maherbeg

Thank you for being explicit with the math!

So yes, at that speed for sure. But if the speed goes up? or the ability to batch at the same speed goes up? The economics start to shift. The gap is much closer, and you'd end up with a box you can still use or sell later.

Subscription pricing is still the best though!

I think this is a bit of a crazy statement. Everyone expects Apple to somehow build a category leading product every year. I'd expect something innovative every couple of years

* the iPhone * the iPad * apple watch * airpods * unified memory laptops and computers

Those are all products that either created a category or changed that industry.

If I can get Sol level capabilities on a $20k machine, then it is well worth it for my employer to buy me that machine for work as a workstation. When you start paying in tokens vs subscription costs due to enterprise agreements, you really start to see how much cash utilizing frontier models at the frontier costs (and I'm efficiently using luna and other models where possible!)

I'll say that Gastown sounded absolutely crazy, but the idea of having orchestrator threads to manage your work and keep tabs on it, having validators to validate the other work etc. were generally the right shape. I think GasTown probably could have been really successful if there was a pared down version with more obvious names rather than the fun names.

I'm going to be thinking about this blog post for a while though because if you squint and tear it apart, there are probably really good generalizable pieces in here to take home for future models.

Some of these feel like solved problems effectively, so having them in the standard library is nice (at the expense of keeping these forever for backwards compatibility once a new tech replaces it). I do think having a larger standard library for common things (like golang) is the way to go. If a dependency seems to basically be installed by default everywhere, maybe it should go in the standard library.

I think part of this is to enable the interface on the web and other devices where you might not have a terminal. But yeah, agreed, it's wayyyy too many tools.

The greatest gift the childless have is not knowing how amazing having children is.

This is a great name! It ties back to git nicely (git pull origin/main), and sounds human.

Cursor was also a great name given that it evolved into an advanced AI assisted auto complete. I think their team does a solid job with naming.

Does anyone have a https://tenstorrent.com/hardware/cards to try it on?

The Codex desktop app lets you highlight anything in the transcript and add it to the chat as an annotation.

I mean at the rate of model releases happening, I think a lot of these will collide more often than expected!

The token usage is really interesting. I would imagine the most efficient thing is to keep the state of everything persisted, and past the cache expiration window, to automatically start a new session with the previously persisted state instead of just a long running conversation.

If someone solves this part of continual effective compaction + selective resetting at cache expiry, they're going to make a ton of money. Right now, only the token insensitive can use these sweet features.

The 3090s also don't have enough VRAM to run larger models too. It really comes down to how you like to develop (synchronously with lots of steering vs async agentic)

GasTown was crazy but had a rough shape that was ahead of its time. I think https://github.com/kunchenguid/firstmate is a much better, and easier to understand version of agent orchestration.

Now that even the smaller models from labs (Luna, deepseek v4 flash) are getting powerful, I think the orchestrator pattern of a smart model coordinating smaller models for work will end up being the way to go.

Yes, that's true, but in the post, they even mention that building the spec is the scarce resource.

  For that to work, the swarm has to actually follow the spec, which is what much of this post is about. We gave the swarm 835 pages of prose and it came back with a database. What was scarce in this experiment, and what we expect to be scarce in software engineering going forward, is the right description of intent.

Yeah, agreed. It makes everything else look pretty bad. I still like to manage my work in a more structured form, but Codex can just rip on a goal end to end in the same thread.

You also might like https://github.com/nicobailon/pi-boomerang that lets you effectively do some work, compact that section with a summary and continue on.

I'm pretty annoyed with how fast this feels. Wish MacOS was this fast launching things.

oh yeah, sorry, I should've started my comment with "this is awesome" because it is, you did a great job explaining all the layers and the interactions within. I love stuff like this even if it's not necessarily prod ready by default!

A few of us built nearly the exact same thing for a Hackathon which was fun. This definitely can work. There are a couple of other approaches too that are interesting like

  - xata - https://xata.io/blog/xatastor-zfs-nvme-of-for-millions-of-postgres-databases
  - neon - which has a more sophisticated architecture that builds abstractions at the Postgres layer

But separating compute and storage sucks and the performance you get out of EBS and friends is mediocre. The elasticity is nice, but if you have High Availability and can move instances around, you can still expand your cluster relatively easily, just not easily in an emergency scenario.

This is really pretty, but I'm surprised it doesn't have a microphone. I know it's just rebranding the existing work louder creator keyboard, but a mic would really, really help with this product. Especially one that is really effective at Wispr Flow-type speaking.

What's better about the dedicated player out of curiosity?

I wouldn't care where my data went if it was to help my children.

I would say one thing I've enjoyed about the latest frontier models from US labs is that you just work at a higher level of abstraction. You can talk about the end goal and it'll just rip. You'll add scaffolding to constrain the patterns etc, but I do way less baby sitting than I expected on 5.6 vs 5.4 vs Deepseek v4 Pro.

I wouldn't call this minor. The 2 big features seem to be integrated memory + ambient proactiveness. This requires pretty intense tuning to not be annoying.

I've found the prompting needs are drastically different from the latest frontier models to the latest open weight models. I can be much more vague and talk about an end goal with the frontier models vs needing to be more prescriptive + have a workflow on the open weight models. This gap continues to close, but the level of abstraction I'm working on with the latest models continues to move much higher.

Cursor had a specific licensing agreement that allowed them to brand it how they want.

I wonder if new models will be trained with speculative decoding as a core feature allowing fewer experts to be needed for a pass.

I'm a big PGDog fan! It really helped us scale our connection proxy needs pretty substantially and it has great features like auto mode to support Aurora failovers neatly. It's infra that just works.