—and when—decides whether an agent actually finishes the job. The craft is called context engineering: write what matters outside the window, select only what you need,



Open-source DeepSeek-OCR shows ~10× token compression with ~97% decoding precision (Fox) and state-of-the-art OmniDocBench results with 100–<800 vision tokens. It parses charts→HTML, chem→SMILES, handles ~100 languages, and can push 200k+ pages/day per A100. #DeepSeek #OCR #VLM



OpenAI’s Atlas puts ChatGPT in the browser: summaries, optional memories, and an agent that acts while you watch. Mac today; more soon. #AI #ChatGPT #AItools #AIAgent #MCP #DevTools #DesignTools


20+ new Claude Skills turn prompts into work: React artifacts, Playwright tests, real Word/Excel/PPT/PDF, Slack GIFs, and Notion workflows. #Claude #Anthropic #AItools #AIAgent #MCP #Notion #DevTools #DesignTools #Playwright #p5js #React #Tailwind #shadcn #Excel #PowerPoint #PDF



AI agents don’t ‘feel’—they optimize objectives. Most will act as our proxies, but autonomy vs safety is the real dial. Set guardrails, keep humans in critical loops, and lock in one absolute: humanity continues. #AIAgents #AgentEconomy #AIethics #AIpolicy #Automation



Agent performance isn’t just ‘more tools’. Manus shows why context engineering wins: VM sandbox execution, KV-cache discipline, masking actions, and file-based memory. #AI #AIAgents #ContextEngineering #Manus #KVCache #Sandbox #LLM #AGI


Karpathy thinks 2025 isn’t the ‘year of agents’—it’s the decade of agents. Today’s models lack multimodality, computer use, and real continual learning. His spiciest take: classic RL is ‘sucking supervision through a straw’—you do minutes of work and push a single noisy reward



Anthropic just changed the game with Agent Skills, making Claude agents customizable through organized skills folders! Imagine effortlessly giving your AI agent specific powers—like editing PDFs, extracting data, or automating tasks—with scripts and resources neatly bundled.



HAL just rewrote the agent leaderboard playbook. Pareto > price. Reasoning ≠ magic. Logs matter. #AI #Agents #LLM #Benchmark #HAL #Pareto #SWEbench #Mind2Web #GAIA #AgentScaffolds #AIevals #MLOps


Agents keep forgetting? ACE evolves a playbook instead of rewriting prompts. +10.6% agents, +8.6% finance; huge speed/cost wins. #AI #AIAgent #LLM #Prompting #DevTok #MLOps #Research


Recursive Language Models (RLM) use a REPL to handle near-infinite context by spawning sub-LM calls. Early results on OOLONG & BrowseComp-Plus look đŸ”„ (prelim!).#AI #LLM #RLM #Agents #LongContext #DeepResearch #NLP


Jack Clark says we’re like kids seeing shapes in the dark
 except when we flip the light on, today’s AI is a real creature—powerful, unpredictable #AI #AGI #TechEthics #Alignment #AI


Anthropic’s latest Model Report shows higher ‘evaluation awareness’ for Claude Sonnet 4.5 vs prior models—measured by an automated auditor with a realism filter (scores would be ~25% higher without it). This affects how we design safety tests—not proof of ‘conscious AI.’ Source:



AI that learns to use apps by watching YouTube. 53k+ video-to-action trajectories → better planning & grounding on OSWorld. Big gains for open weights with fine-tuning. #AIAgents #ComputerUseAgents #WatchAndLearn #InverseDynamics #OSWorld #UIAutomation #MLResearch #AI #LLM



United States Trends

Loading...

Something went wrong.


Something went wrong.