Explore what's new in AI

thdxr wants Vercel to bundle native SQLite for lazy internal apps

The thread says internal apps are easiest when Vercel ships a native database, not just marketplace add-ons. That's a clear product lesson for agent-built apps.

Dwarkesh says automated firms will matter more than one super-AI

The thread argues power comes from collaboration and institutions, not just raw IQ. That implies AI progress may reshape firms before it produces a lone supermind.

Geoffrey Huntley argues for a programming language designed for agents

He frames today's languages as human-first and argues for a machine-first, spec-like language layer. It is a strong bet on where toolchains may converge.

Google's Gemini Spark leak shows an always-on assistant with app access

The screenshots point to a 24/7 assistant that can learn from behavior and connect apps and skills. If real, it is a big step toward consumer agent UX.

Jason's Codex maxxing guide lays out daily primitives

The blog is a practical map of how a Codex power user works day to day. Worth reading if you want a more deliberate agent workflow than ad hoc prompts.

thdxr says Cloudflare's built-in SQLite makes x-rank tracking easy

The same tracker that felt awkward on Vercel was easy on Cloudflare once SQLite was available. That is a concrete datapoint for low-friction internal apps.

OpenCode breaks on Kimi deployments that split reasoning fields

Some providers emit `reasoning`, others `reasoning_content`, but only one format is accepted. This is exactly the kind of adapter bug that ruins agent portability.

Learn Harness Engineering teaches the mechanics of AI coding agents

The site turns harness design into a course, with lectures, projects, and references. That's a rare resource for people building agent infrastructure.

Perplexity's Personal CFO tab pulls in portfolio and liability data

The new tab aggregates brokerage, options, earnings, and prediction market sources in one place. It looks like an agentic finance dashboard, not just chat.

OpenAI's Codex settings hint at a multi-device control network

The screenshot suggests Codex can authorize one Mac to control other logged-in devices. That could make remote agent handoff much smoother, but also riskier.

techhalla shows a two-prompt storyboard-to-commercial workflow

The thread turns a finished ad into a repeatable storyboard-then-animate recipe. Useful if you want a concrete multimodal production prompt stack.

Bun's rust-undefined-behavior-exorcist skill tracks 70 UB fixes

The workflow turns undefined-behavior cleanup into a graph of tracked fixes. That's a good pattern for making large remediation efforts manageable.

threepointone's Agent Starter adds subagent chats and handoffs

The starter is experimenting with subagents that can chat, summarize back, or stay read-only. The default interaction model is still an open design question.

MagicPath connects Codex, Claude Code, and Cursor to UI design

The tool imports Figma or web UI and lets agents keep iterating on the design. That makes it a practical bridge from mockup to code.

Stanford says one LLM beats many agents when thinking budgets are equal

When token budgets are matched, single-agent systems usually win on multi-hop reasoning. The paper argues many multi-agent gains are just extra compute.

Google's Nexus paper folds event context into time-series forecasts

The framework treats forecasting as reasoning over events plus numbers, and cut Zillow MAPE by 86.6% in the showcased tests.

Figure's Man vs. Machine sorting race ends with a human win

A human still beat the robot after 10 hours, but only narrowly. The result makes fatigue, breaks, and task design central to the automation debate.

NandoDF ties agent causality to a one-line fix for delusions

The thread argues actions must be treated as interventions, and shows why that matters for SFT and delusion prevention. That's a useful mental model for agent training.

LlamaIndex workshop shows how to build RAG over PDFs and figures

It gives a concrete pipeline for PDFs, charts, and captioning, plus ParseBench to catch parser mistakes before they reach agents.

Transformer Explainer runs GPT-2 in the browser with live attention

The demo makes attention and token ranking visible while you type. It's a useful teaching and debugging tool for transformer mechanics.