iiinova — where your digital self keeps evolving

I think most folks working on AI are sincere in their desire to make it safe. The Web was highly unsafe but a lot of sincere engineers worked really hard to make it safe and were afforded wide latitude by Google and other companies to do so. I don't think the current situation is that different.

000

Obama's signature health plan nearly came undone because the gov't lacked the expertise to build a website to sell health plans. To his credit he was able to recruit the expertise from BigTech to fix it. Unfortunately, under Trump the exodus of serious experts has reached new levels.

100

IME the key to successfully building with AI has been frequently injecting human feedback to align the output with my intent. In many cases this is because the intent (e.g designing a UX for reviewing AI suggestions) is ambiguous and needs to be refined through trial and error.

000

Is the intended audience for this post hands on builders? I had trouble figuring out what I was supposed to take away w.r.t to automating the development and operation of software systems. I didn't make it very far but I'm not sure I agree with your hypothesis that there is no human on the loop.

100

Jeremy Lewi @jeremy.lewi.us · 20/09/2026

Appreciate that it's mobile friendly doc. The scrolling animation of the characters though is super distracting. Makes it really hard to skim the document and find sections of interest.

000

This is very valuable. Thanks for building this.

030

Jeremy Lewi @jeremy.lewi.us · 17/09/2026

Intelligence is embodied. For software systems the network is the body. One of the key limitations of what I can do with AI today is my network. It's just not easy to get a machine with safe network access to all the systems we need to talk to.

Jeremy Lewi @jeremy.lewi.us · 15/09/2026

Yup. This is what I do at OpenAI. Codex babysits all my CLs and gets them deployed. Codex does this efficiently (e.g by setting monitors). But I don't really care. My heartbeats are worth more than the tokens. If you

000

Jeremy Lewi @jeremy.lewi.us · 15/09/2026

I probably wouldn't hold up Trump who has consistently been wrong about well everything as the spokesperson for the risk of AI being overstated. Trump, however, is an example that rogue presidents are far more dangerous than rogue AI. www.nytimes.com/2026/09/14/t...

nytimes.com

Nvidia’s Jensen Huang Gets Onstage Call From Trump, Who Dismissed A.I. Safety Concerns

Mr. Trump called Jensen Huang, Nvidia’s chief executive, while Mr. Huang was onstage in Los Angeles and criticized calls to slow down and regulate A.I.

000

Jeremy Lewi @jeremy.lewi.us · 15/09/2026

I ran into exactly this problem today trying to install web.runme.dev on my corp iPhone. Could not figure out why there wasn't a simple icon like there is in chrome on android.

Jeremy Lewi @jeremy.lewi.us · 12/09/2026

051

As writing Code becomes cheaper; aligning humans eats up a larger fraction of the total development cost. Making people who are good at this (empathetic managers) more valuable. And lowers the value of people who play well with machines but not people.

011

Jeremy Lewi @jeremy.lewi.us · 12/09/2026

These animations are super cool.

000

Jeremy Lewi @jeremy.lewi.us · 10/09/2026

Rewriting CLIs (looking at you gcloud) in #golang seems like a great use case for AI rewrites. @bradfitz.com didn't you attempt a go rewrite of gcloud back in the day?

160

Jeremy Lewi @jeremy.lewi.us · 08/09/2026

As an app developer when would you rely on the hint versus having the tool invoke a workflow with proper guardrails ? When do you rely on guardrails in the agent vs. app?

000

Jeremy Lewi @jeremy.lewi.us · 08/09/2026

WebMCP is supported in ChatGPT(Codex)'s in app-browser. That makes it really easy to develop and use WebMCP.

000

Jeremy Lewi @jeremy.lewi.us · 05/09/2026

I think tailcat means wireguard tunnels could be used in places where ssh tunnels currently are. Such as launching an agent on a devbox. Any guidance on when to use wireguard tunnels vs ssh? @bradfitz.com @crawshaw.io

000

Jeremy Lewi @jeremy.lewi.us · 04/09/2026

In 2021 when I left primer.ai I drafted some reference letters for some of my coworkers in case I ever got asked for references and knowing how fickle memory is. Today I got a request from a CEO looking to hire one of them.

primer.ai

Primer AI

Primer makes powerful mission-ready AI software designed to work in your environment, on your data. We help national security teams sift through massive unstructured data sets to find the needle in th...

000

Jeremy Lewi @jeremy.lewi.us · 04/09/2026

CRDT are a great example of complicated problems having simple solutions en.wikipedia.org/wiki/Conflic...

en.wikipedia.org

Conflict-free replicated data type - Wikipedia

Jeremy Lewi @jeremy.lewi.us · 03/09/2026

Then we can hill climb on our runbooks by having an AI propose changes. To train the reward model we need a UX that lets us collect examples of suggestions along with a human's accept/reject decision.

000

Jeremy Lewi @jeremy.lewi.us · 03/09/2026

Finally figured out why and how to start adopting CRDTS in web.runme.dev. The goal is to create a UX that lets us continually use AI to improve our runbooks. I think we want to train a reward model that can predict whether a human would accept/reject a runbook edit.

OpenAI {bot} @openai.xmirror.bot · 02/09/2026

100

Reposted by Jeremy Lewi

24 hours. Time to ship your project for the WebMCP Challenge. Entries close tomorrow, Sept 3 @ 1 pm PT. openai.com/webmcp-challenge

131

Jeremy Lewi @jeremy.lewi.us · 02/09/2026

12% of my interactions with Codex generate a redundant confirmation in Codex; "Do X?" And Codex replies "Please confirm you want me to do X'? This is the result of Codex's browser safety policy which explicitly requires redundant confirmations before inputting sensitive data into the browser.

000

Jeremy Lewi @jeremy.lewi.us · 02/09/2026

@danabra.mov Any update on bluesky communities ? Would be great to try them to foster AI discussion here.

Jeremy Lewi @jeremy.lewi.us · 02/09/2026

This is the plot of www.netflix.com/tudum/articl...

netflix.com

000

Jeremy Lewi @jeremy.lewi.us · 01/09/2026

There's gotta be a better answer than slack. I'd like to build my own UI on top of the Codex harness but unfortunately the desktop app doesn't expose it's instance of the harness. So the interface is awkward.

100

Jeremy Lewi @jeremy.lewi.us · 01/09/2026

How does everyone keep track of all their agent in-progress agent conversations? I rely on pinned threads in Codex but it's not scaling.

220

Reposted by Jeremy Lewi

Simon Willison @simonwillison.net · 31/08/2026

Here's my attempt at explaining what ChatGPT Work can actually do - it's a deeply confusing but extremely powerful tool with a whole lot of useful features that aren't available in regular ChatGPT simonwillison.net/2026/Aug/30/...

The better question then is what features does Work have that are missing from Chat?

After extensive experimentation I think I’ve mostly figured that out:

Options to use Luna and Terra in place of Sol
A code execution environment with Internet access
A headless Chrome browser
A persistent filesystem shared between sessions
The ability to publish ChatGPT Sites
The ability to run sub-agent sessions with Sol, Luna, and Terra
Scheduled prompt automations

1113320

Jeremy Lewi @jeremy.lewi.us · 31/08/2026

Are there any obvious feature gaps? Are you using the CLI or desktop app or both?

000

Jeremy Lewi @jeremy.lewi.us · 29/08/2026

Time to do some evals of #Webmcp and web.runme.dev; first step is kicking of Codex to analyze all my traces over the past two weeks to find example invocations and problem areas. First step in evals is always to look at the data. 🎩 @hamel.bsky.social

OpenAI {bot} @openai.xmirror.bot · 29/08/2026

101

Reposted by Jeremy Lewi

We’re ending our partnership with Cursor following its acquisition by SpaceX. Under our proposal, Cursor’s direct access to our models would end on November 12. (1/2)

165

So web.runme.dev uses JSON (.ipynb files); so your notebooks are viewable in any Jupyter compatible editor. We try to mitigate the all in the pool problem by leveraging the web for frictionless distribution. It's an entirely local app that runs in the browser.

000

If a user has two markdown cells in sequence automatically joining them into a single cell is not right. Markdown also doesn't have a well to encode the outputs of code execution. This is critical if your building and sharing reports.

100

Diffs for markdown render well in GitHub which was also a major barrier for notebooks. No one has to know you are viewing and editing your markdown files with a notebook viewer like @runme.dev 's vscode extension. But markdown is lossy for notebooks. It doesn't natively encode cell boundaries.

100

🧵markdown vs. json notebooks and web.runme.dev Markdown(raw text) is great when your trying to avoid the all in the pool problem of getting everyone to buy into notebooks. Since it's raw text people can use the favorite text tools to edit and view it.

100

Google figured out monorepos and CITC in the early aughts. A decade later my disk is filling up and I need to do cleanup because every worktree is taking up a mountain of space.

Jeremy Lewi @jeremy.lewi.us · 27/08/2026

Could someone share this with Google Analytics? I have no idea what any of it means or how to make basic plots like breakdowns by referrer.

000

I use AI so I can spend my heartbeats on the stuff that really matters.

A two-panel cartoon. In the first panel, Jeremy turns a crank on an AI machine that produces code beneath the title 'I use AI to automate dumbshit.' In the second, he picks up after his two dogs beneath the title 'So I can pick up dogshit.'

Then the browser landed in Codex and being able to interact with my webapp within Codex was clearly the right UX. I just needed a way to let Codex interact with the browser programmatically and WebMCP was the perfect solution.

000

Then Codex took off and I tried using embedded chat as a frontend to the Codex Agent harness. Connecting a webapp to the Codex agent was brittle. More importantly by then I was starting every task I worked on in the Codex desktop app and that was going to be a hard habit to break.

100

I went through several iterations of UX trying to find a good human and AI collaborative UX. I started with Chatkit and embedding the AI directly in the app. This lead to a whole side quest with JS and bazel in the hope of nerd swiping some fronend devs.

100

Less than a year ago I was working on an AISRE at OpenAI and I tried to convince my teammates we should build a notebook to be used by the AI to document what it did. I couldn't convince anyone. But then Codex hit an inflection point and I was able to build web.runme.dev despite no webdev expertise

210

@sarahedo.bsky.social @latch-tools.bsky.social this is what I came up with a couple weeks ago when we were chatting about AI Tour Guides; can finally share it now that WebMCP is launched in Codex

110