HN user
Developer for 35 years, 10 years at IDEO as a Software Designer, project lead and Design Director. Currently working on agentic native apps, and on the hunt for a profitable SaaS company to acquire via infinitecake.net
I think it would depend on how you measure it. Adding a review step will definitely increase cost, but theoretically that cost is lower than re-doing the subagent's implementation later when you discover it's broken.
The issue with this is just how much context it consumes. Generating the plan could easily get me to 30% of a 1M context window—the point at which I normally clear. Most of that context fill is not necessary for execution, and the model degrades as context fills. So starting from 30% (or passing a fresh agent all of that context) is not ideal. But maybe I'm misunderstanding what you're proposing?
Super ambitious, but maybe that's its downfall. It's hard to imagine incrementally adopting this.
Yes, and I'm arguing that the post is over-simplifying. As I mention, for small plans, it's probably easier to have the big model just do it. But for large, multi-session plans, it's cheaper to let smaller models execute the bulk of the work.
Yes, this is a core part of my workflow; use the smartest model (e.g. Fable) to generate a large hierarchical implementation plan, then clear the context and have a lesser model (e.g. Sonnet) track and execute the plan [0], often in parallel.
It all depends on how much work you're doing; if it's a simple task, there's virtually no need for plans at all; just have the smart agent do it. But if the work is going to take 5, 10 or 100 sessions, there's real value in the "smart agent plans, cheap agent executes" model.
I think it makes more sense if you think about an ASIC pipeline which is updated on an annual basis, like Apple's M or A-series chips.
Sure, an ASIC model will always be behind (12 months? 18 months?), and a hosted flagship will be significantly better. But as time marches on, would I use a flagship model from last year if it came on a PCI card and cost $1000? Without a shadow of a doubt.
It's undeniable that some of these models generate a ton of thinking tokens, but it's arguable whether that makes them "much less effective."
For example, Kimi 2.7 has been really effective for me despite having verbose thinking blocks, simply because it runs so fast. Speed-wise, it feels about like Sonnet, possibly faster.
Yeah, it's an unfortunate convention from the very first "1 bit" model. But to be clear, Bonsai comes in both ternary and actual 1-bit variants.
Exactly this. For whatever reason, Claude likes to talk about the "shape" of "load-bearing" "seams," but if that's the internal jargon it needs to plan and execute its work, who am I to judge?
But if I'm reading what is supposed to be someone's original thoughts, it's a huge bummer to see an obvious AI tell. You might say that "it's not just disappointing—it's disrespectful."
The phenomenon I'm describing is people switching to consumption-only mode, or only posting privately. I think you're right that engagement is going up, as people treat social media more like TV and less like microblogging. I define that as "checking out."
It's just a huge change from the early days of social media (Instagram 1.0, Flickr, Twitter, o.g. Facebook), where nearly everybody who was active on the platform was also posting.
Yes
This is why I just use my name… it also helps me dial in my tone when I know I’m putting my name behind a comment.
There's honestly something to this [0]… you should do it!
From what I’ve observed, people are increasingly checking out of social media (or turning their profiles private), because it’s just unappealing unless you’re actively promoting something.
People have said this about virtually every new writing technology, from pencils to typewriters, from computers to spell-check.
Personally I would never use AI for writing, but it's absolutely a power tool for coding. I still need to think about what needs to be generated, and why. For now, I still need to guide the model based on my experience. And in the end, I'm on the hook to approve what's created. If anything, I have more time to think about what I'm making, because I'm not busy building the CRUD for a web app for the 100th time.
There are certain things I enjoy coding by hand, in exactly the same way I like using a manual coffee grinder. Coding can be enjoyable.
But most of the time coding is a means to an end. I bet your friend in HVAC was not told to use a manual drill to have a deeper understanding of how to make holes for installation. AI is simply a power tool.
This article starts off by saying AI can be useful for question-answering, but then quickly states that if you're using AI to generate code "you're wasting your time:"
The reason AI is a bad tool is that generally speaking, it is completely opaque.
The author is missing the fact that you can use those great question-answering and knowledge distillation capabilities to ask the AI questions about the code it wrote.
Later, they write:
Who is going to verify that it is doing what it is supposed to?
What? If you hire a writer to write a blog post, who is going to verify that they didn't just fill in gibberish? It's you. You there, in the front. You're the one.
If you can't find any way to independently verify that the software is doing what you think it should do, I don't think you have any business generating it, whether you use AI or code it by hand. That's like commissioning a portrait of someone you've never seen.
My friend, there are over 2000 VFX shots in Fury Road, and at least 600 in The Fall Guy.
I get that people are tired of CGI, but it’s a tool that is used in virtually every film that reaches theaters, for reasons as prosaic as matching skies between shots.
It really depends on the tool… CRMs and EHRs are often designed to be customizable. But end-users typically don't want to spend a lot of time configuring and customizing the interface itself. All the choices quickly become overwhelming.
The effect of the interface becoming "invisible" is actually a function of time spent in the interface. I think what the author is reacting to is discretionary friction; designers or product folks adding features or complexity. The thing is, that friction may be necessary in order to achieve a certain task (think about resolving a merge conflict). And given enough time in the interface, even those "disruptive" steps fade into the background.
To give a concrete example, the console of a 737 is incredibly dense with controls. The airplane itself has many different modes, and there are many moments of intentional friction.
However, if you interview a pilot with 10+ years in a 737, they will tell you the interface has become invisible.
The same goes for the supposedly "bad" Bloomberg terminal. You'll find the same thing in Healthcare, where an interface cluttered with buttons is exactly the right solution for someone who spends 8+ hours/day in a MR scanning software and wants instant access to all the controls.
As programmers, I think we're too quick to generalize our own experience and preferences and try to apply them to others.
Source: I spent 10 years designing consumer and professional software at IDEO
Data you collected or data you found? Any idea where it was?
I think it was a lengthy blog post citing manufacturer data, found via the FreeNAS forums. The added safety of RAIDZ2 was a good tradeoff. Someone else on this thread reports they lost two SSDs in rapid succession, which I hear happens sometimes with drives from the same batch. In any event, I felt no need to launch an independent investigation of the math and underlying assumptions.
Zpool status gives you a per-disk error count. I kind of assumed someone doing a resilver would use that command.
Yes, I did indeed use zpool to do the resilver… I don't recall if there were any read errors on other drives at the time of those failures, but probably not, as the scrub probably would have flagged it. They've all been replaced with higher capacity drives, so at the moment I see "No known data errors"
This is a common claim, but honestly, citation needed.
This is based on the data I had available 10 years ago when I set up this array… If you have data that points to drives having fewer read errors per GB in recent years, I'd love to see that citation.
And how many times did you have a read error on one of your other disks during a resilver?
I don't know, because I have RAIDZ2, so that read error would probably be corrected without it being reported to me.
The risk of z1 is that if you get a read error during resilvering, that data is permanently corrupted. The odds of this happening go way up the larger your individual drives are. This is why I chose RAIDZ2 for my NAS. I've had to resilver 2-3 times over the past 10 years, and never lost a byte of data.
The amount of leverage AI tools give small interdisciplinary teams to apply their existing skill sets is kind of crazy.
This is 100% true, and this type of building is where AI is adding real value beyond the "I automated my entire business" hype.
Unfortunately, most companies don't have small interdisciplinary teams who have autonomy to scope and ship software. I spent 10 years at IDEO telling clients this was the way, but it's virtually impossible to replicate on the client side.
Yes, and this is especially true of enthusiast communities, which usually have evergreen topics. A user who is new to the Leica M system can head to rangefinderforum.com and get value out of lens reviews or camera comparisons that might be literally 20 years old.
Ah, so it’s impossible to make a robot unless it looks like an animal
Real question: what about 3 legs? Is tripedal locomotion a viable compromise?
“It’s a big step forward to this holy grail of making a living thing out of dead components,” said Sijbren Otto, a systems chemist at the Stratingh Institute for Chemistry in the Netherlands who was not involved in the work.
That is the holy grail? I get that the goal is to "grow" biofuels, plastic, fertilizer, drugs, or whatever else we can imagine. But is that worth the many apocalyptic sci-fi outcomes we can imagine?
The article is based on running Qwen 3.6 on a 128GB MacBook Pro. For reference, a 128GB MBP currently starts at $6699 USD [0]
Some people will be happy to pay that premium for privacy, but at roughly 10X the cost of a MacBook Neo, that money could also buy a lot of credits on OpenRouter or frontier labs.
[0] https://www.apple.com/shop/buy-mac/macbook-pro/14-inch-space...
People need to Google Jevons Paradox, smdh