Ox-Alpha Model, ChatGPT iMessage, WisprFlow — AI Daily Aug 21

348 messages · 70 active members

348
messages
70
active members
@jcartu, @aa12on, @fmill1
top contributors

Overview

Today's chat opened with deep testing of the mystery 'ox-alpha' stealth model on OrcaRouter. @samtome ran it head-to-head against Opus 5 on a 47-shot video analysis pipeline and found richer field coverage and sharper OCR (0.719 vs 0.698), but at 11× the latency and roughly double the hallucination rate. Meanwhile OpenAI's new Apple Messages plugin for the ChatGPT Mac app sparked a practical discussion — builders flagged cross-thread search as the standout unlock over autonomous sending, which fed directly into a resource-rich roundup of iMessage-for-outreach providers led by @ecom2023's 17-link vendor list. The voice dictation debate returned, with WisprFlow taking a clear win over Superwhisper — builders cited better learning from user speech, Zoom/Slack/Meet meeting summaries, an iPhone app, and a $112/year price (or ~$5 with edu email). Otter surfaced as a phone-based alternative to dedicated devices like Plaud via MCP, while FluidVoice and FreeFlow got mentions as free/local options. On the model/infra side, builders continued migrating workflows to Grok for speed and to GLM 5.3 for local runs (4× RTX Pro 6000s needed). Claude Code is increasingly refusing scraping builds, pushing users toward Grok, Apify+CC, or Kimi. @samarh90 also hit a useful Claude safeguard blocking plan-mode edits to harness config files, and @tvrxm shared Tencent's HYCreator for long-form video generation.

Topics

@samtome benchmarked the ox-alpha stealth model on a 47-shot video analysis pipeline against Opus 5. It edged OCR recall (0.719 vs 0.698) and produced 2.8× richer descriptions, but ran 11× slower (81s vs 7.5s per call) and hallucinated ~2× as often — inventing punch-ins, conveyor belts, and 'revealed' objects. Verdict: real strengths, but fabrication-under-confidence disqualifies it for generation-feeding layers.

OpenAI's new Apple Messages plugin for the ChatGPT Mac app can search, summarize, draft, and send iMessage/SMS/RCS threads inside Codex and ChatGPT Work — with cross-thread search flagged as the biggest unlock. That fed into @ecom2023's 17-provider roundup for iMessage outreach, including Clawmessenger ($50/mo for 500 msgs/day) and @Andrenator8's success with Tuco.ai, with deliverability and per-message margin as key watchouts.

The chat converged on WisprFlow as the winner for voice dictation: better speech learning, Zoom/Slack/Meet meeting summaries, iPhone and Android apps, and $112/yr (or ~$5 with edu email). @rockdm switched from months on Superwhisper and isn't going back, though @robinroy flagged latency on Apple headphones. FluidVoice and FreeFlow surfaced as free/local alternatives.

Builders continue routing by model strength — @mattkowald and @jcartu run GLM 5.3 locally (~4× RTX Pro 6000s, still available direct from NVIDIA at ~$14k), @arielletolome uses Gemini Flash and Cerebras-hosted Qwen for high-TPS subagents, and Grok is the go-to for legal writing. Claude Code is refusing scraping tasks, pushing users toward Grok, Apify+CC, or Kimi. @samarh90 also hit a useful Claude safeguard blocking plan-mode edits to harness config files.

New member Jeff kicked off a Plaud debate. @jcartu wears his daily but argues the subscription is unnecessary at your desk — just record with your mic and hand the transcript to your agent, unlocking uses like meeting-video psychological profiling. Otter surfaced as a phone-based alternative that connects to LLMs via MCP, avoiding a separate device. Pocket was noted as a no-subscription hardware option.

Key Takeaways

  • Ox-alpha edges Opus 5 on OCR and description richness but hallucinates ~2× more and runs 11× slower — dangerous for pipelines feeding downstream generation.
  • ChatGPT's Apple Messages plugin's biggest unlock is cross-thread search, not autonomous sending; iMessage outreach can beat SMS reply rates but watch deliverability and margins.
  • WisprFlow is the community's clear voice dictation pick at $112/yr (or ~$5 with edu email); meeting-notes integration with Zoom/Slack/Meet differentiates it from Superwhisper.
  • Claude Code increasingly refuses scraping builds — Grok, Apify+CC, and Kimi are becoming the fallback stack; build harness-level guardrails that block AI from editing its own config.
  • You don't need a Plaud subscription at your desk: record locally or use Otter+MCP, then hand the transcript to your agent for deeper analysis.

Hot Threads

@william_hodgesstarted

WisprFlow vs Superwhisper voice dictation showdown

20 replies14 participants
@ecom2023started

iMessage outreach providers and deliverability

10 replies3 participants
@samtomestarted

Ox-alpha stealth model benchmarks vs Opus 5

10 replies6 participants

Linked Items