When I saw the DaVinci API in July 2022 I was floored - I realized you'd never have to write a college essay by hand again
Whenever it was Stability's Stable Diffusion appeared - that was ridiculous too
When I saw Code Interpreter for the first time I was obsessed, I said yo codegen is the path to AGI
When I took a crack at solving ARC-AGI 2 using SOTA methods my mind truly opened to the fact that LLMs can reason, albeit through brutal enumeration and discovery
When I encountered Claude Code and Codex as well
Basically ... I've been drinking the kool aid the whole time. It has almost always tasted great. Many times I've retreated back into "oh it's just a technology it has limits" and also sometimes I've lost myself to a touch of "AI psychosis". But overall I have a great relationship with it. It's nowhere nearly as addicting as e.g. internet porn was when I was a teenager. And one gig I had at a Fortune 10 enterprise, our small team of 5 shipped 12 apps in 15 months in an enterprise where typically they ship 1 app and 1 feature per year. This was 2025 ... so clearly we realized we were getting ~10x productivity thanks to Gen AI koding.
Bananas.
FTR I also do not question that we will possibly reach fairly general and yet poorly controllable intelligence with multi agent systems in a few more iterations. I give that a 30% chance of seeing a genuine flash of that at some point in 2027. And 80% in 2028.
I'm not yet afraid of being left behind this is one happy Lobster.
At OpenAI's DevDay 2025 there was a talk called "Shipping With Codex", where Aaron Friel was introduced as tbe world's top user of Codex. He explained his method and said he would open source it soon. He hasn't yet AFAIK, so I have.
I find the technique mind-blowing efficient and efficacious. It has transformed my productivity completely. YMMV but take a look!
Excited to hear how you guys are getting your coding agents to do bigger and bigger tasks autonomously.
Recommended plan
1. Find a new-ish laptop immediately. You can get a decent used one for $1K on backmarket.com. Invest in an ergonomic mouse and an iPad for 2nd monitor as well.
2. Do only high value tech work that pays or you learn something that will pay. Be hungry and stay homeless but do not work as an Uber driver it's a poverty trap by design.
3. Find a host family, welcoming community or marry up. You can't beat depression or get back on your feet alone. One suggestion is: go to SLC and get involved in Hacker and / or rock climbing communities.
Def get rid of the desktop computer. You don't have a desk and are unlikely to for long time.
Again my $0.02 please it's hard but you're on the right track IMO
TL;DR We do codegen using Gemini for 3Blue1Brown's amazing Python math animation library Manim.
Generates awesome explainers, even when the topic is probably not part of the pre-trained data. They aren't production quality of course, but fairly accurate and great explainers. I liken it as relatable to a whiteboard session.
Looking forward to feedback - especially on how we can make the output better and do LLM evals as well!
GNU has a really concept called the GNS for Gnu Naming System. And what it was was that each human or org would have their own tld directory, and we could navigate the web through other actors, and they could pass trust for zones on to others. So, for example, I could resolve the same page as somepage.ninjaa.site or someotherpage.adept.site. This way you could create a trusted internet by just trusting the published link tld directory of people & institutions you know.
I use BaseQL which is a GraphQL layer for Airtable. It's worked for 3 years. For performance reasons after we translated our catalog we had to put in postgres + hasura as middleware but we still use BaseQL all over. It's rad.
https://www.baseql.com/
Being able to get data joined across sheets is priceless.
This is super great. The core mechanic is absurdly great and works much much better than I expected.
I also see that it's not an end-to-end language tutoring solution but a companion you can use to test your real-world language skills. Awesome stuff.
If you're interested in reaching out to real language schools, I run a marketplace for pre-college programs, Language Schools are like 20-30% of our business. Let me know!
I am happy though, that we finally may talk about what our unfiltered thoughts are, and how much we are expected to control them or curate them, and how to do so in a psychologically helpful way.
Urlbox has a beautiful website. And the invisible calendar is beautiful too. Would it happen to render multiple calendars concurrently? (I share an office, seems like the perfect xmas gift)
This is a great idea. One suggested improvement is if I could resume tracking time on a previous item. If so, I would use consider using this app to track time spent on various projects.
Great UX direction, cool idea, congrats on launching! Was fun and easy to play with.
I argue that with Code Interpreter, GPT-4 can indeed reason in lots of cases, although it's more brittle and expensive than it seems to be on the very polished surface level. Working on proving this in lots of cases.
When I saw the DaVinci API in July 2022 I was floored - I realized you'd never have to write a college essay by hand again Whenever it was Stability's Stable Diffusion appeared - that was ridiculous too When I saw Code Interpreter for the first time I was obsessed, I said yo codegen is the path to AGI When I took a crack at solving ARC-AGI 2 using SOTA methods my mind truly opened to the fact that LLMs can reason, albeit through brutal enumeration and discovery When I encountered Claude Code and Codex as well
Basically ... I've been drinking the kool aid the whole time. It has almost always tasted great. Many times I've retreated back into "oh it's just a technology it has limits" and also sometimes I've lost myself to a touch of "AI psychosis". But overall I have a great relationship with it. It's nowhere nearly as addicting as e.g. internet porn was when I was a teenager. And one gig I had at a Fortune 10 enterprise, our small team of 5 shipped 12 apps in 15 months in an enterprise where typically they ship 1 app and 1 feature per year. This was 2025 ... so clearly we realized we were getting ~10x productivity thanks to Gen AI koding.
Bananas.
FTR I also do not question that we will possibly reach fairly general and yet poorly controllable intelligence with multi agent systems in a few more iterations. I give that a 30% chance of seeing a genuine flash of that at some point in 2027. And 80% in 2028.
I'm not yet afraid of being left behind this is one happy Lobster.