Codex Overtakes Claude, NY AI Ad Law, DeepSeek V4 Pro — May 25

336 messages · 73 active members

336
messages
73
active members
@mb29266, @Wootbro, @tidemid
top contributors

Overview

Today's conversation was dominated by a clear migration trend: builders are shifting their primary coding work from Claude Code to Codex, citing generous and frequently-resetting usage limits, better instruction-following on precise prompts, and a maturing MacOS app with subagent visibility and phone continuity. Several members described hybrid workflows — Claude or Claude Design for creative/UI thinking, Codex for execution — though others traded horror stories of Codex introducing 'gaping security holes' in auth code, reinforcing that agents should be treated as rotating tools with planning and review phases. On the legal and policy front, @Wootbro shared a detailed breakdown of New York's new synthetic performer disclosure law, which appears to cover static images, narrated videos, and AI stock imagery — not just talking-head avatars. Builders running AI-heavy Meta ads debated enforcement realities, with most expecting platforms to bolt on automatic AI disclaimers rather than mass-disable ads. @arielletolome also shared a summary of the 2027 ACA Final Rule for insurance marketers. Infrastructure chatter included DeepSeek V4 Pro's permanent 75% price cut undercutting GPT-5 and Opus 4.7, cmux workspace orchestration, a TrapDoor supply chain attack hitting npm/PyPI/Crates.io, a phased MacOS bootstrap script for syncing dev environments across machines, plus Groq's Whisper V3 Turbo, NVIDIA's 0.6B speech model, and Grok v9 rumors.

Topics

Multiple builders reported switching primary coding to Codex, citing continuously-resetting limits, better behavior on precise prompts, and a maturing MacOS app with subagent inspection and phone handoff. Caveats surfaced too — one member flagged Codex introducing security regressions in auth code — so most are rotating between Codex and Claude as each regresses.

@Wootbro shared a detailed legal analysis showing New York's new law covers any visual depiction of AI-generated humans — including static images, silent figures in narrated video, and AI stock imagery — not just talking-head avatars. Builders debated enforcement, expecting Meta to auto-append AI disclaimers rather than mass-takedowns.

DeepSeek made its promotional pricing permanent, undercutting GPT-5, Claude Opus 4.7, and Gemini 3.5 Flash on frontier reasoning workloads. Builders flagged it as a serious option for enterprise-style workloads and lined up Bedrock comparisons against Nova Pro.

Builders agreed Codex is weak on raw frontend design but excellent when fed exported Claude Design zips or Mobbin MCP UI references. Recommended polish tools included impeccable.style, the 'make-interfaces-feel-better' skill, and Open Design as a local Claude-Design alternative running on CC or Codex tokens.

@geilt shared a phased setup.sh/bootstrap.sh handling brew, iCloud symlinks, nvm, SSH, kube, nginx, and agentic CLIs (claude-code, codex, opencode, grok, cursor, kimi), with configs kept separate from the command repo. @tidemid and others dug into cmux for managing parallel agents and floated inter-agent comms; links also surfaced for Groq Whisper V3 Turbo, NVIDIA's 0.6B speech model, and Grok v9 rumors.

Key Takeaways

  • Codex's resetting limits, instruction-following, and maturing MacOS app are pulling builders away from Claude Code — but rotate agents and always plan/review AI-generated auth code.
  • NY's synthetic performer law likely covers static AI images and silent AI figures in narrated ads — not just talking-head avatars — and applies across mediums.
  • DeepSeek V4 Pro's permanent 75% price cut now sits below GPT-5, Opus 4.7, and Gemini 3.5 Flash for frontier reasoning workloads.
  • Best frontend pattern emerging: design in Claude Design, export the zip, then implement in Codex; use Mobbin MCP and impeccable.style for polish.
  • Audit the TrapDoor supply chain attack across npm/PyPI/Crates.io; Socket CLI helps catch malicious package updates before production.

Hot Threads

@justingacinastarted

Switching from Claude to Codex — context, limits, and workflow

18 replies8 participants
@tidemidstarted

cmux workspaces, inter-agent communication, and wired headphones

16 replies9 participants
@Wootbrostarted

NY synthetic performer law: what counts as an AI human in ads

11 replies7 participants

Linked Items