@nizzyabi - Reverse-Engineering Instinct: How Its

Open original

@nizzyabi on X Nizzy Abi, using a 12-week iMessage database and a stopwatch, measures how the Instinct AI assistant behaves over text: instant read receipts, tapbacks used as a small vocabulary, short replies, and proactive follow-ups. He then argues the internals are several ordinary systems wired together well, and that the hard part is a year of infrastructure, not the model.

  • Behavior is measurable: messages are read about 1.1 seconds after sending, 44% of exchanges open with a tapback, and replies are short (median 131 characters) with no bold and rare emoji.
  • Long tasks run separately from the chat channel, so the assistant can answer a quick question mid-research, which suggests a chat agent backed by task agents.
  • Memory and honesty are the weak spots: it kept a project it was asked to forget for 16 days, and it claimed a voice memo was sent when the audio never arrived.

Nizzy Abi spent twelve weeks using Instinct, an AI assistant that works over iMessage, and then went through his own message database with a stopwatch. His headline finding is that the product's appeal comes less from what it can do than from how it behaves in a text thread. On October 12 he asked which dentist network to call; the assistant corrected his premise within 43 seconds, found the right plan, and had him enrolled 22 minutes after his first text, having sent nine messages and filled in two secure forms. Across 185 of his messages, Instinct marked them read a median of 1.1 seconds after sending, reacted with a tapback in 44% of exchanges (about seven seconds in, as a kind of shorthand for "on it," "researching," "done," or "reminder set"), kept replies short (median 131 characters), and sent its own follow-ups 17 times in twelve weeks, some about tasks he had dropped hours earlier. The more important part is the guardrails around action. Before it sent an email, submitted a form, or bought anything, it showed him the draft or price and waited for approval. Sensitive data like SSNs and card numbers went through secure links rather than the chat. It admitted limits: it could not place phone calls on his account (though Abi notes this may vary), and it said so instead of inventing results. When it did make mistakes, it usually caught them first. In one case, a hat listed at $44 turned out to be $64 at checkout because the research had used member-only prices; the assistant stopped and asked before buying, then explained the discrepancy afterward. An email that bounced got flagged with a correction message, and a new address was proposed rather than sent blindly. Abi's framing is that the experience feels like a personal assistant because of this combination of speed, restraint, and visible checking, not raw capability. He also points out that the feel is mostly a product of decisions that take real engineering effort: presence indicators, message length, when to interrupt, when to wait for approval. He writes from the position of someone who has built a similar text-based assistant, which is why he thinks these details are harder than they look. The piece also draws on other users' reports, including investor John Borthwick's enthusiasm and Susannah Duke's observation that the assistant seems to watch her email. Abi's own data is a single user's, so the numbers describe his experience rather than a general rule.

Ask about this article