remich

Yeah the rose colored glasses some people seem to have about how fun doing everything by hand was in the past and how well most people did those things continues to astonish me

It's also not $1,500 per month per engineer. It's that per month per engineer per tool. Which means it could easily be at least $3,000 (Claude Code and Cursor) or $4,500 if Codex was also an option on top of those two.

And as you have written on your blog it's a soft cap that can be exceeded with justification.

No reason to doubt his journalistic integrity? He's not a journalist for starters. He's a PR flack who does PR for AI startups on the side while blogging on substack. There is every reason to doubt his journalistic integrity.

Perhaps they aren't, but not currently viable !== always unviable.

Because it doesn't. Not for the tasks where using Opus instead of a lower tier model is appropriate, at any rate. Benchmarks show this, as do revealed preferences of actual users. To believe that Qwen is as capable as Opus at 1/20 the cost you have to believe that every person who does not make the choice to use Qwen over Opus for a given task is some mix of ignorant or delusional. This is certainly an opinion you can hold about other engineers, but it's definitely a questionable one at best.

Fiduciary duty but for AI, interesting. I think there's some potential there, though of course you'll end up confronting the classic sci-fi trope of "what if the system judges what's best for the user in a way that is unexpected / harmful"? But, solve that with strong guardrails and/or scoping and you might have something.

I'm starting to get to the point where I'll only listen to AI energy use critiques if the commentator tells me up front they abstain from all forms of social media, especially video-based social media, first.

Are they? Or do you just mean that it's few and far between that we hear about them? If it's the former, I think there's a much bigger universe of this kind of stuff than most people realize. Otoh, if you're just commenting on the lack of coverage, then, yeah I agree I wish more publicity was paid to small software like this. Maybe we need a catchy term - "organic software"? "Locally grown software"?

Yeah and then when that library stops being maintained or gets taken over, everything breaks.

We're in a transition phase, but this will shake out in the near future. In the non-professional space, poorly built vibecoded apps simply won't last, for any number of reasons. When it comes to professional devs, this is a problem that is solved by a combination of tooling, process, and management:

(1) Tooling to enable better evaluation of generated code and its adherence to conventions and norms (2) Process to impose requirements on the creation/exposure of PRDs/prompts/traces (3) Management to guide devs in the use of the above and to implement concrete rewards and consequences

Some organizations will be exposed as being deficient in some or all of these areas, and they will struggle. Better organizations will adapt.

I think this is an interesting point, my one area of disagreement is that there is no "anti-LLM sentiment" in the programming community. Sure, plenty of folks expressing skepticism or disagreement are doing so from a genuine place, but just in reading this site and a few email newsletters I get I can say that there is a non-trivial percent in the programming world who are adamantly opposed to LLMs/AI. When I see comments from people in that subset, it's quite clear that they aren't approaching it from a place of skepticism, where they could be convinced given appropriate evidence or experiences.

Get AWS to actually support pgvectorscale and timescaledb for RDS or Aurora and then maybe... sigh....

What if the instructions they give you would be to submit to them while they assaulted you, sexually or physically? Are you supposed to comply and then challenge them in court later?

That is a thing that happens. Rarely, I suppose, and #notallpolice and all that, but the idea that we should live in a country where everyone just has to "comply" with the instructions or be murdered is ridiculous.

Okay, I certainly agree with you that different use cases can dictate different outcomes when using AI tooling. I would just encourage everyone who thinks similar to you to be cautious about assuming that someone who experiences a different result with these tools is less skilled or dealing with a less difficult use case - like one that has no edge cases or has greater tolerance for bugs. It's possible that this is the case, but it is just as possible that they have found a way to work with these tools that produces excellent output.

Have you considered that it's a bit dismissive to assume that developers who find use out of AI tools necessarily approve of worse code than you do, or have lower standards?

It's fine to be a skeptic. Or to have tried out these tools and found that they do not work well for your particular use case at this moment in time. But you shouldn't assume that people who do get value out of them are not as good at the job as you are, or are dumber than you are, or slower than you are. That's just not a good practice and is also rude.

I have had a lot of success lately when working with Opus 4.5 using both the Beads task tracking system and the array of skills under the umbrella of Bad Dave's Robot Army. I don't have a link handy, but you should be able to find it on GitHub. I use the specialized skills for different review tasks (like Architecture Review, Performance Review, Security Review, etc.) on every completed task in addition to my own manual review, and I find that that helps to keep things from getting out of hand.

I don't think that anyone actually believes that writing code is only for junior developers. That seems to be a significant exaggeration at the very least. However, it is definitely true that most organizations of this size are hiring people into technical lead, staff engineer, or principal engineer roles are hiring those people not only for their individual expertise, or ability to apply that expertise themselves, but also for their ability to use that expertise as a force multiplier to make other less experienced people better at the craft.

As a former tech lead and now staff engineer, I definitely agree with this. I read a blog post a couple of months ago that theorized that the people that would adopt these technologies the best were people in the exact roles that you describe. I think because we were already used to having to rely on other people to execute on our plans and ideas because they were simply too big to accomplish by ourselves. Now that we have agents to do these things, it's not really all that different - although it is a different management style working around their limitations.

I get what you're saying, but I would say that this does not match my own experience. For me, prior to the agentic coding era, the problem was always that I had way more ideas for features, tools, or projects than I had the capacity to build when I had to confront the work of building everything by hand, also dealing with the inevitable difficulties in procrastination and getting started.

I am a very above-average engineer when it comes to speed at completing work well, whether that's typing speed or comprehension speed, and still these tools have felt like giving me a jetpack for my mind. I can get things done in weeks that would have taken me months before, and that opens up space to consider new areas that I wouldn't have even bothered exploring before because I would not have had the time to execute on them well.

Please do. I'm trying to help other devs in my company get more out of agentic coding, and I've noticed that not everyone is defaulting to Opus 4.5 or even Codex 5.2, and I'm not always able to give good examples to them for why they should. It would be great to have a blog post to point to…

Any particular papers or articles you've been reading that helped you devise this? Your experiments sound interesting and possibly relevant to what I'm doing.

Can't speak to Claude Code/Desktop, but any of the products that are VS Code forks have workspace restrictions on what folders they're allowed to access (for better and worse). Other products (like Warp terminal) that can give access to the whole filesystem come with pre-set strict deny/allow lists on what commands are allowed to be executed.

It's possible to remove some of these restrictions in these tools, or to operate with flags that skip permissions checks, but you have to intentionally do that.

Valid questions, but I'd say that it's hard to know what the future holds when we get models that push the state of the art every few months. Claude sonnet 3.7 was released in February of this year. At the rate of change we're going, I wouldn't be surprised if we end up with Sonnet 5 by March 2026.

As others have noted, Google's got a ways to go in making it easier to actually use their models, and though their recent releases have been impressive, it's not clear to me that the AI product category will remain free from the bad, old fiefdom culture that has doomed so many of their products over the last decade.

The way I always like to think about it is: "a computer shouldn't be able to do this."

I'm an SWE working in AI-related development so I have a probably higher baseline of understanding than most, but even I end up awed sometimes. For example, I was playing a video game the other night that had an annoying box sliding puzzle in it (you know, where you've got to move a piece to specific area but it's blocked by other pieces that you need to move in some order first). I struggled with it for way too long (because I missed a crucial detail), so for shits and giggles I decided to let ChatGPT have a go at it.

I took a photo of the initial game board on my tv and fed it into the high thinking version with a bit of text describing the desired outcome. ChatGPT was able to process the image and my text and after a few turns generated python code to solve it. It didn't come up with the solution, but that's because of the detail I missed that fundamentally changed the rules.

Anyway, I've been in the tech industry long enough that I have a pretty good idea of what should and shouldn't be possible with programs. It's absolutely wild to me that I was able to use a photo of a game board and like three sentences of text and end up with an accurate conclusion (that it was unsolvable based on the provided rules). There's so much more potential with these things than many people realize.

Is this something that can happen? We just ran into this limitation and I really want to keep using pgvectorscale... am exploring other solutions on EKS but RDS would be so much easier. From my reading it seems like this isn't something we can get done as a single AWS customer though.

This just made me think of that wall-mounted singing fish. Someone should make a cursed version that connects to Claude.

Right, words don't have semantic meaning on their own, that meaning is derived from surrounding context. "Cat" is both an animal and a bash command.

But, isn't improving tools and the LLM's integration with them improving the model?

Caveat that we don't fully understand how human intelligence works, but with humans it's generally true that skills are not static or siloed. Improving in one area can generate dividends in others. It's like how some professional football players improve their games by taking ballet lessons. Two very different skills, but the incorporation of one improves the other as well as the whole.

I would argue that narrowly focusing on LLM performance via benchmarks before tool use is incorporated is interesting, but not particularly relevant to whether they are transformative, or even useful, as products.

Having a large audience does not imply being the most well informed or correct.

The question isn't whether it makes a difference, the question is whether the model you're working with / the platform you're working with it on already does that. All of the major commercial models have their own system prompts that are quite detailed, and then the interfaces for using the models typically also have their own system prompts (Cursor, Claude Code, Codex, Warp, etc).

It's highly likely that if you're working with one of the commercial models that has been tuned for code tasks, in one of the commercial platforms that is marketed to SWEs, that instructions similar to the effect of "you're an expert/experienced engineer" will already be part of the context window.