Wiredframe Radar logo

Category intelligence

Social Media Briefing — February 21, 2026

509 current items analyzed and ranked.

Executive synthesis

Social Media Summary

A landmark day for AI infrastructure and security discourse. Andrej Karpathy defined "Claws" as a new conceptual layer atop LLM agents—handling orchestration, scheduling, and tool calls—sparking massive community engagement (754K views). Meanwhile, Anthropic dominated security conversations as Claude Opus 4.6 found 500+ open-source vulnerabilities, with the team also unveiling safeguard probes for cyber misuse detection.

  • An Anthropic engineer revealed that prompt caching must be the foundational design constraint for AI agents, reshaping how developers think about agent architecture (676K views)
  • NVIDIA's Jim Fan announced DreamDojo, an open-source robotics world model pre-trained on 44K hours of human video, declaring 2026 the year of Physical AI
  • Sam Altman met India's PM Modi, revealing India is OpenAI's fastest-growing market for Codex globally (3.2M views)
  • The llama.cpp team joining Hugging Face marked a major consolidation in the open-source inference ecosystem
  • Demis Hassabis showcased Gemini 3.1 Pro capabilities, while Greg Brockman framed inference compute as the new driver of software productivity
  • Karpathy further envisioned a post-app paradigm where discrete software dissolves into ephemeral, AI-assembled code paths

Key Themes

Claws: New AI Infrastructure Layer · 5Prompt Caching as Core Agent Design Principle · 10Claude Opus 4.6 Vulnerability Discovery · 6World Models and Physical AI · 6Post-App Software Paradigm · 5llama.cpp Joins Hugging Face · 1OpenAI Global Growth and Codex Adoption · 2Gemini 3.1 Pro Launch and Reception · 7Claude/Anthropic Product Launches · 3Coding Commoditization & AI Workforce Shift · 6

Primary evidence

Top Ranked Signals

Social Twitter Feb 20

Bought a new Mac mini to properly tinker with claws over the weekend. The apple store person told me...

By @karpathy

95 score
AI Analysis

Following yesterday's News coverage, Karpathy provides an extensive analysis of 'Claws' - a new layer of the AI stack on top of LLM agents handling orchestration, scheduling, context, tool calls, and persistence. He discusses security concerns with OpenClaw (400K lines of vibe code being actively attacked), praises NanoClaw's ~4000-line auditable codebase and novel 'skills-as-configuration' approach, and describes the aesthetic appeal of local AI setups on physical devices like Mac Minis.

Bought a new Mac mini to properly tinker with claws over the weekend. The apple store person told me they are selling like hotcakes and everyone is confused :) I'm definitely a bit sus'd to run OpenClaw specifically - giving my private data/keys to 400K lines of vibe coded monster that is being actively attacked at scale is not very appealing at all. Already seeing reports of exposed instances, RCE vulnerabilities, supply chain poisoning, malicious or compromised skills in the registry, it feel
AI Infrastructure EvolutionClaws EcosystemAI Security ConcernsLocal AIVibe Coding Risks
Social Twitter Feb 20

one of the biggest realizations I've had working on Claude Code is that you fundamentally have to de...

By @trq212

95 score
AI Analysis

Continuing our coverage from yesterday's Social post, trq212 (Anthropic) shares major insight: agents must be designed for prompt caching first, as nearly every feature touches caching. Links to detailed article based on months of Claude Code learnings

one of the biggest realizations I've had working on Claude Code is that you fundamentally have to design agents for prompt caching first, almost every feature touches on it somehow I wrote this in a day but it's the culmination of months of learnings, hope you enjoy it
Prompt CachingAgent ArchitectureClaude CodeAI EngineeringSystem Design
Social Twitter Feb 20

Announcing DreamDojo: our open-source, interactive world model that takes robot motor controls and g...

By @DrJimFan

92 score
AI Analysis

Jim Fan announces DreamDojo, NVIDIA's open-source interactive world model for robotics. Pre-trained on 44K hours of human egocentric video, it uses latent actions to make videos 'robot-readable,' supports real-time inference at 10 FPS for live teleoperation, policy evaluation, and model-based planning (+17% real-world success). All weights, code, and datasets are open-sourced.

Announcing DreamDojo: our open-source, interactive world model that takes robot motor controls and generates the future in pixels. No engine, no meshes, no hand-authored dynamics. It's Simulation 2.0. Time for robotics to take the bitter lesson pill. Real-world robot learning is bottlenecked by time, wear, safety, and resets. If we want Physical AI to move at pretraining speed, we need a simulator that adapts to pretraining scale with as little human engineering as possible. Our key insights:
RoboticsWorld ModelsPhysical AIOpen SourceNVIDIA ResearchSimulation
Social Twitter Feb 20

Opus4.6 found 500+ vulnerabilities in open-source code and we've begun reporting them and contributi...

By @trq212

92 score
AI Analysis

Claude Opus 4.6 found 500+ vulnerabilities in open-source code; Anthropic has begun reporting them and contributing patches

Opus4.6 found 500+ vulnerabilities in open-source code and we've begun reporting them and contributing patches quick excerpts from some of them 🧵
AI Vulnerability DiscoveryClaude Opus 4.6Open Source SecurityAI SafetyResponsible Disclosure
Social Twitter Feb 20

Great meeting with PM @narendramodi today to talk about the incredible energy around AI in India. I...

By @sama

90 score
AI Analysis

Following yesterday's News coverage of the Delhi AI summit, Sam Altman met with Indian PM Modi to discuss AI in India, revealing that India is OpenAI's fastest growing market for Codex globally, with weekly users up 4x in the past 2 weeks.

Great meeting with PM @narendramodi today to talk about the incredible energy around AI in India. India is our fastest growing market for codex globally, up 4x in weekly users in the past 2 weeks alone. 🇮🇳! t.co/MRbw0UkotJ
AI Global ExpansionOpenAI Business GrowthAI GeopoliticsCodex Adoption
Social Twitter Feb 20

@kepano I think this group of reactions are still fundamentally rooted in a scarcity mindset of soft...

By @karpathy

82 score
AI Analysis

Continuing yesterday's Social thread, Karpathy envisions a future where software becomes so cheap and abundant that discrete 'apps' no longer make sense - code paths assemble for hyper-specific purposes and get deleted after single execution. He describes it as moving from 'castles of bricks of code' to 'boiling soup of code.'

@kepano I think this group of reactions are still fundamentally rooted in a scarcity mindset of software. 2 years ago AI was botching autocomplete, today it is almost one shotting browsers and C compilers. Where is it in 2 more? 10? 20? Software so insanely cheap and abundant that discrete “apps” make no sense in today’s sense. It’s just code paths that assemble for a hyper specific purpose, just to get deleted after a single execution. You don’t need to know anything or exercise any creative di
Future of SoftwareAI-Generated CodeSoftware AbundanceParadigm Shift
Social Twitter Feb 20

This is incredible btw - using Gemini 3.1 as a city builder. I used to dream about this when painsta...

By @demishassabis

80 score
AI Analysis

Continuing yesterday's Social coverage of the Gemini 3.1 Pro launch, Demis Hassabis (Google DeepMind CEO) highlights someone using Gemini 3.1 as a city builder, comparing it to his own experience painstakingly building virtual cities for simulation games like Republic.

This is incredible btw - using Gemini 3.1 as a city builder. I used to dream about this when painstakingly making virtual cities for simulation games like Republic.
Gemini 3.1 ProAI Creative ApplicationsGenerative AI
Social Twitter Feb 20

We've been working on this for a while -- it's impressive (and scary) to see the kinds of security i...

By @bcherny

78 score
AI Analysis

Building on yesterday's Reddit discussion of Boris Cherny's work, Anthropic's bcherny announces a security scanning feature they've been working on, describing it as impressive and scary what security issues it identified. Rolling out as research preview for Team and Enterprise customers.

We've been working on this for a while -- it's impressive (and scary) to see the kinds of security issues it has identified. Rolling out slowly, starting as a research preview for Team and Enterprise customers.
AI securityAnthropicClaudeproduct launch
Social Twitter Feb 20

All those products where building an "AI agent" meant defining a series of basic prompts linked toge...

By @emollick

75 score
AI Analysis

Emollick argues that AI agent products from 2025 that relied on deterministic flowcharts of linked prompts with separate RAG inputs are looking dated, implying the field has moved well beyond this approach.

All those products where building an "AI agent" meant defining a series of basic prompts linked together deterministically through a flowchart with separate RAG inputs are looking pretty dated right about now (yes, that is basically every agent product released in 2025)
AI AgentsProduct ObsolescenceAgent Architecture Evolution