refactor_master

I think it’s relevant to first read [1] to see why they’re doing this. It’s basically done as a meme.

[1] https://motherduck.com/blog/big-data-is-dead/

If A vibes, and B is overwhelmed with noise, how does B reliably go through it? If using AI, this necessarily faces the same problems that recording all A's actions was trying to solve in the first place, and we'd be stuck in a never-ending cycle.

We could also distribute the task to B, C, D, ... N actors, and assume that each of them would "cover" (i.e. understand) some part of A's output. But this suddenly becomes very labor intensive for other reasons, such as coordination and trust that all the reviewers cover adequately within the given time...

Or we could tell A that this is not a vibe playground and fire them.

Here’s a talk about leaning into the garbage flow. And that was a decade ago.

https://youtu.be/E8Lhqri8tZk

I can’t imagine the number being economically meaningful now.

Yes, even ferries in Scandinavia can be roads when it makes for a better map… Or not ferries in the Mediterranean when it doesn’t.

Honestly I think this part

> The resulting images bring insights into the ways in which road infrastructure reflect regional, political and geographical situations.

should just be taken as pseudo-profound art fluff. Are you telling me all Greek commercial transport goes straight through non-Schengen instead of just ferrying across?

Yes, that seems to hold for rocks. But that doesn’t shut down the original post’s premise, unless you hold the answer to what can and cannot be banged together to create emergent intelligence.

I think it's meant to be taken more in the abstract. Yes, LLM can refuse your request, and yes you can ask it to prepend "have you checked that it already exists?", but it can't directly challenge your super long-range assumptions the same way as another person saying at the standup "this unrelated feature already does something similar, so maybe you can modify it to accomplish both the original goal and your goal", or they might say "we have this feature coming up, which will solve your goal". Without proper alignment you're just churning out duplicate code at a faster rate now.

It’s interesting then to ask if this will behave the same as big orgs? Eg once your org is big and settled, anything but the core product and adjacent services become impossible, which is why 23 often see a 50-person company out-innovating a 5k person company in tech (only to be bought up and dismantled, of course, but that’s besides this point).

Will agents simply dig the trenches deeper towards the direction of the best existing tests, and does it take a human to turn off the agent noise and write code manually for a new, innovative direction?

But the code you’re writing is guard railed by your oversight, the tests you decide on and the type checking.

So whether you’re writing the spec code out by hand or ask an LLM to do it is besides the point if the code is considered a means to an end, which is what the post above yours was getting at.

Maybe I was stupid or maybe it just doesn’t hit the same way if you don’t grow up in the US, but I remember not being terribly fascinated by it as a 90s kid. In fact, I found it kind of uncanny that the world felt so… disconnected. I later learned this was called “modernist architecture”.

Hell, my bed is on the floor, and my sofa is now also a pillow on the floor.

I still get made-up Python types all the time with Gemini. Really quite distracting when your codebase is massive and triggers a type error, and Gemini says

"To solve it you just need to use WrongType[ThisCannotBeUsedHere[Object]]"

and then I spend 15 minutes running in circles, because everything from there on is just a downward spiral, until I shut off the AI noise and just read the docs.

I think this is a classic old-vs-new tale. I started my PhD in biochemical research where analyzing data by hand was definitely a "craft" in some aspects. Later I forewent going to the lab entirely and instead spent all my time on developing machine learning for automated data analysis. But just like field work, you still need people in labs who can continue the craft.

The article should perhaps introspect a bit more instead of setting up a false dichotomy between "rainforest field work or computers".

But that’s more of a theoretical truth than a practical one, isn’t it? High quality novels are easily found. TikTok videos of equally high quality and depth? Perhaps not so, or exceedingly rarely.

Infinite monkeys with infinite time could surely also produce something spectacular and eye-opening, statistically speaking. But umm, you’d have to wait infinite time for it to be done, so it’s not really efficient when time is a finite resource.

So now every fraudster with $5 appears legitimate?

Remember blue check marks? The EU is not happy about those.

https://ec.europa.eu/commission/presscorner/detail/en/ip_25_...

I’m interested in earnings correlating with feature releases. Maybe you’re pushing 100% more bugs, but if you can sell twice as many buggy features as your neighbor at the same time, it could be that you could land more contracts.

It’s definitely a raise to the bottom scenario, but that was already the scenario we lived in before LLMs.

I second the persistence. Some of the most persistent code we own is because it’s untested and poorly written, but managed to become critical infrastructure early on. Most new tests are best-effort black box tests and guesswork, since the creators have left a long time ago.

Of course, feeding the code to an LLM makes it really go to town. And break every test in the process. Then you start babying it to do smaller and smaller changes, but at that point it’s faster to just do it manually.

The crusade against gluten probably did it. Tofu lives as un-refrigerated grey blobs and tempeh never even made it to the shelf, probably because of hormone-disrupting soybeans. But hyper-engineered single cell meat? Now that’ll sell.

Gemini routinely makes up stuff about BigQuery’s workings. “It’s poorly documented”. Well, read the open source code, reason it out.

Makes you wonder what 97% is worth. Would we accept a different service with only 97% availability, and all downtime during lunch break?

I think group 3 is a bit of a reach. Most people just treat it as a commodity. You need a break after shopping? Coffee. Meeting someone to talk over something for 30 minutes? Coffee. Need a cozy place to sit and get some work done? Coffee. For none of these do people have to engage with the community or be caffeine addicts.

> We shall be able to rid ourselves of many of the pseudo-moral principles which have hag-ridden us for two hundred years

Still haven’t gotten rid of work for work’s sake being a virtue, which explains everything else. Welfare? You don’t “deserve” it. Until we solve this problem, we’re not or less heading straight for feudalism.

> Go to any Asian country and houses, healthcare and cars are priced like commodities, not luxuries.

What do you mean? Several Asian cities have housing crises far worse than the US in local purchasing power, and I'd even argue that a "cheap" home in many Asian countries is going to be of a far lower quality than a "cheap" home in the US.

If you've ever read a blog on trading when LSTMs came out, you'd have seen all sorts of weird stuff with predicting the price at t+1 on a very bad train/test split, where the author would usually say "it predicts t+1 with 99% accuracy compared to t", and the graph would be an exact copy with a t+1 offset.

So eye-balling the graph looks great, almost perfect even, until you realize that in real-time the model would've predicted yesterday's high on today's market crash and you'd have lost everything.

Originally I used Pandera, but it had several issues last

* Mypy dependency and really bad PEP compliance * Sub-optimal runtime check decorators * Subclasses pd.DataFrame, so using e.g. .assign(...) makes the type checker think it's still the same type, but now you just violated your own schema

So I wrote my own library that solves all these issues, but it's currently company-internal. I've been meaning to push for open-sourcing it, but just haven't had the time.

Should have done GME stocks only. Now THAT would’ve been interesting to see how much they’d end up losing on that.

Just riding a bubble up for 8 months with no consequences is not an indicator of anything.

I honestly don't get why you'd hate pandas more than anything else in the Python ecosystem. It's probably not the best tool in the world, and sure, like everybody else I'd rewrite the universe in Rust if I could start over, and had infinite time to catch up.

But the code base I work on has thousands and THOUSANDS of lines of Pandas churning through big data, and I can't remember the last time it lead to a bug or error in production.

We use pandas + static schema wrapper + type checker, so you'll have to get exotic to break things.

There's been a translate button for years which hooks deep into every nook and cranny of the website's HTML. It works great, it's built in and many restaurants even advertise it for tourists, because it's a zero-effort translation of their existing menu. Plus, it's low-data when you're inside a 1-bar basement restaurant.

Using an LLM to translate the visible part of a PDF on a mobile... seems like the worst possible solution to the problem.

Actually, big orgs are doing exactly that; slapping a chatbot onto their support ticket backlog. Being really, actually “data driven” is hard, and must happen from the bottom up. So instead there’s chatbots in their frontend and support backend, but the backend doing the actual lifting probably hasn’t changed one bit.

I wonder how many market inefficiencies this creates. People with worse education, people who cheated their way to a job opening compared to a better candidate, etc. Basically counteracting the productivity gains AI was supposed to bring.

> There were some UI glitches

Interesting that a next-gen open-source-based agentic coding platform with superhuman coding models behind it can have UI glitches. Very interesting that even the website itself is kind of sluggish. Surely, someone, somewhere must have ever optimized something related to UI rendering, such that a model could learn from it.

IO and global state is also just inputs that can be part of arrange-act-assert. Instead of mocking your database call to always return "foo" when the word "SELECT" is in the query, insert a real "foo" in a real test database and perform a real query.

Again I've heard "but what if my database/table changes so rapidly that I need the mock so I don't need to change the query all the time", in which case you ought to take a moment to write down what you're trying to accomplish, rather than using mocks to pave over poor architectural decisions. Eventually, the query fails and the mock succeeds, because they were completely unrelated.

So far I've only seen mocks fail eventually and mysteriously. With setups and DI you can treat things mostly as a black box from a testing point of view, but when mocks are involved you need surgical precision to hit the right target at the right time.