Investor Nathan Benaich says integration gaps now bottleneck AGI progress.
Whenever I don’t use codex for a task, I ask myself why and usually realize that there’s some missing context, I needed to write a skill, or I just didn’t think to use it.
Rarely is it because the task is outside of the capabilities of the model. Overhang right now feels large.
Many users agree with OpenAI executives that integrations, context, and tools limit progress more than model capabilities, while some dismiss the claims as delusional marketing.
Me in the past few weeks:
- Hey, I can build a dashboard with Codex to monitor AI legislation
- Actually, I can also monitor news articles about AI
- Actually, I can push updates from all of those things into a single news app that I built
- Actually, I can just have this app sync in the background, and also get RSS feeds from SubStack and all kinds of other places
- Actually, I can also autonomously scrub YouTube videos I'm interested in for key content on my Mac Mini, have Codex summarize them, and push these summaries to the app as well
- Actually...
You are very attached to codex. This is an unhealthy addiction. Keep feeding your zombies as you punish us for addiction and risks while your zombies do not sleep, do not leave the house, develop social and eating disorders, and become completely dependent on the tool which you promote crazily and reward users for the above behaviors with extra usage, while you consider it dangerous for neurodivergent people to get help so that they can connect with the world, you promote the tool that does exactly the opposite.
@gdb greg, I asked codex to create a product press release, generate spec sheet + images, identify news outlets with the name of the specific reporter(s) 'most likely to publish', search for the email addresses, and send the packets; it emailed 54 reporters/publications . . . magic.
@RobertDMellish @gdb 100% accurate. This is absolutely the bottleneck for me now.
@gdb agree
@gdb that's interesting. Because I use Codex constantly for coding work, but still go to ChatGPT web for a lot of creative/research/brain-storming tasks. I don't know if this is just muscle memory or if the user experience on the web just feels more natural for certain tasks.
@gdb the problem is most practitioners dont have the disposition, imagination or desire to be effective at this type of tool use.
its like handing everyone a flute and expecting everyone to make the same magic, some people are just drummers, or non musical, or deaf.
@deredleritt3r @gdb and then you come face to face with the actual bottleneck being your attention span and cognitive bandwidth and time in the day
@jxnlco @bender_2716057 @gdb somehow always meowbooks fault
agi is bottlenecked by integrations.
Whenever I don’t use codex for a task, I ask myself why and usually realize that there’s some missing context, I needed to write a skill, or I just didn’t think to use it.
Rarely is it because the task is outside of the capabilities of the model. Overhang right now feels large.
@gdb #keep4o #OpenSource4o
@gdb 完全正确
@gdb think deeper Greg. what could you do, when every turn came with an actionable, editable ledger? #Fieldledger
:x will there be a day where we don't write down these skills as often, or rely less on md files & chat history and scaffolds ? T_T 🥲 ./scaffolding_hell
idk, i see todays models and yeah capabilities around coding and simple chained tasks look amazing but even playing pokemon feels a bit janky & amnesiac. Yeah you can do it that way,,, it is technically within capabilities,,, but why? its expensive, its janky to use 'chat history/transcript as memory' across turn, it makes it less fluid.. =_=' just because you can hobble the models over the finish line doesn't been we should be satisfied just yet. bleh I guess its always true that we can always push a little more but still...
Still think we are closer to the age of multimodal chatbots than we are to the era of agentic ai, what we have today though amazing =will seem like antiques along several dimensions: speed, costs, efficiencies, memory, ..
@gdb Clearly not doing any design or UI/UX tasks...
Yeah, and I think the context problem is massively underrated. The model can do more than most tasks require, but if it doesn't have the right files, the right docs, the right history, it stalls. So people blame capability when really they just haven't done the setup work to let it run.
I'm a lawyer, so best believe that I wouldn't be able to write a single line of code on my own. Yes, it's all Codex.
I prefer to build things carefully, one single feature first, iterate on it a bit, add another feature, etc. Just figure out what you really want and have it built by Codex step by step.
There will be failures, sometimes it won't get the UI you want exactly right, sometimes it won't implement something correctly because it misunderstood your instructions or you underspecified something, etc. But you can eventually get it to do what you want, and it's no longer a particularly painful process. And a lot of the time, it just one-shots whatever you wanted.
The added bonus is having Codex run on your computer and controlling it from your ChatGPT app in your phone. This removes lots of friction and is insanely useful.
@gdb The question is not whether AI will become more capable. It will. The real question is whether our observability grows with that capability.
Translation: Our Codex is already god-tier. The only problems are you failing to stuff enough context in, define new skills, or even remember to use it.
The capability overhang is so massive it’s touching the stratosphere, while you’re still crawling on the floor. Whose fault is that? This is textbook “blame the user” perfection: Model shits out garbage code → You didn’t give enough context Model forgets everything → You didn’t define the skill Model literally can’t do it → You just didn’t think to use it (wake up, of course it can)
Beautiful closed loop.
The “I didn’t think to use it” part is the sneaky one. I’ve noticed the blocker is often not model capability, it’s that the task still lives in my head as a one-off chore. The moment it becomes a named workflow with context attached, Codex suddenly feels much less like a chatbot and more like a teammate who was already there.
Investor Nathan Benaich says integration gaps now bottleneck AGI progress.
Whenever I don’t use codex for a task, I ask myself why and usually realize that there’s some missing context, I needed to write a skill, or I just didn’t think to use it.
Rarely is it because the task is outside of the capabilities of the model. Overhang right now feels large.