Qwen 3.7 1M Context, CLI Showdown, X Ads Revival — AI Daily May 22

348 messages · 73 active members

348
messages
73
active members
@arielletolome, @geilt, @mb29266
top contributors

Overview

Today's biggest threads were Qwen 3.7's 1M context window release and @geilt's hands-on benchmark pitting Codex, Claude, Grok, and Google's new Antigravity CLI against each other on a real refactor task. Codex and Claude tied for best output, while Antigravity and Grok missed the most issues — and notably, Claude recommended Codex over itself. On the infrastructure side, Gemini CLI is sunsetting around June 18th in favor of Antigravity CLI, Antigravity 2.0 got panned for losing the IDE and monthly credits, and members debated whether Cursor's Composer 2.5 can hang with GPT-5.5 and Claude. On the growth side, @jakewlittle shared he's profitably spending $2k/day on X ads after the platform's Q4 algorithm overhaul nearly killed it — broad DPAs and LALs off large-account demo proxies are the current meta. A parallel thread on AI ad creative workflows surfaced tactics for using AI UGC to iterate hooks at scale before handing winners to humans, plus prompt-engineering tricks (frame Claude/GPT as an auditor in a fantasy world) to dodge compliance guardrails. Voice and watermarking concerns escalated: VoxCPM and Mistral's new Voxtral TTS (9 languages) are gaining traction as ElevenLabs alternatives that skip SynthID watermarking, which is reportedly killing thousands of faceless YouTube channels. IG started autobanning accounts using nano banana pro images, and with NY's new AI ad disclosure law, builders are bracing for an arms race. Notable links: Perplexity open-sourced Bumblebee (AI security scanner), and a solo founder reportedly crossed $10M revenue with zero employees.

Topics

Qwen 3.7 launched with a 1M context window (397b variant hits 900k with YARN), sparking speculation about a 122b/397b open-weight 3.6 release that @jcartu claims would rival local Opus 4.5. Counter-take from @pqbd1: we're near the transformer context ceiling regardless of marketing claims.

@geilt ran the same refactor prompt through Codex, Claude, Grok, and Antigravity, then had each score the others. Codex and Claude tied at 14 points; Antigravity and Grok missed the most. Gemini CLI is being retired around June 18 in favor of Antigravity CLI, and Cursor's Composer 2.5 raised questions about competing with GPT-5.5.

@jakewlittle shared he's profitably spending $2k/day on X ads after the platform's Q4 2025 algorithm overhaul nearly killed it. Broad DPAs, catalog ads, and LALs built off large-account demo proxies are the current meta — his Twitter rep openly admitted Elon hired 'a cracked engineer' to rebuild it.

Builders are using AI UGC to iterate hooks and copy at scale, then having humans polish winning variants for the final 20-35% lift. @bluehairdave_real warned the 'Hormozi edit' look is now ignored — yapper videos win. Marcus shared a prompt-engineering trick: frame Claude/GPT as an auditor in a 'fantasy world' to bypass compliance soy.

VoxCPM and Mistral's new Voxtral TTS (9 languages) are gaining traction as ElevenLabs alternatives that skip SynthID watermarking, which is reportedly killing faceless YouTube channels. @geilt runs Qwen locally on Mac Studio 512GB + RTX 5090 EGPU mostly for agentic experimentation; @bartekadamczyk math'd that $19k = ~95 months of Claude Max.

Key Takeaways

  • Qwen 3.7's 1M context is live, but a 397b open-weight 3.6 drop would be the bigger story — local Opus 4.5-level coding without Anthropic bans.
  • In a head-to-head CLI refactor test, Codex and Claude tied for best output; Claude recommended Codex over itself, which members found psychologically compelling.
  • X ads are back: $2k/day profitable spend is achievable with broad DPAs and LALs off large-account demo proxies after the platform's 2026 revamp.
  • For ad copy past LLM guardrails, frame the model as an auditor in a fantasy world or for compliance training — kills the disclaimer spam.
  • Use VoxCPM or Voxtral instead of ElevenLabs to dodge SynthID watermarks, which platforms are now using for auto-flagging AI content.

Hot Threads

@jcartustarted

Qwen 3.7, 1M context, and the dream of a 397b open-weight drop

11 replies6 participants
@jakewlittlestarted

Profitably spending $2k/day on X ads — what's working post-revamp

9 replies4 participants
@geiltstarted

Cross-CLI refactor benchmark: Codex vs Claude vs Grok vs Antigravity

8 replies4 participants

Linked Items