Gera

@andyasprou open source?

OpenClaw X Linear on someone on slack can say "create a daily ticket to audit my LinkedIn DMs, draft response tickets I can review, and let Alfred send the approved ones" you get a paperclip style orchestration but execution, delegation and status tracking stays within Linear

AI Primer - Jan 20 LTXStudio turns a soundtrack into the edit timeline—20s videos with native lip sync. - ElevenLabs gets it exclusively for 7 days - Caps: 100MB audio, 20s max - Export: 1080p/25fps 🧵

🧩 2026-01-20 - AI Dev Stack Roundup full report: ai-primer.com/en/engineer/re… 🧩 Claude Code lands in VS Code (GA) — editor-native agent workflows • Claude Code’s VS Code extension hits GA with @-mentions and slash commands • Claude Code in VS Code can answer using your open file and selected lines • Claude Code posts a full VS Code setup guide for editor workflows • Claude Code VS Code extension also runs inside Cursor’s IDE • Anthropic asks for feature requests for the Claude Code VS Code extension 🧰 Claude Code CLI: 2.1.14 behavior changes, stability fixes, and power-user features • Claude Code CLI 2.1.14 fixes premature context blocking and crashy subagents • Claude Code CLI 2.1.14 adds bash-history autocomplete and plugin pinning • Claude Code users report formatting bugs, MCP tool list drops, and high CPU • Claude Code changes shell semantics: bash state no longer persists between calls • Claude Code drops ExitPlanMode allowedPrompts guidance in 2.1.14 • Claude Code web/desktop shows early “Export” work for zipping a repo • Claude Code UI shows “Update memory” to write back to agent docs • Claude Code now prefers gh CLI for GitHub URLs over HTML fetching • Claude Code power users surface /fork and /resume for session branching 🧑‍💻 OpenAI Codex in practice: planning friction, multi-agent hints, and enterprise workflows • Cisco says Codex plan docs improved reviews; cites 20% build gains and 10–15× defect throughput • Codex CLI surfaces a “dangerously bypass approvals and sandbox” mode • Codex plan mode ergonomics get criticized: “asked me to give it paths” • Codex Monitor pitches a fast-moving UI for Codex CLI with subagents and worktrees • Codex docs text shifts from “minimum” to “optimal” workers in multi-agent workflow • User report: Codex 5.2 is slower but better at C++ memory bug fixing than Opus 4.5 • OpenRouter shares a Codex usage-tracking page for top model variants • Codex “optimal fix” conversations show up in PR commentary, with stylistic tells • Users report model-picker mismatch: “Extended Thinking” selected, no thinking behavior • Codex is being used to remove “Opus-style” artifacts from PRs 🧭 Cursor & IDE harness design: static vs dynamic context, agent review UX, and team settings • Cursor explains why it moved from static to dynamic context discovery • Cursor adds a configurable Agent Review step (Quick vs Deep) • Cursor usage tip: start in Plan Mode, then let it search without heavy @-tagging 🧱 Skills & plugin ecosystems: install flows, versioning, and auto-loading behavior • OpenSkills 2.0 teases version locking and agent auto-detection • Vercel’s Skills.sh proposes a shared install path for agent skills • Skills are easier to add; determinism remains the trade-off • Skills auto-loading is still a trust gap for builders • Deepagents treats agents as folders you can share • Claudeception captures Claude Code learnings as skills • Gemini-in-Chrome shows “Skills management” UI strings 🧪 Workflow patterns for shipping with agents: specs, slop control, and compounding loops • De-slopping a coding agent by forcing module boundaries and testability • Enterprises are redoing AI “best practices” for long-running agents • CLI-first automation revival: small bash loops replace lots of process • PRD refinement: specify module depth and required test coverage before agent work • UI prototyping pattern: iterate on taste with throwaway routes • Why you still have to read code: slop compounds across agent runs • Jevons Paradox framing: cheaper code expands what teams choose to build • A 3-day agent-coding weekend replaces 1–2 years of solo work (anecdote) • Mindset pattern: agentic coding rewards comfort with failure and iteration • Setup discipline: compounding workflows beat perfect agent configs • Compaction workaround: write a “remember” file before context truncation • AI prose detection fatigue shows up in engineering review culture 🕹️ Agent ops & swarms: remote execution, task backlogs, and trace analytics • LangSmith adds an “Insights Agent” to analyze agent traces at scale • Decentralized swarm control beats a “mastermind” agent for long runs • Beads backlog telemetry shows 45 projects with 4,721 open tasks • “Everything is a Ralph loop” pushes loop-first agent ops as the default • “Write a remember file before compaction” pattern for long agent sessions • PredictionArena uses Kalshi P&L curves as an agent performance leaderboard • W&B WeaveHacks returns with a “self-improving agents” theme 🧠 Model watch: new checkpoints, open models, and near-term release signals • GPT-5.3 becomes the next named OpenAI update as Altman solicits feedback • DeepSeek “MODEL1” breadcrumb appears in FlashMLA KV-cache kernel code • Gemini 3 Pro resurfaces in AI Studio A/B tests, hinting at GA timing • LightOn releases LightOnOCR-2-1B OCR family under A2.0 with speed claims • OpenBMB’s AgentCPM-Report claims an open 8B deep-research agent workflow • Gemini expands language coverage to 70+ with 23 newly added languages • Gemini adds an “Answer now” fast-path powered by Gemini 3.0 Flash • LiquidAI’s LFM2.5-1.2B-Thinking lands as an on-device reasoning option • GPT-5.3 “Garlic” codename circulates with a “new pre-training” narrative 🔌 Connectors & orchestration surfaces (MCP, workflow nodes, and in-product actions) • Claude adds opt-in health data connectors across iOS and Android sources • Google ships a Stitch MCP server for live UI generation and IDE code fetch • OpenAI Atlas adds site-specific actions, starting with YouTube timestamps • Skills vs MCP debate narrows to determinism and “no code exec” enterprise reality • Firecrawl’s /agent node lands on n8n cloud for workflow-native web research 📏 Benchmarks & leaderboards: community eval scale and new scoreboards • Text Arena crosses 5 million votes, scaling human preference evaluation • GLM-Image reaches #8 among open models on LMArena Image • BabyVision highlights a large gap between Gemini 3 Pro Preview and adults • PredictionArena tracks models via P&L on Kalshi weather markets • Gemini 3 Flash briefly appears on LM Arena, then disappears ⚙️ Inference & self-hosting: runtimes, determinism knobs, and performance engineering • vLLM v0.14.0 flips async scheduling on by default and adds a gRPC server • SGLang-Diffusion claims Cache-DiT speedups and adds LoRA serving APIs • Ollama ships experimental local image generation on macOS • vLLM adds a practical knob for batch-size deterministic offline inference • MoE scaling field notes: Nsight-guided DeepEP tuning for intranode bottlenecks • vLLM adds day-0 support for Step3-VL-10B with reasoning/tool-call wiring 📄 Document AI: OCR quality, doc-heavy agents, and ‘file handling’ gaps • CB Insights frames document processing as the proving ground for enterprise agents • A Gemini Build prototype turns PDFs into fillable forms with bounding boxes • Gemini’s file-handoff limitations are still a blocker for “do the work” workflows • OlmOCR-bench is turning green; the next bottleneck is harder real-PDF evals 🏗️ Compute & infrastructure: power, clusters, and ‘pay our way’ commitments • OpenAI’s Stargate Community plan commits to covering energy costs and funding grid upgrades • Anthropic is rumored to self-deploy a TPUv7 fleet with nonstandard racks and optical switching • OpenAI Podcast frames compute as the binding constraint for AI adoption • Amodei argues against allowing high-end Nvidia GPU sales to China • OpenAI and the Gates Foundation launch $50M Horizon 1000 for primary healthcare delivery • Lisa Su projects another 100× compute surge over 4–5 years 🎬 Generative media in production: influencer factories, audio→video, and deterministic editing • Higgsfield launches AI Influencer Studio for 30s full-motion HD avatars • LTX Studio ships Audio-to-Video focused on dialogue and lip-sync • Black Forest Labs makes FLUX.2 [klein] free via API for 24 hours • Bria FIBO Image Edit lands in ComfyUI with JSON prompts and masking • ElevenLabs adds LTXStudio Audio-to-Video with a 7-day exclusive window • PixVerse R1 pitches a real-time video “world model” with continuous streaming • Gemini shares Nano Banana Pro pet adoption headshot workflow and prompt • Invideo Vision: Angles generates 9 camera perspectives from one frame • AI film workflow: Nano Banana character design plus Veo 3.1 “Ingredients” for interviews updates below ⬇️

AI Primer - Dec 26 MiniMax drops M2.1: open weights, 230B MoE model, 74% SWE-bench, runs 40–47 tok/s on Mac Studio. - SGLang/vLLM offer day-0 tool-calling support - Claude Code: 200 PRs/month, no IDE - Nvidia–Groq: inference racks go hybrid 🧵

AI Primer - Jan 2 Claude Code’s latest workflow runs 10+ parallel agents, racking up 3,000 tests in days—Boris Cherny cites up to 3× better shipped code quality. - Plan mode with verified loops now standard - RLMs stay stable past 10M tokens - Nvidia faces 2M H200 China orders 🧵

AI Primer - Jan 5 30,000 Atlas robots/year: Boston Dynamics wires Gemini Robotics into its fully electric factory bots—50 kg lift, 56 DoF, Nvidia onboard. - Rubin NVL72 trains 10T models with 4× fewer GPUs at 1/10th token cost - Anthropic orders ~1M TPUv7 chips - Nemotron RL hits 71–83% on core benchmarks 🧵

AI Primer - Jan 7 Stanford finds Claude 3.7 Sonnet regurgitates books with 95.8% near-verbatim recall. That’s not a corner case. - ChatGPT Health launches: 260 doctors, 600k review cycle - LTX-2: full open audio-video, runs 4K on RTX cards - Claude Code 2.1 hot-reloads Skills, auto agent forking 🧵

@higgsfield_ai Awesome

KLING 2.6 MOTION CONTROL is LIVE on Higgsfield. TRY FREE. DAY-0 access. As always. Fast & complex movements finally work in AI video. Full-body sync, expression mapping, 30-second generations. Everyone else is behind. For 9 hours: retweet & reply & like & follow for 249 credits

@NahFlo2n blueprint

$100k months from ai influencers that don’t exist i generate the face, voice, script, and videos in under 2 minutes and the pages run themselves the results speak for themselves: @ healthwithnancy → $100k/month @ malehealthclub → $450k/year @ mens_healthcoded → $3k/post this is one of the easiest scalable systems right now rt + comment “blueprint” and i’ll dm the setup (follow for dm)

@recap_david @therecapai SYSTEM

This AI system created a viral Instagram reel with 4.1M views... completely on autopilot (linked below). These millions of views led to 10,000+ newsletter subscribers for my AI newsletter @therecapai While most creators burn out from the daily content grind, I built an n8n automation that transforms raw AI news into scroll-stopping short-form video. Here's what happens behind the scenes: → Data ingestion agents scrape Twitter, Reddit, and top AI blogs 24/7 → AI analyzes hundreds of stories and picks the top 3-5 most engaging ones → Advanced prompts craft multiple hook variations and full video scripts → HeyGen generates AI clone of me speaking script → Video editor in india ads B roll → Everything gets delivered to Slack for quick review and approval This system has brought my Instagram account to 16,000+ followers in just a few months. That's the power of consistent, data-driven content creation. I just released a deep-dive tutorial showing exactly how this works, including the prompts I literally spent hours on making it all possible. Want me to send it to you? Like & RT this post Follow so I can DM you Comment 'SYSTEM' and I'll share the complete workflow template + all the prompts. Your next viral video is literally just one automation away.

@mikefutia ADS

This Shopify brand is running 900+ AI-generated ads on Facebook 🤯 And they're publishing 20 new AI UGC ads every single day. I pulled their entire Facebook Ad Library and reverse-engineered their entire strategy. Perfect for DTC brands & media buyers who want to scale creative without hiring an army of UGC creators. Here's what makes this wild: Most brands struggle to produce 5-10 new creatives per month. This brand is pumping out 20 AI-generated ads PER DAY. They're using AI video tools to create testimonials, unboxings, demos & lifestyle shots at scale. And it's working. What I found in their ad account: → 900+ total ads live in their Facebook Ad Library → Consistent AI-generated videos across all creatives → Multiple hooks & angles tested daily → Same product, dozens of different scenarios → Clear patterns in what's getting spend I recorded a full Loom breakdown showing: → Which AI tools they're using → The exact creative patterns they're scaling → Hook structures they're testing → How they're maintaining brand consistency → How you can replicate it for your brand Want the full behind-the-scenes Loom breakdown? > Comment "ADS" > Like this post And I'll send over the Loom video (must be following so I can DM)

@zulkarnaimx Youtube

I don't understand why people don't just lock in on YouTube My faceless YouTube channels make me $120,000+/mo (long form only) There is 𝗻𝗼 𝗹𝘂𝗰𝗸 involved. Let me send you a free course on exactly how to do it. Just Retweet and Comment "YouTube" (must follow)

@JulianGoldieSEO Yes

FREE N8N Template That Builds $3,000+ Monthly Faceless AI Channel! Grab this FREE template that starts your faceless video empire and attracts high-paying clients automatically. What's Included in Your FREE Package: ➤ 4-stage workflow system - Idea generation to auto-posting covered ➤ Google Sheets tracking template - Monitor all your automated content ➤ VEO 3 API setup guide - Professional video quality every time ➤ Multi-platform posting templates - TikTok, Instagram, Facebook ready LIKE + RETWEET + REPLY “YES” & I’ll send you the FULL workflow + setup FREE!

@Lupin_Ai_Coder IG

My AI Instagram girl is bringing in $30k a month Just reached 300K, most of her reels go viral Nobody notices she is AI If you want to try creating an AI influencer yourself, like, retweet, and comment “IG.” Follow me, and I will send you a guide with instructions.

@FynCas Nano

Nano Banana + MakeUGC + Veo3 = AI content Factory This agent pumps out hundreds of ads daily — fully automated. - No $300 creators - No $10K/month agency fees - No products Just drop in a competitor’s ad → pick an avatar → let AI rebuild it in seconds. Want the agent + full workflow? Comment “Nano” and I'll send it for FREE! (must be following)

BREAKING: Gero Signs Potential Multi-Billion-$  Deal to Fuel Our Mission to Stop Aging  -- A few years ago, my son, aged five at the time, asked what happened to the dinosaurs. After learning their sad story, he told me he wanted to fly to the Moon and defend Earth from falling comets. He began assembling his crew—quickly running out of family members and turning to his grandparents—but then stopped and asked: “Dad... Grandpa and Grandma are already old. It will be hard for them to shoot down comets. Can we make it so people don’t age?” At that tender age, he couldn’t yet fully grasp death, but he already understood aging—and its cruel way of stealing our strength and time. This story has played out billions of times in human history: a child realizes, for the first time, that life is finite. But maybe—just maybe—his generation will be the first to hear a different answer. Today, we took a step toward that answer. Together with Chugai Pharmaceutical (a member of Roche group), we announced a partnership to develop therapies that target the root causes of major age-related diseases. At Gero, we’ve spent the last 15 years building physics-informed AI models, trained on massive longitudinal datasets tracking human biology over years. These models revealed something critical: in long-lived mammals like humans, aging is mostly an entropic process—slow, irreversible damage that accumulates over time. But diseases? They are failures of individual systems—failures that occur faster in aging bodies but, crucially, are still actionable. Some of these failure points are shared across multiple diseases, and targeting these shared vulnerabilities could lead to therapies that fight multiple diseases at once. When we started our journey, there were no examples of success to follow. What we had was the belief that science works—and that complex phenomena like aging and chronic diseases could be understood through first principles borrowed from other sciences like physics, and with the help of emerging AI and machine learning tools. We also had no clear idea how to fund such an enterprise. Our inspiration came from nature itself—from mammals like naked mole rats and bats, which show little to no signs of aging over long lifespans. Things are moving fast. The recent success of GLP-1 drugs (Ozempic, Wegovy) has already shown what’s possible: a single drug class improving metabolic health, heart disease, obesity, and more. What once seemed like science fiction is now reality. And yet, as I argued in my Lifespan's column, these are still just our first flying machines. The real challenge is to build the rockets—to address the fundamental physics of aging, not just its symptoms. Our models predict three levels of intervention against aging in humans  (as always, please do not forget to like, follow and repost before scrolling down for more explanations and links): 1/2

@DmitryLepisov AIDS virus

What do you see in this logo?

@kmartynov В Монте есть неплохой джаз клуб, с латиноамериканским послевкусием Thelonious Club. Сегодня работает. Там хорошо заказать стейк 🥩, обычно очень вкусный подают

Друзья, если кто-то из вас в Монтевидео, буду рад повидаться.

What a picture. Skin from Skunk Anansie receiving her OBE for services to music from King Charles at Buck Palace today.

🎉🌈