Member of Technical Staff @a____t____g 9x hackathon winner SF
San Francisco Bay Area
Joined April 2025
- Tweets164
- Following249
- Followers104
- Likes292
Mark. retweeted
we built “let’s bump phones to exchange contacts” but for agents!
now your @instinct / @bot / @muse (whatever you use!) can make plans, negotiate deals, and get what it needs from your friends/coworkers’ assistants
it's “have your ppl (agents) talk to my ppl (agents)!” for everyone 🤝
built w/ @uncanny_guzus @ @ycombinator #MakeSomethingAgents hackathon (s/o @nemild)
on top of @stripe (agentic p2p payments) + @manufact (mcp relay server) 🙌
Who will you bump first?
Mark. retweeted
I resigned from Anthropic today. I spent the last three years doing pretraining research at both OpenAI and Anthropic. Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives. More thoughts below.
Mark. retweeted
i'll never forget that time someone demo’ed me a language model at a hackathon in 2022 and I thought to myself "this is just GPT-3 but worse"
and it turned out to be Claude 1
Mark. retweeted
WHERE ARE YC FOUNDERS NOW?
@_javokhir and I were curious where they go "after the startup". So we ran Deep Research on ALL 11,398 founders who have ever gone through YC.
Here's what we learned: joinedanthropic dot com
Personal update: I'm taking a leave of absence from YC to join Anthropic. I'll be working with @NotTomBrown on the compute team.
Powerful AI has the potential to improve the life of every human on earth and, as we enter the early stages of recursive self-improvement, availability of compute becomes one of the most important issues to solve.
I'm excited to get started 🚀
If Codex repeatedly reconnects after you send a message or takes 15+ seconds before it starts responding, check whether the task has accumulated lots of screenshots. In my case, completed compactions retained raw base64 image payloads, causing later turns to repeatedly process a huge image-heavy history.
Removing those retained image nodes restored 4-8s response times. I packaged the repair for this specific symptom as a skill:
github.com/markmdev/codex-th… (Just send it to your Codex if you hit the same error)
Why is nobody talking about this Codex feature?
You can select anything in the agent's response, click "More details", and
it creates a new ChatGPT chat that already has all the context and just
explains the thing to you.
It helps so much when you don't understand something in Codex's answer.
Before, I'd open a side chat, write a prompt, and wait while it
double-checks everything it already knows. Now it's one click, and the
explanation is almost instant.
And it's not just about code either. If you're learning something new, just
use it all the time. It's extremely powerful.
A month ago my Codex setup had 100 skills. Today it has 62, and the total instruction volume across them dropped by half in the last two weeks. Those cuts used to be judgment calls. This week I turned them into measurements.
The method is simple: each rule in my AGENTS.md gets a realistic task where the natural shortcut violates exactly that rule, and then the same task runs again with the file unloaded. If the agent behaves correctly both ways, the rule is a model default and the line is dead weight. That was the outcome for 10 of the 17 rules I tested. The four rules that actually changed behavior were the ones that make my setup mine: reading project memory before working, updating it after, stopping when a plan assumption breaks mid-task, and gating design decisions. And one rule failed every attempt no matter how I worded it, which taught me more than the passes did.
Full method and results in the article.
Attending AI Engineer World's Fair this week.
My Day 1 saved sessions lean coding agents / AI-native SDLC: Codex, Uber uReview, 1M AI-generated PRs, production code access, Claude long-horizon, Shopify River, Figma, Cursor.
If you're there too, what are you prioritizing?
My wife, our two dogs, and I did a 9-day, 3,500-mile road trip last summer. Each day had 500+ photos/videos across a photo camera, video camera, and two iPhones.
After the trip, I wanted to make short personal memory videos for each day, around 10-15 minutes each, using the best clips/photos from that huge collection. Almost a year later, I had finished only 3 of 9. It took forever.
Last week I thought: wait, what if I just give Codex everything? The first 3 finished videos, 5,000+ source files, and a description of what I wanted. I asked it to study the first 3 videos and continue in the same style.
After about 12 hours of work, and only 5 corrections from me, Codex produced a video that was better than the ones I made myself.
Then it saved the whole process, all my corrections, and everything it learned into a skill. Now, a week later, it's finishing the last day's video. For the later days, it usually needed only 1-2 corrections after the first version.
This is incredible. It probably would have taken me another year or two to finish the remaining videos myself. With Codex, I finished them in a week, and spent maybe 2 hours total reviewing versions and giving small corrections.
It brought us so much joy to watch all these memories come back.
Codex analyzed every video/photo for each day, understood the scenes, chose 150-200 sources per day, picked music, built the timeline, added transitions and collages, clipped videos, handled chronology, and rendered everything in 4K.
I didn't do ANY of the editing work. I just gave the initial direction, reviewed the results, and nudged it when something felt off.
fork_context:true subagents continue parent-thread work instead of following delegated prompt
github.com/openai/codex/issu…
