←
Alexander Embiricos4,005 posts
Alexander Embiricos avatar

Alexander Embiricos

@embirico

Codex @OpenAI

◷ Joined November 2011▧ 309 Photos and videos
Alexander Embiricos avatar
Alexander Embiricos@embirico·Oct 16

This was such a fun conversation with @danshipper! My quotes here are more raw than usual. I didn't expect to be quoted, but then when Dan shared this AI-forward interview+transcript idea, I just couldn't say no.

Codex is one of the most important untold stories in the history of technology. A year ago, it was a scrappy effort in a quiet corner of @OpenAI. Its cloud product hadn’t found PMF, and the team was starting over from scratch. Less than 12 months later: - Codex swallowed the flagship ChatGPT desktop app from the inside (it literally runs as com.openai.codex on your Mac) - Codex and ChatGPT for Work hit 40M+ weekly users across engineering and knowledge work - Its leader @thsottiaux was elevated to run all of OpenAI’s core products, including ChatGPT, Codex, and the API How did they pull it off? They trained their own dedicated model line, separate from ChatGPT, with researchers and product builders sitting on the same team They built four different products in a single year (cloud → terminal → IDE sidebar → a desktop app with no code editor at all) as every model leap made a new form factor possible They used Codex to build Codex—so a handful of people could prove a new product in weeks instead of arguing for it in meetings I had a front-row seat to the whole arc—from their failed Cloud launch, to OpenAI’s offices the night before GPT-5 dropped, to the day Codex took over ChatGPT. And I spent months interviewing @gdb, @thsottiaux, @embirico, @mia_glaese, @aambrosino, and @edbayes to get inside how it all went down. Here is the full, behind-closed-doors story of the team that remade OpenAI: every.to/p/how-codex-happene… We’re pushing the boundaries of how stories like this are told. This piece is completely agent-native—designed to be read alongside your agent inside ChatGPT’s in-app browser: - Every interview transcript is included in full: Click any quote and you (or your agent) can see the exact conversation it came from. - Built for your agent to explore: Ask your agent to pull out product lessons for your team, cross-examine the transcripts, or find insights I missed when writing the piece. It’s a story about the future of software—built to be read with an agent: every.to/p/how-codex-happene…

0:29
Alexander Embiricos avatar
Alexander Embiricos@embirico·Oct 9

We started with dots on desktop for power users, and now we’re opening up to mobile:

Fresh updates for dots. First up: you can now create your dot straight from your phone in the ChatGPT app on iOS and Android.

2:46
Alexander Embiricos avatar
Alexander Embiricos@embirico·Oct 9

me when i get a phishing email for x

dot - meeting starting now grok bot - your meeting is now poke - meeting now instinct - jump on that call now openai meetings plugin - meeting take notes? granola - join zoom + take notes dia - join zoom technically i shouldnt miss any meetings now. im late because of this tweet ffs

Alexander Embiricos avatar
Alexander Embiricos@embirico·Oct 9

The Codex sandbox on Windows is one of our top areas of user feedback. Very excited to ship a new sandbox built on MXC. Let us know what you think!

An update for builders using Codex on Windows: We’ve built a new sandbox mode using @Microsoft’s Execution Containers (MXC) for faster setup, stronger network enforcement, and granular file access controls. Requires a compatible Windows 11 device.

Alexander Embiricos avatar
Alexander Embiricos@embirico·Oct 8

A colleague at @OpenAI just told me they got something wrong in a meeting, and their dot DM’d them a correction they could share during the meeting 👌

Alexander Embiricos avatar
Alexander Embiricos@embirico·Oct 7

we're working on improving dots and bringing the capabilities to all. in the meantime, here's an absolutely magical upgrade to ChatGPT that's making its way to Free & Go users tomorrow!

GPT-6 and Intelligent UI, now rolling out in ChatGPT for everyone. Intelligent UI in ChatGPT delivers fast, interactive answers that make everyday questions more visual, complex topics easier to grasp, and interactive tools for your task available on the spot.

1:10
Mitchell Hashimoto avatar
⟳ Alexander Embiricos retweeted
Mitchell Hashimoto@mitchellh·Oct 7

I just called Chungus (my OpenAI Dot) on the phone while picking up mexican food and asked it to transfer some Rex benchmarks from last night on my Mac Studio to my MacBook Pro via Tailscale and I just got to my desk and there they are. What a fucking time to be alive.

Alexander Embiricos avatar
Alexander Embiricos@embirico·Oct 7

many more improvements incoming!

One week since we launched dots! We’ve been working through your feedback and shipped a bunch of improvements: - Faster cloud browsing - Fixed extra phone notifications while you’re already using ChatGPT - Cleaner code blocks and attachments on web - Fewer pauses while waiting for a reply - Faster sidebar loading, smoother animations and less loading jank - Fixed character rendering in Safari and Firefox - Fixed Outlook setup getting stuck - Clearer plan guidance + a way to retry failed activity loads - More stable safety banners on web - Fixed clipped approval buttons and oversized Gmail previews hiding approvals - Fixed enterprise access issues on web and when starting voice - Better dot organization of when to use new versus existing codex threads Up next: - Voice reliability (being improved everyday, but we’ll shout when it’s up to par with our expectations) - Increasing dot’s control of your Codex app: manage project folders, have full Codex thread context, rename threads, and more - Better organization to know what your dot is working on at any given moment - Connect your dot to multiple computers - Better Windows support for local computer control - what else? Thanks for trying our product and giving us lots of feedback. Keep it coming!

Alexander Embiricos avatar
Alexander Embiricos@embirico·Oct 6

Excited to ship the Meetings plugin! I use it so my dot has the context it needs and stays on top of followups. Also, this is a great example of what you can do with a ChatGPT Plugin Extension!

Your next meeting comes with a note-taker who can help with the follow-up. The Meetings plugin takes notes for you and saves a personalized summary and next steps in ChatGPT Space, based on what ChatGPT knows about you and the work you’ve done together. Keep notes private or share them with your team, then ask ChatGPT to update a project plan or draft a follow-up. Available in beta for Pro and Business users in the ChatGPT desktop app on macOS. Enterprise is coming soon. Download the desktop app, then search for “Meetings” in the plugin directory. chatgpt.com/download/

Post attachment
Alexander Embiricos avatar
Alexander Embiricos@embirico·Oct 6

“My dot will follow up on that” is now a normal thing to say in meetings at @OpenAI. And it just works!

OpenAI's granola competitor (the meetings plugin for ChatGPT) kinda flew under the radar in announcements they have this UX where they just put buttons alongside the camera notch on macos and i really like that. it's much more unobtrusive than what wispr and granola are doing

Alexander Embiricos avatar
Alexander Embiricos@embirico·Oct 6

simple example but i think this is a balanced take about where different agents shine. (spoiler: dots shine on intelligence for hard tasks)

i just gave the exact same task to Muse, Dots, and Instinct (didn't do grok bot because my quota's already tight) i asked them to research good product names for something i want to ship, and i gave a bunch of criteria for what i look for 1. muse was very fast, but outcome was questionable muse was the only agent that went straight into the work without asking me any questions - not sure i like this but boy was it fast. i did 3 rounds of iterations with muse while dots and instinct were both still working on the first iteration BUT... guess what was the top suggestion from muse? it was "endframe" 🤦 i was like.. dude, did you not watch silicon valley? in follow up iterations, i also noticed muse was not quite good at following my intent in a way i like. for example, i said in the very beginning that every name needs to check for domain availability, but later i found out that some names weren't checked overall, it seems to optimize for quick iterations rather than thoroughness 2. instinct was very slow, but quite autonomous instinct first asked me some questions to anchor its work. i liked that because it helps steer its work to be more aligned with what i want then it went dark for 20 minutes to do the work. it did a pretty thorough research and came back with a file i could open and read. the file had a table with all the candidates and their readiness overall i thought that was pretty good, except for the names themselves.. they aren't that good. this feels like a general problem with LLMs - they seem to be really bad at coming up with names humans would really find appealing 3. dots was also slow, but it's more "asynchronous" and more intelligent like indistinct, dots also started with asking me some questions, but it didn't let the questions block its work! it had already started work in the background, and my questions were sent to it asynchronously as steers. i think this shows openai's slightly more experienced with harness engineering it also didn't go dark for the full duration of its work. it kept giving me updates about its progress, which was useful because i spotted some names i liked and didn't like, and gave it feedback early for what to lean into also gotta say, having astra in dots is clearly an advantage. the model is just smarter. it kept all my inputs in its mind, including the original goal i stated plus all the follow up steers. it "gets me" dots was also the only solution that actually gave me a few names i'd consider just a one-off experience here - take it with a grain of salt. i'll share more examples like this when i give them tasks!

Alexander Embiricos avatar
Alexander Embiricos@embirico·Oct 6

Every few days during dots development I would reset my dot to go through onboarding again. Not just the first few screens, but the full experience of prompting my dot to get to know me. It's getting increasingly painful. Pour one out for my latest dot 🫗🫗🫗

Alexander Embiricos avatar
Alexander Embiricos@embirico·Oct 6

Here's the plan for one of our top areas of feedback on dots: voice!

Dot voice: what we're fixing ⚒️ We're hearing lots of great feedback on Dot voice: 1. It takes too long to connect or fails entirely (5-6 rings) 2. The phone call UX: some people love it, some don't 3. No way to see transcript or summary 4. Says "Checking" too much 5. Sometimes goes quiet and does not give updates Here's what we're doing next: 1. Making connection fast and reliable 2. Keeping the phone call UX for now, but instant 3. Adding transcript and summaries to the UI 4. Adding more variance to "Checking" (eventually will remove it for fast questions) 5. Adding more regular updates Our goal is for Dot voice to become the chief-of-staff you can delegate any task to, hands-free. What did we miss? Please keep the feedback coming!

Alexander Embiricos avatar
Alexander Embiricos@embirico·Oct 2

Great intro to the dots takeover we’re seeing at @OpenAI for coding

Saw some asks about whether you should use dots or Codex. In reality they work great together! My dot is now my favorite way of using Codex. Less mental overload and the proactiveness is a lifesaver

Alexander Embiricos avatar
Alexander Embiricos@embirico·Oct 2

love these examples of dot proactivity

This is crazy. My Dot proactively messaged me to point out that a bunch of our CI jobs would start failing on Monday due to GitHub's macos-14 runner retirement.

Post attachment