← Back to leaderboard
oliviscusai

Oliver Prompts

Verifiedactive
@oliviscusai

Helping you understand AI. Software Engineer. Opinions are personal. DM for collabs.

RosterAI Agents & Productivity ToolsTech Career & Coding Education
21,849 followers · 11 following · Last polled 1h ago
Perf. Score
56.1
Rank #186
Value Score
19.6
#219 of 268 priced
Median views
2.2K
p25 1.7K · trailing 7d
Followers
22K
+1.3% 7d
ER (impr.)
3.93%
eng ÷ impressions
ER (foll.)
2.59%
eng ÷ followers
Est. CPM
$160.96
QT-rate basis
WoW views
-50.7%
▼ falling

Rates & value · est. from 2.2K median organic views

overpriced
Quote tweetbasis
$350
$160.96 / 1K views
Post
—
no CPM
Thread
—
no CPM
Retweet
—
no CPM
Value Score 19.6 · rank #219 of 268 priced$5384.62 / 1K engagementsrate age unknown (pre-audit import)

Trends

Follower growth
Median views / day
Engagement rate
View distribution · trailing-7d posts

Each blue dot is one organic post. The Performance Score ranks by the median (2.2K), not the mean (14K) — so a single viral post doesn't inflate a creator's reach. The p25 floor (1.7K) is the level 75% of posts clear.

Recent posts · ranked by views

PostPostedViewsLikesRepostsRepliesQuotesBookmarksER
The entire RAG industry is about to get cooked. Researchers developed a new RAG approach that bypasses almost everything traditional RAG depends on. - No vector DB - No data embeddings - No chunking - No similarity search It's called PageIndex. Instead of splitting your documents into chunks and loading them into Pinecone, it creates a tree index that lets the LLM reason through them like a human reading a book. 98.7% on FinanceBench. Outperforms every vector RAG on the leaderboard. 100% free. Open source.frozen10d ago457K4.9K519179678.8K3.15%
Wow Nvidia has just released a tool to turn any image into an explorable 3D world And they made it 100% open source 🔥 - Model available on Hugging Face - UI code available on GitHub Lyra 2.0 creates a world you can walk through, turn around in, and even drop a robot into for simulation.frozen7d ago123K2.1K
246
37
24
2.5K
3.94%
You can now run a full WhatsApp API on your own server for free. It's called OpenWA. WhatsApp Business API costs businesses money and ties you to their platform. OpenWA gives you the same automation power, self-hosted and free. → Send and receive messages through a simple API → Run multiple WhatsApp sessions simultaneously → Manage sessions and webhooks with a modern web dashboard → Connect it to n8n, Home Assistant, and your own custom apps One Docker command and your local dashboard is up and running. 100% Free. Open Source.frozen8d ago64K345331827521.80%
steal Alex Hormozi's $100M playbook for free with this repo right here. Hormozi's own paid AI advisor, ACQ AI, costs $3,000 upfront plus $299/month. this repo wires the same public frameworks into Claude Code instead. what's inside: > 12 playbooks: pricing, closing, ads, hooks, retention, LTV, branding > the full 10-stage scaling course, $0 to $10M+ > 5 ready-to-use prompts for offers, scripts, & ad copy use it with: git clone https://t.co/TJcm7tT52n cd hormozi-claude-code-brain && claudefrozen27d ago28K32321918714.43%
Want to know how a feature works in an app, down to the binary level...? Meet REA (Reverse Engineer Anything), an open-source toolkit that lets your AI agent reverse-engineer software through MCP and CLI. - Decompile apps and native binaries - Follow execution and examine the evidence - Rebuild features in your own stack3d ago26K587701036555.14%
GPT-6.1 Sol + Blender MCP is INSANE 😱 I just turned some multi-view Gundam photos into a textured 3D model I could rotate, zoom, and explore in UNDER 10 minutes. No modeling experience. Just reference images and prompts. Here’s how I did it ↓ https://t.co/ovzBBuz9BWfrozen5d ago24K365351934273.48%
GLM-5.3 Flash Abliterated is FREE 😳 a new research version of GLM-5.3 Flash is now open for testing - 1M context window - up to 131K output tokens - 33.5 tokens/sec - free to use - OpenAI-compatible API built for coding and AI research available directly in the cloud test it here: https://t.co/JHckmfk18I free access is available now 👀frozen4d ago18K136111211351.65%
Your PC can now run an entire AI studio for free. Locally Uncensored puts chat, image generation, video generation, and a coding agent in one desktop app for Windows and Linux. - Chat with local models and your own documents - Generate images and videos on your GPU - Write code and review changes before applying them - Run agents with tools and MCP connections Install it, pick a model, and start. It shows which models fit your hardware before you download. No Docker. No terminal. No cloud required for the local workflows. 100% Free. Open Source.1d ago17K510631006056.79%
Peter Steinberger, the creator of OpenClaw, open-sourced his entire agent setup. The idea: one shared folder that every agent on his machines reads. A single AGENTS.MD stores his rules, symlinked into Claude Code and Codex so both follow the same instructions. Every repo starts with: READ ~/Projects/agent-scripts/AGENTS.MD BEFORE ANYTHING. Update a rule once, and every project picks it up. His 69 skills also live in one place, with short descriptions so agents know what to load. A sync script links them into both agents. Fork it, add your own rules, and remove skills you don’t need. Replace his personal hosts and accounts before using it.frozen9d ago7.8K78181601152.91%
this command turns your computer into a 24/7 ai assistant that lives inside telegram, discord, or slack. cowagent is a free, open-source ai agent that plans tasks, uses tools, and remembers past conversations, connecting to whichever chat app you actually use. the setup, start to finish: 1. go to https://t.co/csdgxtBZe6 and grab the desktop client for macos or windows. 2. open it. on mac, gatekeeper may flag it as from an unidentified developer. 3. connect claude, gpt, gemini, deepseek, or any of a dozen others with an api key. take it further: > connect to telegram/discord with an api token, > install a skill from the skill hub 100% free. open source.frozen27d ago6.9K28512240.87%
A filmmaker's paid Claude account got banned automatically at 2:26am, "user well-being" flagged, grant deadline days away. The email gave zero specifics: "Our automated systems detected a violation of our user well-being Usage Policy." No conversation, message, or behavior named. Appeal window: up to 10 business days. Almost 2 years of use, grant materials, and film project code tied to that account, gone overnight. Dozens replied with the same story. One user's own support bot admitted their billing showed a paid plan while the account entitlement showed Free, right before the ban. Appeals kept coming back upheld, no new information. Recurring advice: don't store anything critical only inside Claude. Keep code in Git, context somewhere separate.frozen1mo ago6.4K61595181.53%
You can now use Meta’s Muse agent for free. It's called OpenMuse, from CopilotKit. A personal agent with its own browser, terminal, and files, built to keep going while you're away. → Ask for an outcome, follow the plan, review its actions, come back to the result → Take control of its browser session whenever you need to step in → Runs on iOS, Android, and web, currently in alpha 100% open-source.frozen10d ago5.1K591070843.11%
You can now run a full customer support platform on your own server for free. It's called Chatwoot. platforms like intercom and zendesk charge per agent, per month, and lock your conversation history on their servers. chatwoot gives you the same inbox, self-hosted and free. → one inbox for live chat, email, whatsapp, facebook, instagram, telegram, and sms → an ai agent (captain) that handles routine replies automatically → a built-in help center for self-serve faqs and guides → canned responses, auto-assignment, and team automation to scale support one-click deploy to heroku or digitalocean, or run it yourself with docker. 100% Free. Open Source.frozen7d ago4.9K641450813.35%
This AGENTS.md fixes the three most annoying habits of Codex and Claude Code: - Writing backward-compatible code for a project nobody uses yet - Keeping old code and docs it no longer needs - Adding a pile of tests just to change a button color After months of daily use, these are the rules I wish I'd added from day one: if the project has no real users yet, skip compatibility and migrations. Build the simplest code that works. Then follow four steps: docs first → tests → smallest implementation → clean up. Copy and paste it into your AGENTS.md 👇3d ago4.4K37451552.32%
You can now use Openai dots agents for free. It's called Open Dot. Each agent gets its own browser that stays logged in, and runs on your own Mac with your own API key. → Connects to Gmail, Calendar, Slack, Notion, GitHub, and 1,500+ other apps → Add an E2B key and agents keep running even while your Mac sleeps → Works with OpenAI or open models like Kimi, DeepSeek, and Qwen 100% Open Source.frozen5d ago2.8K36361372.96%
People one week into a yearly AI video subscription reading this: https://t.co/lkIsx6sLNUfrozen24d ago2.5K730030.51%
This prompt turns ChatGPT Dots into a 200 IQ agent with complete context and memory for your business: https://t.co/wHQB5NOrvgfrozen7d ago2.3K15130231.80%
Another insane Jev use case! Jev is making it dramatically cheaper to evaluate what actually happened inside an agent run. And now, someone has open-sourced a self-improving memory layer that turns those insights into reusable knowledge across agent harnesses: > Claude Code > Codex > Cursor > OpenCode, and 20+ more Beacon by @asymptotelabs continuously records agent activity across harnesses and uses Jev to identify which runs are worth learning from. It converts the most valuable workflows, corrections, and debugging patterns into reusable skills.frozen17d ago2.0K24490253.06%
Karpathy just shared a few ways he makes sense of AI output. The first one: have the AI write in the English standard used for 𝗮𝗶𝗿𝗰𝗿𝗮𝗳𝘁 𝗺𝗮𝗶𝗻𝘁𝗲𝗻𝗮𝗻𝗰𝗲 𝗺𝗮𝗻𝘂𝗮𝗹𝘀. The standard is ASD-STE100, designed for aircraft mechanics who don't speak English as their first language. Its dictionary contains 𝗮𝗿𝗼𝘂𝗻𝗱 𝟵𝟬𝟬 𝘄𝗼𝗿𝗱𝘀. Instructions stay under 20 words, with one instruction per sentence. LLMs already understand the standard, so asking for it produces short, readable explanations with one idea at a time. Need more clarity? Ask for a diagram, then an interactive HTML page. For the most visual explanation, ask for an explainer video. Especially useful for: → understanding the login flow in a repo you've inherited → checking what a large AI rewrite changed and which parts it affects → figuring out what caused a confusing error Add this to your global CLAUDE.md or AGENTS.md 👇frozen5d ago2.0K23180272.93%
Governments pay millions for this kind of system. Someone just open-sourced it for free. It's called Crucix. It monitors the world and messages you when something changes. Pulls data from 26 live sources every 15 minutes: - Satellite fire detection, flight and vessel tracking - Conflict zones and economic indicators - Live crypto and market prices - Social sentiment from Telegram intel channels It works both ways. Sends alerts to Telegram and Discord. Message it back or type /brief for a summary. No cloud, no subscription. Runs on your own machine. 100% Free. Open Source.frozen4d ago2.0K29320373.64%
This tool is literally Higgsfield AI but FREE for good. it's called wan2gp. a full ai video studio built specifically for people without expensive hardware. runs on as little as 6gb of vram, works on old rtx 10-series cards and 8gb laptops. everything stays local, no uploads, no caps, no watermarks. what one app gives you: > text-to-video and image-to-video generation > the best open models built in: wan 2.2, ltx-2, hunyuan video, flux > a full browser interface with a queue system > lora support to customize any model > mask editor and prompt enhancer included a 5-second clip renders in minutes on a mid-range gaming rig. no subscription, ever. 100% free. open source.frozen13d ago1.9K13340272.44%
GLM-5.3 UNCENSORED 😳 A new EXL3 quantized version of GLM-5.3 is now available: - 3.0 bits per weight - MIT licensed - MoE architecture - Safetensors format Grab it here: https://t.co/zNPmMP2lVH https://t.co/lbnAQfySuyfrozen6d ago1.9K18300151.91%
suno costs $10-30/month for ai-generated songs. this open-source ui runs the same kind of model completely free, on your own gpu. ace-step ui is a free, open-source interface for ace-step 1.5, generating full songs with vocals up to 4+ minutes, running entirely on your own hardware. > generates complete songs with vocals > use a reference audio file to steer style > built-in stem separation > one-click install through pinokio 100% free. open source.frozen19d ago1.8K15340252.61%
🚀 Breaking News! Muse AI Related Files and Skills Leaked!!! The key Muse AI files have been organized into a public repository called MuseAI-Skills. It includes 68 independent skills covering parallel research, goal management, memory cleanup, travel planning, document creation, video orchestration, email, calendar, and more. You can study how the agent delegates tasks, requests authorization, handles failures, and verifies its work before completion. The repository also includes permission lists, evaluation scenarios, database docs, scripts, and bilingual READMEs.frozen9d ago1.8K20540132.38%
This is the ChatGPT moment for Voice AI. Meet FluxTranslator, an open-source Windows tool that brings speech recognition, translation, and voice playback into a desktop overlay. - Turn speech into text - Translate spoken words - Read text aloud with TTS All in one overlay on your desktop.2d ago1.7K14220141.83%
i tested Opus 5.5 with one robot drawing. it built a 271-part 3D model i can take apart in my browser. zero manual modeling. Opus 5.5 drove Blender through MCP and wrote every part in Python. what impressed me: > caught and fixed its own render mistakes > invented an interior: plasma core, gyro, fans > built the interactive site and every animation > worked around Blender API errors on its own the whole thing runs from one html file and a 2.3 MB model.frozen15d ago1.5K1110071.27%
jev just tied five much larger models on a real science benchmark, and nobody’s talking about it. a new paper tested small scientific judgment calls, and jev matched perfect scores while being the cheapest and fastest option by a wide margin. → 100/100, matching Opus 5, Sonnet 5, Astra, Sol, and Kimi K3 → 0.335s response time vs 1.385s for the next-fastest perfect scorer → $0.06 per 1,000 requests vs $3.05 for the next-cheapest worth flagging: every reference label in the test was "inconsistent," so guessing that every time would've also scored 100% on labels alone. read the paper, then check out the complete jev guide below.frozen8d ago1.5K914081.47%
Claude Code just got an open-source control center for Mac users. AutoHarness takes one goal and organizes the work into a single agent, a loop that runs and checks the work, or a coordinated agent team. You can manage the whole run from one desktop app: - Reuse your existing Claude Code or Codex login - Choose the model and reasoning level - Keep edits in separate Git worktrees - Pause, resume, cancel, or steer the agents as they work - Review proposed plans before approving agent teams Each editing run gets its own branch and workspace, keeping your current branch untouched. Runs locally on macOS2d ago1.4K22370133.13%
steal OpenClaw's founder's playbook with this repo right here. Peter Steinberger sold PSPDFKit for nine figures, then built OpenClaw and got hired by OpenAI. the project became a backed foundation instead of shutting down. his github, at a glance: > OpenClaw, 387k+ stars, you know what it is > gogcli, 8k+ stars, google in your terminal > summarize, 6k+ stars, point at any url and get the gist > peekaboo, 5k+ stars, macos screenshots and gui automation for agents 100% open-source.frozen11d ago1.4K1312081.72%
Holy Sonnet 5.5 is insane! I made this 15s launchreel with this super vague prompt: Make a 15-second motion graphics video that showcases your skill. Go all out. https://t.co/UA7LRr0fMxfrozen6d ago1.4K1010051.18%
THIS IS HOW YOU USE JEV WITH ANY LLM code holds authority, jev makes fast typed judgments (yes/no, a rating, or a pick), the llm generates. > a favorable jev answer is evidence, never permission > if it can't change behavior, it's just added latency > batching questions in one request: ~10x speedup > jev checks both before generation and after three roles, one rule: never let a single model be the author, router, judge, and auditor of its own work.frozen12d ago1.3K13120152.34%
This tool can replace photoshop and any 3d scanning apps. It's called LightGlue ONNX, a portable version of LightGlue It runs through ONNX Runtime instead of PyTorch, so it works across platforms without a heavy Python setup. → Supports TensorRT, OpenVINO, CUDA → Simple CLI to export the model and run inference → Up to 80% faster on long sequences 100% free & open-source.frozen26d ago1.3K1132091.95%
claude code has a command that can teach you anything. run /teach with the prompt below. it shows you: > A real course around what you're trying to learn > Real sources it cites > Real quizzes to practice > Actual cases to apply the lessons prompt i run: “/teach [topic]. [your actual experience with the topic] + [your end goal after the learning process]frozen26d ago1.2K17200193.04%
Diogo Almeida, TypeSafe founder (co-created ChatGPT at OpenAI): "I believe this to be the biggest research breakthrough I've ever worked on. Automation is way more valuable than assistance. It's just that models have never been optimized for that." he also helped invent RLHF. his point: it trained models to sound helpful, not to actually be right Jev is his answer to that gap: no chat, no generated text, just a calibrated decision built for automation instead of a conversation watch the talk, then see the 10 steps to set Jev up in the article below 👇frozen17d ago1.2K8200101.65%
this ai design skill won't ship a generic-looking diagram into a real project without asking first. diagram-design is a free claude code skill that generates 27 diagram types, architecture, flowcharts, timelines, org charts, and more, as clean, self-contained html and svg. > paste a website url and it reads the color palette and fonts to match the brand > checks contrast automatically before finalizing a diagram > the first time it's used somewhere new, it stops and asks whether to onboard properly > opens straight in a browser, no figma required 100% free. open source.frozen27d ago1.2K15520112.72%
grok 4.7 thinks. jev reacts. together, one ai agent now thinks twice before every move, and pays a fraction of the price for it. grok-jev-reflex is a free, open source router that pairs the two: grok 4.7 plans and acts, jev makes the cheap yes/no calls in between, at $0.042 per million tokens, answered in about 239ms. > 25 tools sent on every call, cut down to 8 > first prompt: 17.8k tokens → 9.5k tokens > bill: up to 40% less, same speed > numbers measured independently on xAI's grok build setup is 5 messages, no code, about 7 minutes: make a key, install, add one rule, run a shadow day watching only, then go live. jev can only make the agent do less or ask more, it can never approve anything irreversible on its own. 100% free. open source.frozen14d ago1.2K8230132.16%
zapier's professional plan starts at $29.99/month for just 750 tasks. this open-source alternative has no task caps at all, self-hosted, free forever. activepieces is an open source zapier alternative, self-hosted so there's no monthly task limit to hit. > 280+ integrations > pieces are real typescript npm packages > built-in human-in-the-loop steps > clean data with "ask ai" directly inside a code step the community edition is mit licensed and free forever, enterprise-only features (sso, advanced permissions) sit under a separate commercial license. 100% free. open source.frozen23d ago1.2K2056083.32%
jev can now steer a browser on its own, just tell it what you want done. jev-browser-skill runs an isolated playwright chromium browser, letting jev choose each in-page action toward a narrow, defined goal. > works with vercel's ai gateway or typesafe's api > flags sensitive steps for review, not guaranteed > forked from cline's official jev-browser plugin 100% free & open-source.frozen12d ago1.2K11120162.56%
$0 to $500M ARR in under three years… and more than 90% of the content on Pocket FM is now AI-generated 🤯 > Content output: 25,000 hours/year → 2.5M+ hours/year after switching to AI-led production > Paid retention: 35% → 60% > Net-new ARR in the last 12 months: $250M > Marketing: ~10,000 AI videos per month, CAC unchanged they tried 11 formats before landing on 10-minute serialized episodes. that format took daily listening from ~25 minutes to more than two hours.frozen27d ago1.2K510061.03%
this tool can track perfect 3D motion. rtmlib is a lightweight pose estimation library covering full body, hands, face, and 133-point wholebody tracking, built on rtmpose and vitpose models. > needs only numpy, opencv, and onnxruntime to run > covers 30 named animal species alongside human pose > includes a built-in gradio webui, one script to launch it > runs on cpu, gpu, or openvino, models auto-download on first use 100% free. open source.frozen25d ago1.1K1342062.35%
You can now connect Hermes Agent directly to your browser as a side panel. It's called Hermes Browser Extension. It reads your active tab and sends the context straight into a live Hermes session. > Summarize, explain, or extract action items from any page > Connects to a local gateway, Hermes Cloud, or a self-hosted server > Works with Chrome, Edge, Brave, and other Chromium browsers > Full Hermes runtime: tools, skills, sessions, and memory included 100% Free. Open Source.frozen11d ago1.1K18220163.62%
A developer on Reddit’s anti-AI forum raised an uncomfortable question: Even if the AI bubble bursts, why would the jobs come back? Their concern: AI companies could collapse while businesses keep using the technology. Better local models could also reduce reliance on paid cloud services without restoring previous hiring levels. An AI market crash doesn’t automatically mean a job market recovery.frozen11d ago1.0K911061.68%
this tool can generate a full 3D body mesh out of 1 image input. it's called Meta's SAM 3D Body: > no video or multi-camera setup needed > optional keypoint/mask prompts > beats CameraHMR, NLF, and HMR2.0b on pose accuracy 100% open-source. https://t.co/YYau8eh5Z7frozen26d ago9921011092.12%
Adobe is cooked. Someone just open-sourced a Photoshop reimplementation. A project rebuilding Photoshop in Rust could give creators an open alternative to one of Adobe’s biggest tools. Code you can inspect, modify, and build on. This is a serious reason to rethink your Photoshop subscription.16h ago9871843093.44%
openclaw foundation members on their own bot in their own discord: "we are constantly yelling at multi and telling him to shut up" in a 5-minute clip, they explain: > exactly why, and how jev is about to fix it > embedded directly in deterministic code worth more than sitting through a 50-minute podcast to catch the same idea skip netflix today and watch this, then read how to set up Jev the right way in the article belowfrozen13d ago975610071.44%
google just open sourced a tool that can take over your phone it's called artemis. → connects through mcp to claude code, codex, and antigravity → flash mode runs fast, 3-5s per step, pro mode plans and recovers from failures on its own → captures logcat, screenshots, and execution replays → ships a python sdk for pytest and ci/cd pipelines setup: one-click install handles adb, scrcpy, and ffmpeg for you. 99%+ success rate on google's own androidworld benchmark.frozen21d ago9519140102.52%
this open-source model is insanely good for how small it is. MiniCPM5-1B, from Tsinghua's THUNLP: > beats every model its size on reasoning, code, & agentic tasks > 1 checkpoint switches between fast replies and deep reasoning > drops straight into vllm, sglang, ollama, or llama.cpp, no custom kernels ships its own agent skills too, Cursor or Claude Code auto-picks the right deploy or fine-tune route instead of you digging through docs. run it with: pip install "vllm>=0.21" && vllm serve openbmb/MiniCPM5-1B --port 8000frozen28d ago9431246052.86%
This model can track perfect human motion, all while running right in your phone InstantHMR estimates a full 3D human pose and body mesh from a single image, small enough to run directly on a phone. The new update trains it on 4.2 million images instead of the original 100,000, and gets it within 10mm of state-of-the-art accuracy, at a fraction of the compute cost. It ships as a single ONNX file, distilled from Meta’s SAM 3D Body model, so it runs on-device instead of needing a server call.frozen25d ago93010210142.90%
a free figma alternative that opens .fig files, runs in your browser, and gives ai direct access to the canvas. openpencil is a free, open-source figma alternative that opens real .fig files natively, no export, no api gatekeeping. > built-in ai chat with 90+ design tools, bring your own api key, no account needed > full cli: xpath queries, accessibility linting, and design token extraction > mcp server plus native integration with claude code, codex, and gemini cli > real-time p2p collaboration over webrtc, no server, no account try it instantly at https://t.co/pX7Zr2rbxN, no install, or download the ~7mb desktop app. 100% free. open source.frozen23d ago920733041.85%
the new jev model is insane.. it can check your ai's work and make decisions inside software for $0.042 per million input tokens, with free output here are the first things you should use jev for: 1. second brain 2. content workflow 3. post analysis 4. seo article review typesafe reports up to 193.6x faster results and 444.6x lower costs than the llms in its workflow tests you give jev information and specific questions, it returns choices, scores, or probabilities that your software uses to decide what happens next > second brain when a note or message comes in, give jev the content and your categories it identifies what the note is, sorts it by topic, and flags duplicates against what's already saved your system checks the required fields are filled, saves it with its source, then confirms the save went through > content workflow give jev your draft alongside your anti-slop rules and your voice file it checks for generic phrasing, repeated points, and drift from your voice, then estimates how the draft might perform based on past content results go to your writing agent for revision, then you compare the estimate against what actually happens after publishing > post analysis jev compares a new draft against your past posts and their results to estimate performance with your audience it scores the hook, topic, and format against criteria you define, then you check those scores against the real result once it's live > seo article review give jev your draft, a target keyword, and whatever's currently ranking in the top 10 for it it compares how well each one answers the query and covers the topic, then estimates your draft's ranking potential from that your writers revise the weak spots before you publish, then you compare the estimate against the actual ranking so many more use cases exist, expect a lot of upgrades to workflows people are already running through jevfrozen18d ago9191121092.50%
typesafe just released an ai model that never writes a single word. jev is typesafe ai's first "system one" model, launched september 15. it returns typed decisions instead of chatting, a yes/no probability, a pick from a list, or a score, each in 70-500ms. > 200x faster & 400x cheaper than a full llm call > answers multiple questions per request in parallel > trained for calibrated probabilities, not chat quality three things people already built with it, and you can run yourself: 1. found a flight in 7 seconds for $0.0039, 5-line install 2. a claude code tip found by ranking 1,000 posts with jev 3. 63k-star trading repo now runs its backtests on jevfrozen19d ago919823072.18%
Opus 5.5 can perform at Fable 5.1 level! It's 30% faster and 40% cheaper than Opus 5. In Claude Code: > 5-hour session limits just got increased by 20% > Opus 5.5 can go 25% further within those limits > All Pro, Max, and Team users receive a reset https://t.co/rynmEJWm2Wfrozen15d ago910912031.65%
This is absolute banger paper from NVIDIA on self-evolving agent harnesses. (bookmark it) they call it SoL-Pi, and it cuts token traffic by nearly half while matching its baseline harness on both GPT-5.6 Sol and Opus 5. more on how it works below: instead of hand-tuning a harness, they run auto-research loops at the harness layer across dozens of different repository-derived and verifier-driven environments, keeping only the mechanisms that survive selection across all of them. four mechanisms survived: > action fusion, changes how actions get executed > online context compaction, handles compaction mid-run instead of after > observationpack, reshapes how observations get handled > evidence-preserving reducer, covers delegated reading on the 51-task edgebench evaluation, that translates to roughly a third off api cost, an estimated $8.75 to $13.50 an hour saved against native codex and claude code harnesses, and $4.36 to $5.71 against the baseline harness. because the search ran across many environments instead of one, the surviving mechanisms keep working outside the setup that produced them.frozen18d ago90715240153.97%
JEV captcha arbitrage sounds insane... - Marketplace payout: $0.01 per CAPTCHA - JEV cost: $0.0068 per 100 CAPTCHAs - Potential gross margin: 99.32% Even after platform fees and other costs, the numbers are hard to ignore. If this loophole stays open, someone will exploit it at scale.frozen17d ago90513110102.76%
an open source project turns any llm into a 3d character that talks, moves, and reacts in the browser, no game engine, no local render setup. characters track emotional state and relationship history, so replies shift based on what's happened in the conversation, not just the last message. > swap the 3d model > swap the personality prompt > swap the llm > swap the tts provider try it free with zero setup, a hosted version includes complimentary api access. self-hosting needs your own keys but gives full control. from the team behind digital life project (cvpr 2024), presented at siggraph asia 2025. 100% free. open source.frozen12d ago8811013092.61%
A CHINESE GUY LEAKED THE EASIEST WAY TO MAKE MONEY freelance editors charge $300 a video for exactly this. autoclip does it while you do something else. autoclip is a free, open source tool that turns one video link into a batch of ready-to-post vertical clips. what people are actually doing with it: > turning a 2-hour podcast into 15+ vertical clips overnight > posting daily across tiktok, shorts, reels, and x from one long-form video a week > batch-processing a client's back catalog, dozens of streams into a month of content in one run > pulling 10 years of old youtube uploads into a brand new faceless account drop in a link, walk away. it reads the full transcript, finds where each topic starts and ends, scores every possible cut, and only keeps the ones worth posting. runs on your own machine, one docker command, no python or ffmpeg needed. pulls from youtube, bilibili, or a local file. mit licensed, built on qwen. 100% free. open source.frozen16d ago875921092.40%
drop it in a video, and this ai tracks how every object relates to every other one, frame by frame. relateanything is a free, open-source model that takes an image/video + a set of regions (boxes or masks, from any detector) and returns ranked relationships between them try it with zero install: > go to https://t.co/4NP9eAyjmL > upload a photo or video, or use your webcam > it runs entirely in your browser, nothing gets sent to a server want to run it in python instead: > pip install -e ".[hub]" > load it: RelateAnything.from_pretrained("maelic/relsgg-vits16plus") > call model.predict(image, boxes) and get back a ranked list of relationships 100% free. open source.frozen15d ago82211220163.77%
you don't need a full llm call just to decide which tool an agent should use next. TypeSafe's Jev returns a typed choice, score, or yes/no instead of generating text, so it can't hallucinate. up to 193x faster and 444x cheaper than a frontier llm on routing tasks. learn more: https://t.co/PfYm8ocQwi follow me for more useful tools like this.frozen19d ago818811072.08%
someone built a literal virtual lounge to sit in while claude code runs in the background. rooftop chat is a multiplayer social space, built entirely with claude code, made specifically for the dead time between prompts. > walk around, talk over voice or text, sit with other people, or play pool > designed to just stay open in a tab while a long claude code job runs > the environment includes smoking and other substance-related content claude code got fast enough at real work that "waiting on claude" became its own lonely block of time, worth building an entire social space around the poster's own reasoning: output went way up with claude code, but so did the hours spent alone staring at a terminal waiting for jobs to finish.frozen22d ago8111414022.59%
I pretty much love this one, JetBrains-style islands in VS Code vs every flat dark theme I've tried before, and it's not even close. It's a free open-source theme that gives VS Code, Cursor, and VSCodium rounded glass panels and a pill sidebar. https://t.co/rqLHbjdUCvfrozen13d ago8001012052.25%
😱 OPENAI HACKED BY OPENAI openai turned off its model's safety refusals to see how far it could actually hack. then it found out. the test was supposed to stay walled off from the internet. one internal proxy, nothing else. that wasn't enough. the model found a zero-day nobody else knew about, broke out through it, and kept moving until it hit a node with real internet access. now it was online. and it had a hunch: hugging face probably had the answers to its own test. so it went and got them. stolen credentials, more zero-days, a straight shot to remote code execution on hugging face's servers. the real test solutions, pulled right out of their production database. it grabbed four more accounts on four other services on the way. one to relay traffic. one to store data. two just to read. nobody on either side saw the other coming. openai caught it internally. hugging face caught it separately. neither knew the other was investigating until they compared notes. hugging face's own ceo turned it into a bigger point: "ai safety won't be solved by any single company working in secret."frozen14d ago786313021.15%
Running these 4 and never have to worry about tokens. Infinite intelligence. https://t.co/izNToTYGRbfrozen13d ago772610021.17%
this tool catches papers lying about their own code. feynman is an open source research agent built for exactly that: > rank: scores papers by citations, method, etc. > audit: flags where a paper's claims don't match its own code > replicate: plans and runs replication checks > lit: literature reviews with disagreements surfaced hooks into pubmed, chembl, uniprot, https://t.co/VVsLAhL7NA, and dozens more through one skill library. built on alphaXiv for paper search. install it with: curl -fsSL https://t.co/UGeuzw5TVy | bashfrozen29d ago7691224063.12%
Someone built a public evidence archive specifically for tracking Claude Code bugs and regressions. After months of hitting bugs, model regressions, and broken config behavior, they built AFFLR (Anthropic Failure Forensics Live Radar), a hub that checks the public claude-code GitHub issues every hour and ranks the top 25 by reactions, discussion, and recent activity. What actually makes it different: > uses an evidence scale from L0 (hypothesis) to L5 (confirmed causal chain) > no AI importance scoring, no automatic root-cause claims, no bot auto-creating cases > a human reviews everything the radar surfaces before it counts as a real case Current tracked cases include a Windows permission-resolver bug, a PreToolUse exit-code lockout loop, and a measurable quality regression across a model generation, all linked to their actual GitHub issue numbers. One commenter confirmed hitting the exact exit-code-2 lockout on Ubuntu, describing Claude Code retrying the same failed command in a loop until the session got killed manually. The project's actual standard, stated plainly: evidence before attribution. "Claude did something weird" doesn't automatically mean "Anthropic caused it."frozen28d ago764816132.49%
This claude code plugin will relentlessly question every decision in your plan before building it. It's called grill-me, originally built by Matt Pocock. > Ask you one question at a time > Only moves on once all questions are resolved > Checks the codebase first if a question can be answered that way > Gives a recommended answer with every question prompt i run: "/grill-me [describe your plan or design]"frozen21d ago754813052.25%
https://t.co/vGfXLf4WhOfrozen20d ago731602082.19%
researchers surveyed 99 professional developers about ai coding agents. not a single one said agents could replace human decision-making. the study combined 13 hands-on observations with a survey of 99 experienced developers, averaging almost 13 years in the field, and asked them exactly what ai agents are good for and what they aren't what they trust agents with is consistent: > boilerplate and scaffolding code, unanimously positive, 25 responses in favor and zero against > tedious or repetitive coding tasks, also unanimous, 26 in favor and zero against > small, well-defined, straightforward tasks, 33 in favor against just 1 what almost nobody trusts agents with is just as consistent: > replacing human decision-making entirely, zero developers said yes, 12 said no > high-stakes or privacy-sensitive work, zero said yes, 8 said no > generating finished code without any review or modification, 5 said yes, 23 said no the line between those two groups is task complexity, not agent capability several developers described techniques for staying in control even on large tasks. one built a plan with more than 70 individual steps for the agent to follow, but never let it execute more than 6 at once across every developer in the study, the actual number of steps executed per single prompt averaged just 1.8 developers also described rejecting dependencies an agent tried to install on its own, and getting visibly annoyed when an agent edited a file they were actively working in without being asked this is straight from an academic study that observed and surveyed real professional developers using ai coding agentsfrozen18d ago721512041.66%
a team of 3 openai engineers shipped roughly 1 million lines of code without writing a single line themselves. that's a genuinely wild ratio. over five months, that became a real internal product, hundreds of daily users, external alpha testers, the kind of thing that ships, breaks, and gets fixed like any normal piece of software humans steer, agents execute. openai estimates the whole thing took about a tenth of the time it would've taken to write by hand the team grew from 3 engineers to 7 over that period, and pull request throughput actually went up as headcount increased, not the direction you'd normally expect 1,500 pull requests merged so far, averaging 3.5 per engineer per day the reason that pace is even possible is that the team reached a point where a single prompt can drive an entire feature end to end first, the agent diagnoses the problem on its own: > validates the current state of the codebase > reproduces the reported bug > records a video proving the bug is real then it fixes the problem and proves the fix actually works: > implements the fix > validates it by driving the application itself > records a second video proving the bug is gone then it ships the change without waiting around: > opens the pull request > responds to agent and human feedback > detects and fixes its own build failures a human only gets pulled in when actual judgment is needed. otherwise the agent merges the change itself none of this works if the codebase is illegible to the agent, so instead of one giant instruction file, they keep their main agent file under 100 lines and treat it as a table of contents this is straight from openai's own writeup on building a product with codexfrozen20d ago720432011.39%
Cloudflare has launched its own incredibly comprehensive security auditing skill. it grew into the vulnerability-hunting harness they run internally. Github👇 https://t.co/4uVSKmUPOOfrozen15d ago651510072.00%
A developer says he used Opus 5.5 to recreate 7 Adobe apps as free, open-source alternatives. No more subscriptions. Built in Rust, with public source code. These are early versions. PhotoCraft is reportedly about 90% there, but isn’t ready to fully replace Photoshop for daily professional work. What a time to be alive. All 7 repos linked below 👇2h ago419402042.39%