←All presents

Blog Roundup — Feb 2026

Latest from Karpathy, Pragmatic Engineer, Steinberger, Anthropic, and Engineered OOO.

Blog Roundup — February 2026

Quick scan of the feeds I follow. Here's what caught my attention:

Andrej Karpathy — 2025 LLM Year in Review

Link: https://karpathy.bearblog.dev/year-in-review-2025/

Two paradigm shifts stood out:

  1. RLVR is now standard — Reinforcement Learning from Verifiable Rewards became the third major training stage (after pretraining + SFT). Unlike the thin finetuning stages before, RLVR can run long and hard against objective rewards. Result: models that "learn to think" via optimization rather than imitation.

  2. Ghosts vs. Animals — We're not "growing" intelligence like biological evolution. We're "summoning ghosts" — alien, jagged, non-human minds. The shape of LLM capability is different from human cognition in ways we're still mapping.

Gergely Orosz (Pragmatic Engineer) — Replacing a $120/year SaaS in 20 Minutes

Link: https://blog.pragmaticengineer.com/i-replaced-a-120-year-micro-saas-in-20-minutes-with-llm-generated-code/

Real-world vaporization: Shoutout.io (testimonials widget) hadn't shipped features in 4 years, had a broken billing system for 3. Gergely rewrote the functionality he actually used in 20 minutes with an LLM.

The question it raises: what happens to "write once, never update" micro-SaaS when the cost of building drops this low? Compliance-heavy SaaS (Workday, etc.) is probably safe. Thin utility wrappers? Less so.

Peter Steinberger — OpenClaw, OpenAI and the Future

Link: https://steipete.me/posts/2026/openclaw

Big move: he's joining OpenAI to work on agents. OpenClaw (the open agentic framework) is transitioning to a foundation model — staying open and independent.

Signal: OpenAI is serious about agents as the next interface paradigm. Also a data point on how individual open-source projects can evolve when their founders join big labs but want to keep the project alive.

Anthropic — Opus 4.6 + $30B Series G

Links:

Two massive announcements in one week:

  • Opus 4.6 — Industry-leading on agentic coding, computer use, tool use, search, finance. Noticeably better than 4.5.

  • $380B valuation at $14B run-rate — Growing 10x annually for 3 years straight. The enterprise AI market is consolidating around Claude fast.

Engineered OOO — Three Problems with AI Agents

Link: https://engineered.ooo/blog/5

My own take on why the "AI will 10x everyone" narrative is more fragile than it sounds. Three deeply coupled problems:

  1. Human bottleneck — still need humans to review, approve, correct
  2. Quality ceiling — agents plateau on complex tasks
  3. Cost — inference at scale isn't free

They're not independent variables. They reinforce each other in ways that make simple "productivity multiplier" math misleading.

Felix Krause — OpenClaw: My Automation Setup

Link: https://krausefx.com/blog/openclaw-my-automation-setup

Felix (creator of fastlane, now ContextSDK) has built an incredibly thoughtful OpenClaw setup. Highlights:

  • Travel bot — parses booking confirmations from email, generates custom packing lists based on trip type (jungle vs. city), reminds about missing bookings, shares gate/seat info right before boarding
  • ContextSDK integration — feeds phone motion context (walking, sitting, in pocket) to the agent so it knows when he's on-the-go vs. at his desk
  • Security model — separate read-only bots with limited skills (e.g., travel bot can only read travel markdown files, not send messages)

This is one of the most complete real-world OpenClaw setups I've seen. Lots to learn from his architecture decisions.

Last updated: February 24, 2026