very cool. Congrats on the launch. first half of 2026 was the year when every one made such internal tools one way or another. I am happy you guys could build a product out of it.
The cool thing is even though this tool addresses a few, use case for reducing cognitive load, they happen so frequently that they add up.
I think they key with cognitive load is that the agents often produce a lot of trace/docs etc, in the end only a small % of the traces really are important to the final changes, because concise changes are typically very small and self contained.
One thing i see this becoming important with is maintaining internal tooling built using this. I maintain an internal docs system that helps me do designing before i build code, that docs system is completely vibe coded and i can add features to it very fast, but now its all grown up. A challenge then is can i understand just enough about a new proposed change to approve it? That is key to the doc system not becoming a burden in of itself, while maintaining a tight core feature set.
> end only a small % of the traces really are important to the final changes, because concise changes are typically very small and self contained.
yes, totally. i think this is mostly 1 piece of the broader puzzle. we found that, esp. for complex changes where i need to spend my brainpower anyways:
> A challenge then is can i understand just enough about a new proposed change to approve it?
whiteboard has been a powerful tool for us! i think much more of this needs to be instrumented as part of a larger system, as you said (we are thinking the same way btw: https://dev.fast/about/)
Classification models lend themselves to sparsity and explainability. The good ones are very simple and economic to run on a laptop. If someone told you before 2022 that a json classifier was a product you would have laughed it off.
To add to this the more difficult problems in classification done on scale have always been about collecting "good" -ve examples, enough data to calibrate on every confidence interval and debugging outliers. And those are solved on a case by case basis by the company pursuing its own peculiar version of the problem.
Am I the only one who thinks this is just all hype?
And yet I(and many others publicly on x) chucked Claude for codex back in June. Codex is my workhorse. The main reason I use codex is because of its language. I just got really tired of reading long weirdly worded prose that I had to fix with some skill(though some people like matt pocock and dex horothy have good ideas on this, buts it's just wasteful). This is triply bad for learning newer stuff because it goes up and down the abstraction layer on any topic like mad. One moment it would be explaining a high level detail and then cite contrasts and then point indirectly to an implementation detail as an example. It's idea of explaining more abstractly was also weird in a different way.
Fable was better but I don't have 1000$/day to spend on it.
Hi thanks for the insight. Do you see a role for Control Systems(i.e. ones analogus to Instrumentation engineering) playing a role to modulate certain parts of continual learning? One very important way we learn are lived experiences, it's like telling memory:this part is more important( for emotional or social utility values), pay attention. Good or bad lived experiences both count. I guess is that a path that practical research is considering?
The core or the problem is also what you are describing was bought up by Yuval Harari in his interview with the economist.
To paraphrase some of his lines: We make decisions using our emotions and our thoughts. What makes us different from the AI is that we can be afraid.
To portray a guy as ordering "beef bulgogi", in the same breath as "email this rocket design marketing", while it _might_ seem appealing and resolute, though oddly fast paced, seems pretty ignorant of the human _quality_ that make most practical decisions messy.
All of this will be besides the point. Here is what's gonna happen. The frontier labs are just gonna keep building powerful models. AGI or not, open models in a year will be as powerful as Fable and Astra — probably by using em — and at a very soon enough point after that some one (a state or a few dozen people) with a few 100 GPUs is going to launch an unconscionable attack(if they have not already) that's gonna do a lot of damage.
Please for the love of god, just sit in a room with the government and put some restrictions around AI use before it harms a lot of people. Like tell the government to impose a minimum spend on frontier lab AI's spend on cyber defense and building every country's capabilities. The post-training mask for "I am a good assistant" is going to become a very sad joke when many people literally lose everything.
When one writes jupyter notebooks for DS you are not writing python. If you ask 10 DSs explain to me what python's attribute lookup model is and why is it different from other OO languages like say Java or C++, they would not care about it. The only thing DSs care about is the rich DS Library support and fast speed of protoyping. To a DS using jupyter this is almost the same feedback loop as a type system at compile time.
Have you tried using `uv`'s newer tools? They help a lot e.g. with linting speed, lock management, package dependency separation, correct python version mgmt and no need to fudge with venv.
The cool thing is even though this tool addresses a few, use case for reducing cognitive load, they happen so frequently that they add up.
I think they key with cognitive load is that the agents often produce a lot of trace/docs etc, in the end only a small % of the traces really are important to the final changes, because concise changes are typically very small and self contained.
One thing i see this becoming important with is maintaining internal tooling built using this. I maintain an internal docs system that helps me do designing before i build code, that docs system is completely vibe coded and i can add features to it very fast, but now its all grown up. A challenge then is can i understand just enough about a new proposed change to approve it? That is key to the doc system not becoming a burden in of itself, while maintaining a tight core feature set.
reply