CharlieKey takeaways from your favorite podcasts
← 17 October 2026Latent Space

High Taste Testers: OpenAI Shipped Everything. What Actually Works?

Four heavy AI users review OpenAI's DevDay releases and the current model landscape, arguing the labs struggle with product polish while personal assistants, fast decision models and new frontier models reshape daily workflows.

Watch on YouTube ↗

Charlie's summary

  1. Cloud execution is still unsolved: agents stop working when you close your laptop, despite the “close your laptops” pitch

    The group says OpenAI's mission accomplished pitch hasn't landed: one speaker closes his laptop and his agent stops. They point to Pi Durable, which externalizes agent state into an append-only log so work moves between local and cloud. Codex and Claude Code already keep a JSONL file internally but haven't committed to a stable API for it, so others build their own harnesses.

    ▶ 4:39
  2. The labs are weak on product: DevDay updates felt like regressions and features are buried

    The group says the Codex desktop app got “less and less right” with each update: extra sidebars, worse performance, freezing. Dots and a maps feature are hidden in the UI, which they attribute to protecting millions of mainstream ChatGPT users. They also argue the Codex/ChatGPT merger and enterprise-plus-consumer Dots risk becoming “something for everybody and end up being for nobody.”

    ▶ 8:12
  3. Call them assistants not agents: proactive, always-on helpers are the right idea for non-developers

    One speaker is on a crusade to rename proactive agents assistants. Agents reads as a scary developer word, while everyone knows rich people have personal assistants who schedule things and triage email. Others note you could get much of this by adding cron jobs and storage to a coding agent, but the first speaker counters that he doesn't want to maintain more infrastructure, since the goal is reducing load.

    ▶ 14:18
  4. Trust is the real barrier for personal assistants: only a few companies get access to your email

    Several speakers connected personal email to an agent for the first time with Dots. One trusts only Apple, Google and OpenAI; others hesitate over Meta. Even sceptics worry less about malicious prompt injection than about an agent being helpful in the wrong way, such as telling a stranger the wrong thing or leaking details. Most keep to read-only, with explicit approval before anything is sent.

    ▶ 15:19

Give your email to unlock the rest of the key points — for every episode.

Do more with Charlie

Free — sign in with your email, no password.