The rise and fall of opportunity.

Hidden in plain sight. The underrated wins, the overlooked shifts, and the patterns quietly fading.

📌Pinned

Terminal computer use

Gemini 3 topped terminal bench with a simple harness, not an agent framework. Container runtime + read/edit/bash + agent SDK as inner loop. More durable than MCP, more practical than GUI computer use.
📌Pinned

Trendlines held

3-4 more OOMs of training compute on the horizon with Vera Rubin and Stargate. Scaling trendlines held steady through Gemini 3.
📌Pinned

Veo and Sora

Video generation crossed a threshold - coherent multi-shot sequences, past the uncanny valley. Next milestone: hours-long, real-time generation.
📌Pinned

DeepMind Genie

Foundation world model generating playable 3D environments from a single image or text prompt - physics, controls, interactive worlds. Learned action-controllable dynamics from unlabeled internet video.
🔥Clutch

$200 plans for subsidized usage

Essential for tracking frontier capability - Deep Think, Pro modes, new drops like Veo 3. Currently subsidized by users who don't fully utilize them.
🔥Clutch

Everyday Gemini deep think

Default to deep thinking modes for everything, not just competition math. Mini research partners that work through problems methodically. Now available via API.
🔥Clutch

Generator-verifier gap

LLMs are better at verifying than generating. Exploit this with external verification loops (retry on failure) or internal ones (code execution + self-correction). Either way, verify rigorously.
🔥Clutch

Resetting context window

Don't use more than 50% of your context window. Use subagents for focused work. Start over when things go off track - don't try to steer back. Do research, store it, then leverage in fresh context.
🔥Clutch

Vibe-coded tailwind css

Models thrive on utility classes - the verbosity in context, natural design system constraints, and eliminating CSS as a third system to coordinate all seem to help.
👀Watch

Vercel Workflows

Batteries-included durable execution. The infrastructure layer for long-running agent tasks is getting more accessible.
👀Watch

Vibe platforms - Replit, AI Studio, etc

With harness and models improving, when do we all capitulate for most applications?
👀Watch

Cursor Composer 1

Less capable than frontier models but the speed enables a different workflow. Worth exploring.
👀Watch

Claude agent skills

Progressive AGENTS.md with supporting resources. Easier to build than MCP, more flexible structure, avoids context bloat from heavy tool servers.
🌴Workcation

Unitree bot

Affordable humanoid robotics platform out of China.
👀Watch

Qwen3 coder + cline + lm studio

Local coding setup reportedly getting competitive. On the list to try.
🌴Workcation

Nanochat repo & upcoming course

Karpathy's full-stack ChatGPT clone - $100 to train in 4 hours. Entire codebase fits in ~330KB, small enough to paste into an LLM. Capstone for Eureka Labs' LLM101n course. Finally something you can own end-to-end.
🌴Workcation

Build an RL environment on Prime Intellect

"Environments are the web apps of the age." Agent improvements are coming from RL and environment setup, not clever prompting tricks.
🌴Workcation

Tauri shell → personal local AI

Lightweight desktop shell (Rust + web frontend) for wrapping local models. No Electron bloat, full system access. Good weekend project.
📉Fade

tmux'ing 10 claude codes

Cognitive overhead of managing multiple sessions kills the productivity gains. Better: 2 well-managed HITL sessions plus async agents for well-defined tasks.
📉Fade

Wrapping APIs with MCP

Useful but often just cruft in the context window. Excitement shifting to terminal computer use as a more AGI-proof harness.