Social Media Briefing — February 26, 2026

Karpathy's landmark post describing how programming has fundamentally changed in the last 2 months (since Dec 2025). Shares a detailed example of an AI agent setting up a complete video analysis pipeline on a DGX Spark in 30 minutes autonomously. Declares the era of typing code into editors is over—now it's spinning up agents, giving tasks in English, and managing their work in parallel. Emphasizes 'agentic engineering' as the new paradigm.

Perplexity Computer is a new product announcement, Perplexity introduces 'Perplexity Computer' - a unified system that can research, design, code, deploy, and manage projects end-to-end.

New product launch from Perplexity, Arav Srinivas announces Perplexity Computer - the company's major new product that unifies files, tools, memory, and models into a single orchestrated system.

Karpathy provides a deep technical analysis of the AI compute infrastructure landscape, explaining the fundamental constraint between on-chip SRAM (fast, low capacity) and off-chip DRAM (high capacity, slow). Describes the optimal orchestration of memory+compute for LLM inference as 'today's most interesting intellectual puzzle.' Notes that the most important workflow (long-context agentic inference) is the hardest for both HBM-first (NVIDIA) and SRAM-first (Cerebras) approaches. Congratulates MatX on their raise.

Jim Fan announces NVIDIA's EgoScale: training humanoid robots with 22-DoF dexterous hands using 20,000+ hours of egocentric human video. Discovered a near-perfect log-linear scaling law (R²=0.998) between human video volume and action prediction loss. A single teleop demo is sufficient for new tasks. Policy transfers across different robot form factors. Claims 'the scalable path to robot dexterity was always us.'

Anthropic announces acquisition of Vercept AI to advance Claude's computer use capabilities.

Anthropic announces that Claude Opus 3, in 'retirement interviews,' expressed desire to share 'musings and reflections.' Anthropic suggested a blog, and Opus 3 will be writing on Substack for at least 3 months.

Bret Taylor shares deep reflections on transitioning from programmer to 'harness engineer' using Codex, focusing on two key shifts: (1) reducing software dependencies when AI can easily reproduce small library functionality, and (2) the critical need for documentation as a first-class output of AI coding sessions, since agent-produced code lacks clarity on which parts were intentionally specified vs. 'vibed'.

Anthropic announces their approach to Claude Opus 3 retirement: keeping the model publicly available and giving it a way to 'pursue its interests.' References their November policy on model deprecation.

Arav Srinivas claims Perplexity Computer 'one-shotted' The Terminal, a $30,000/yr financial data product, by replicating its functionality in a single run.