March 6th, 2026
ChatGPT just got smarter
This newsletter was brought to you byAssemblyAIGPT-5.4, no more restarts
gm legends, happy Friday.
GPT-5.4 Thinking lets you interrupt mid-response and steer it without starting over, Woven does five-minute relationship reps that are actually practical, and Saydi handles real-time translation on calls so bilingual conversations donāt turn into a slow-motion mess.
You prompt your LLMs, why not your speech-to-text?

AssemblyAIās Universal-3 Pro introduces a new class of promptable speech modelsābuilt for real-world Voice AI. It handles domain-specific language, multiple languages, accents, and noisy audio with ease.
Unlike traditional ASR, Universal-3 Pro lets developers guide accuracy with prompts, combining the reliability of speech recognition with the controllability of LLMsāso youāre not stuck fixing transcripts after the fact.
AssemblyAI is opening free access throughout February, and the Product Hunt community is among the first to try what promptable ASR can do.
š Try Universal-3 Pro for free
More control, fewer tokens

GPT-5.4 Thinking is now in ChatGPT, the API, and Codex. Itās built for long tasks: deeper web research, better context retention, and you can interrupt mid-response to redirect it without starting over. It also adds native computer use, tool search that cut token usage by 47% in OpenAIās test with large tool sets, and up to a 1M context window in Codex/API.
š„ Our Take: The interrupt feature is the real quality upgrade. You can stop it the moment it drifts and keep going instead of trashing the whole answer. And the computer-use score is the headline if you care about agents doing real work: 75% on OSWorld-Verified, which is already past the human baseline they report.
Feature requests are easy. Choosing is the job.

Nika started a thread about the classic early-founder trap: you take every piece of feedback seriously, ship every ācould you alsoā¦ā request, and suddenly your MVP feels like a junk drawer. She asks how people decide what actually gets built, and how youād convince a team to prioritize a feature users keep asking for when the founders want to focus elsewhere.
The best answers in the thread basically say: stop counting requests and start looking for pain. A churned user saying āI left because of thisā beats 100 people saying ānice to have.ā Others add filters like vision fit, effort vs impact, āfake doorā tests before building, and even whether a strategic partner will pay to bump something up the roadmap.
Five minutes for love

Woven is a relationship training app that makes you actually practice, not just read tips and nod. You do one small daily lesson, around five minutes, built around attachment theory, with stuff like rehearsing hard conversations out loud, watching demo arguments, and spotting the patterns you keep repeating when things get tense.
š„ Our Take: Relationship apps usually turn into daily question spam or inspirational quotes you ignore. This is way more direct. It is basically reps for communication, which sounds a little cringe until you realize nobody taught us how to do this and we have all been freelancing it.
Subtitles, but for real calls

Saydi does real-time voice translation for meetings, sales calls, and events. It has three modes (one-way, two-way, and transcribe), plus an AI context engine for names and industry terms, speaker labels, and auto language detection. Thereās a Chrome extension for Google Meet, Zoom, and Teams, and a QR join flow for events.
š„ Our Take: The boring features are the whole point here. Speaker labels, custom vocab, and not having to tap buttons mid-call are what decide if this is usable or just a neat demo. If it stays accurate when people talk fast and messy, itās instantly valuable.
Monday through Friday
Our ultra-fast Daily: Three takes on new products. Yesterdayās top ten launches. Thatās it.









