Codex Resets, GPT Live Voice, Compute Politics — AI Daily Sep 12
394 messages · 55 active members
Overview
Topics
Members woke up to another Codex reset, with @jcartu burning 14% in an hour and debating why Anthropic keeps giving free resets instead of selling more $200/mo seats or a $500/mo extreme builder tier. @thewildzeno and others speculated it's a compute-constrained PR play, with Anthropic lacking the data center contracts OpenAI locked in early. Dario's slowdown essay landed poorly in the same thread — read as regulatory capture, not safety.
@samb69 discovered OMP has GPT Live built in via Ctrl+L or /live, enabling fluid two-way voice conversations with agents. @leewardbound confirmed shockingly natural interrupt handling, and members are pairing it with Firstmate for managing 20-30 parallel sessions while driving. Multiple builders separately raved about the same voice tech in the ChatGPT iPhone app — natural filler words, near-zero latency, human-like interruptions.
A long thread debated whether the US should subsidize AI compute the way cities fund stadiums, with members arguing data centers price-gouge and Nvidia subsidizes hyperscalers. @leewardbound noted AI economics scrambles the usual left/right divide. Builders also argued frontier labs can'anonymous hold $25/M-token pricing when DeepSeek delivers ~90% quality at $0.50/M, with @jcartu pushing GLM and K3 as even stronger open alternatives.
@jcartu argued OpenAI, Google, AWS, and Apple are all moving to custom ASICs for inference while Nvidia keeps the training market, and questioned whether America stays ahead once China manufactures competitive chips. @Tz1888 noted China only needs 'good enough' chips, not the best, and members pointed to Elon's Texas Terafab as a coming inflection point for robots and automated vehicles.
@rstmaur shared a battle-tested overnight agent playbook: scrape and monitor new ads/landers/competitors, summarize changes, check broken links, score creatives, prep next-day briefs — but never let agents make big decisions unsupervised. @Andrenator8 asked about custom scraping vs spy tools; @rstmaur recommended both. Separately, a builder demoed AI auto-stitching 10 clips with silence and filler-word removal, declaring they may never manually edit video again.
Words worth knowing
Three terms from the glossary.
Key Takeaways
- OMP now supports GPT Live voice via Ctrl+L — a game changer for orchestrating 20-30 parallel sessions hands-free while driving or multitasking.
- Anthropic's repeated Codex resets look less like generosity and more like a compute-constrained strategy — they lack the data center contracts OpenAI locked in early.
- Open-source models (DeepSeek, GLM, K3) at ~$0.50/M tokens are structurally pressuring frontier labs holding $25/M pricing.
- Overnight agent jobs work best when scoped to research, monitoring, and prep work — never let them make big decisions or launches while you sleep.
- Custom ASICs from OpenAI, Google, AWS, and Apple mean Nvidia keeps training but will lose meaningful inference market share to in-house silicon.
Hot Threads
Codex resets, Anthropic's compute strategy, and Dario's slowdown essay
GPT Live voice discovery in OMP and Firstmate orchestration
Government subsidizing AI compute like stadium funding