Just a set of things too minor to move off of it but annoying enough to not want to start with it.
My list:
No `explain (analyze,buffers)`. Instant DDL has some warts (e.g. fk, metadata locks). Query planning bugs (actually... query planning in general is disappointing). Exiting the repl doesn't stop queries. Implicit type casting. Replication lag from large DDL (e.g. creating an index). Lack of two phase DDL (creating constraints NOT VALID and then VALIDATE later). Lack of extensions (e.g. pg_vector). No safe access to inspect buffer cache. AWS Aurora seems to only add shiny new things to Postgres. And more.
Again, none of this is quite enough to migrate off of it for an established system, but certainly enough to avoid it on a new project.
> This is the number one thing I optimize for now: the ability to quickly and safely change significant parts of the code and product.
This was always a good thing. Its value has nothing to do with the advent of AI coding.
> The opportunity cost of not being able to safely restructure has gone up substantially.
This bit is contradictory with everything else you said. Prior to AI coding it would take a lot longer to perform restructures. If anything, the thing you're now optimising for has gone down in value. It's still valuable, but perhaps a little less.
But in seriousness, this intuitively feels like something (as phrased) that would be easily influenced by loud noise and quantity rather than hard facts. The "piloted poorly" part is applicable to any tool use. AI is no different there other than its adoption rate.
The pattern I observe is: "write code, write test, make things green"
This is different in two ways from the classic TDD red-green-refactor suggestion:
1) they don't start with the test first, so the tests that get implemented are after writing the code, and run the risk of the model attention being now more influenced by the just-written code than the original spec
2) they finish when everything is green and don't followup with the "refactor" step unless manually prompted (either directly or indirectly by your own scaffolding/rules/whatever). this results in frequent hyper-local non-ideal-longterm fixes for things that went wrong in the first shot at writing the code pre-test-writing.
(As always, the only person who can ensure the code landing in your repo is good is you.)
But here's the rub - if you want your clanker to do those steps, it's usually a simple matter of adding them to your AGENTS.md and then it always does them.
I'm a big fan of the characterization step step being added. And it can be reasonable to add this before or even after the fact as a commit prior to your actual commit (assuming you're familiar with using tools where that's easy to do - e.g. jj or git with rebase). And the agents can do this - they just don't tend to without you saying to do so.
A lot of engineering practice comes from choosing which elements are reasonable to use given the context of what you're building. Providing that is your job. When you do that poorly, you get poor results. But garbage in garbage out has always been a thing. Any advanced automation amplifies ambient assumptions
I'm not talking about time. I'm talking about safety. The amount of times I've seen "I refactored it, but I'm not confident enough to take it to prod" is significant. Being able to go faster but still not ship it is the huge opportunity cost.
Time, safety and cost are one and the same. Not safe enough? Spend more time increasing confidence. Taking too long? Cheap out now and pay the price later due to added risk.
All of that is orthogonal to AI. All AI did was accelerate the typing code part - which was never the bottleneck or a very significant cost to begin with.
In the broader corporate world, that's not "opportunity cost". All changes are considered "risk".
All deployments must be approved by an advisory board. All work must originate from a clear business need. Analysis of those needs is not concerned with implementation, least of which whether "AI" is used.
What matters far more is that a contract requires work to be done by a deadline. Those deadlines are driven by policy. There will be no adjustment to policy unless tangible benefits are shown from more frequent deployments of code.
I gotta tell you that's extremely unlikely if you're already shipping every other week at the end of the sprint. Most of that sprint is spent in meetings, not writing code. Nobody is doing big refactors because it wasn't built so fast to require them. There's some technical debt, but nothing so severe. Those meetings are preventing risk, not wasting time. The bottleneck is a feature, not a bug.
If you think the future of dev work is to be a bureaucrat, you're right! It looks like the rest of the world outside of SV is ahead of the curve and living in the future.
I mean "We can't build X because our code structure makes that difficult" has an opportunity cost of the value of X.
I don't think the future of dev work is being a bureaucrat. I've done more rigorous engineering the last two years than I did previously. I'm more confident in the things I shop and they were built in a fraction of the time. It's a bright future for software engineering.
> The cost of shoring up behavior with tests ahead of a restructure has gone down because of AI.
Disagree. The growth in brittle AI-generated tests means restructuring is more costly than it was before. Pruning your test suite so that it tests the essence of the problem and not the incidental design decisions is something AIs aren't yet capable of.
Oh, it's worse than that. I've seen a rise in mutation testing, intended to ensure any change in implementation is caught by tests. Think about that for a moment. It's giving a fancy name to creating brittle tests than fail if any line of code is changed.
And this is seen as a good thing, because LLMs are really bad at confining their changes appropriately. Testing is really in a dark place right now.
> The cost of shoring up behavior with tests ahead of a restructure has gone down because of AI.
Yes.
But the ease of not doing that and instead just getting a brittle set of three-quarters-baked tests is extremely high! And many people seem happy to go from "a few human-written mediocre brittle tests" to "a bunch of AI-written mediocre brittle tests" because it is an objective improvement and the people who weren't avoiding speculative structure and looking for the write boundaries before are happy to also not do so no.
So completely agree with the "take advantage of the tools this way" but I also wouldn't claim it's a reason to no longer worry about if you're building the wrong castles in the sky too early, because perfect refactor-proof testing contracts are still usually pretty hard to design.
> perfect refactor-proof testing contracts are still usually pretty hard to design
Not sure about hard, but definitely rare and we as an industry are under-skilled in these areas.
We have decades of research and tools for testing and verification of software. Property tests, dependent types, formal verification and proof, etc. The paths have been there, we have just collectively prioritized other things.
It requires an intentional shift in how we design and build. That shift is the harder part.
> the ability to quickly and safely change significant parts of the code and product.
Hum, this reminds me something... "O: open to extension, closed to modifications". Old things are new again.
From context efficiency with approaches such ad DDD and clean architectures, all the way to items such as this one, AI is not creating new tradeoff, it just acts as a multiplier, multiplying productivity for teams doing things right, and multiplying debt for teams having a low quality bar as far as design and architecture are concerned.
Your increasing thrash betting that AI will fix it for you. The only thing your getting in return is not having to think that hard.
It doesnt cost that much time or effort to think hard, so you will be outcompeted by people levergaing AI as much as you, but thinking enough to not have it be thrashing around
> The cost of implementing a zero downtime migration has gone down because of AI.
You either don't know what that technical term means or you're just wrong. AI does not meaningfully move the needle on that. It only makes backwards compatible deployments easier insofar you're able to do the overhead for splitting the change with less effort then before.
I've always known how to do zero downtime migrations. The question has always been "Is the engineering cost worth it?"
> It only makes backwards compatible deployments easier insofar you're able to do the overhead for splitting the change with less effort then before.
Yes. That reduces the cost of implementing them.
To be clear, I'm not talking about "Split my db migration and my code that depends on the new table". I'm talking about things like "Set up dual writes between an old database schema and a new database schema with a thorough test suite and do shadow reads against both datasets in prod to do differential testing". That's nontrivial engineering effort that would definitely warrant a discussion in the past. Today we just do it. It's fast and lowers risk.
Generation is basically just memory bandwidth math.
Each token has to read all the active weights. I think that's around 40B parameters active. At a 4-bit quant that's 20GB. With 100GB/s (replace with whatever your bandwidth is) and you get 5 tokens per second.
MTP on a MoE is hit or miss. If you're bottlenecked on memory, MTP can increase the number of active experts (like any batch processing would), which can eat away gains from it.
My favorite lens on SQLite is that it is actually two things:
1. A robust durability implementation
2. A library of high performance data structure and algorithms
The fact this it's SQL is nice, but those two attributes are what make it great.
For example, I'm implement an in-process event log that I want to be durable. I started simple, but soon saw some edge cases and instead of playing whackamole I just swapped to using sqlite as an ordered kv store that gives me ACID.
Another example: ingesting multiple inter related datasets. Instead of a dozen hash maps in memory, I load them up into sqlite (no persistence) and then slice and dice as I need to.
I wonder how Sonnet vs Opus stacks up in a similar time-based comparison.
How far behind are open models compared to Sonnet?
It may be that the absolute SOTA models are way ahead of open models, but the gap in the mid tier really does feel like it's compressing. I'd love to see empirical data about it though.
I'm running opencode with qwen3.6-35b-a3b at a 3-bit quant. I also have qwen3.5-0.8b used for context compaction. I run with 128k context.
It's usable. I set it loose on the postgres codebase, told it to find or build a performance benchmark for the bloom filter index and then identify a performance improvement. It took a long time (overnight), but eventually presented an alternate hashing algorithm with experimental data on false positive rate, insertion speed and lookup speed. There wasn't a clear winner, but it was a reasonable find with rigorous data.
Long agent runs make such a difference. We focus a lot on new models and long context, but the bigger impact is in automatic verification.
I've been leaning in more on e2e test suites. They are slow, brittle and inefficient. But that's almost a feature. I can step away and come back an hour later, and use that time to think about bigger problems.
Thank you, I just realised we are talking about MTP. It seems that it's not that clear though.
"Currently, the MTP capabilities are primarily accessible through Google's proprietary LiteRT framework, rather than the open-weights versions... Despite the missing MTP heads in the open release, Gemma 4 (specifically the 26B-A4B variant) still demonstrates high efficiency"
Since it's a procedural language, you can't do things like create a new index implementation or something else super low level. But there's still a lot you _can_ do. Like implement a custom comparator for a custom type and then use that type in a btree index.
My list:
No `explain (analyze,buffers)`. Instant DDL has some warts (e.g. fk, metadata locks). Query planning bugs (actually... query planning in general is disappointing). Exiting the repl doesn't stop queries. Implicit type casting. Replication lag from large DDL (e.g. creating an index). Lack of two phase DDL (creating constraints NOT VALID and then VALIDATE later). Lack of extensions (e.g. pg_vector). No safe access to inspect buffer cache. AWS Aurora seems to only add shiny new things to Postgres. And more.
Again, none of this is quite enough to migrate off of it for an established system, but certainly enough to avoid it on a new project.
reply