DarkNova6

I'd say: A good abstraction is a formal and abstract statement about the nature of the underlying (sub)system. It is self consistent and allows an easier way to reason and interact with greater complexity underneath.

And the more you see an abstraction as a "wad of API calls and functionality", the more leaky the story gets.

Some systems are notoriously difficult to get this statement right because they are volatile and prone to changes. But I think as in live or in academia, our aspirations should be to strive for a better abstract understanding and keep building better abstractions.

The problem is primarily social. When code is cheap and entry barriers don't exist, how do you filter what is valuable?

For private projects, being able to code things quickly is a godsent. Also for professional environments with clear rules and mutual respect, this can be solved. But on the internet, with people you don't know? You must have some hard barriers.

Having basic defensive programming, some simple classes instead of dicts everywhere and avoiding n+1 is a "different world"?

You misunderstand. And that is exactly the problem.

We did do that and that's why our queries ended up being several lines long. But if you missed just one model? You openly walk a knife again.

It's a mess and it only gets longer and longer. I ended the project with having some proper aggregates, only for that to be thrown out of the window by the guy after me.

I seriously don't get why Django ORM is using the Active Record pattern. This is such a stupid footgun that trivially causes horrible performance and BEGS you to cause n+1 problems.

Never in my life did I have a problem with lazy loading causing unbearable performance until I joined a Python Django team. I really tried to find sympathy for the "dynamically typed" folks (please spare me saying Python is technically statically typed), but coming from writing apps and backends in Java, Swift, C#, Objective-C and PHP, Python with Django was the worst experience bar none.

I worked on the project for 10 months, could at least refactor the project to something semi-sane where obvious mistakes (which would not be possible in other languages) could not happen. Then along comes a "good Python dev" and threw it all out of the window and start doing SQL queries all over the place (typically 3-6 lines long), remove the domain objects and cause the same problems I started with to begin with. But his approach was saying that the other developers were "not good enough".

Yeah, have fun with schema changes going forward. Good riddance.

I'm referring to the formal usage of the term "distillation", instead of the informal one that refers to training it on any data as a whole.

https://www.geeksforgeeks.org/nlp/what-is-llm-distillation/

You mean the asian models which just distilled American ones? I'm happy Mistral is doing their own ground up research. SOTA frontier models are a commodity with little room for second places.

Mistral is playing the smart money on vertical products rather than horizontal ones. The former requires finesse, the latter brute strength.

I don't understand this obsession with clear cut categorization. Systems always have a human component and while we like to collectively deny this fact from a tech-sphere, there are always humans that will make the final call and if anything we should encourage the use of human trust and judgement.

We have focused so much on technological advancement, but our understanding of human systems and building an environment of trust that is not eroded by perverse incentives atrophies.

I don't get it. I thought the advantage is that you have a client as a library with integrated auth & other tech stuff handled so you can just focus on using the API? Because otherwise you could just use REST with basic endpoint code.

This sounds like a more elaborate version of Whataboutism.

I'd like to think that it's twice as good as a one-party system.

Indeed. The true irony behind the Dunning Kruger Effect is that the pop culture understanding of it is essentially describing itself. Showing overconfidence in areas that you have little clue about.

I think it keeps going back to "preserving type invariants". Turning an object into a value can have surprising constraints on the integrity of the type. These include: forced empty constructor (for bulk initialiation in arrays), violation of cross-field invariance (think of a Range-type where "end" value is suddenly smaller than the "start" due to a data race), etc.

Java always had a focus on enabling library authors over empowering call-site users. And I think approach has proven its dividend.

I left Swift and SwiftUi behind be 5 years ago when I switched jobs. I had fond memories and always eyed going back.

What were the emerging problems that you talk about? I am actually curious because I was quite fond of the DevEx

Right? It strikes me that the vast majority of people complaining genuinely don't understand the difference between law and code, content and art or can think of anything beyond what they are facing in day to day work realities.

This being downvoted speaks saddens my heart... and proves the exact intent of the message.

Thanks, now I don't need to even read the article!

You are not wrong, but here's the catch:

"Even I - with all my years of experience and knowledge - fell into the trap of hallucinations. I summarised reports using AI tools and worked from those summaries, trusting they were accurate.

He used an LLM to perform research instead of using it as a mere writing tool. I think large language models are great at formatting text, formatting, structuring and giving a stylecheck. But the last pass and review must always be a human. Everything posted by an LLM under your name is still under your name and you are liable for every single word. It's about the outcome and everybody who uses it is responsible for it.

I'd argue that if a satire or parody is strong enough having a small watermark will not detract from its statement.

You completely disregard practical implications.

Can you manually create images that were done via AI? Of course you can.

Can you create AI fake content and misinformation as easily with human labor as you can with AI? No.

Understanding this distinction answers whether you see the implications AI has on the future of democracies or not.

If anything this an argument _for_ clearly labelling manipulated photos and AI generated depiction of supposedly real life events.

You will have to tell me how "labelling machine additions" is applicable at all as an example here. Unless your argument boils down to "computer doing things does not require labelling".

So what are journalistic texts if not docs? At this point, an LLM is just another evolution from the typewriter.

You are still responsible for your work. What is important are the semantics and statements of the text. Less so how the exact arrangement of letters came to be.

More authentic because of what, a missing watermark? It still has 2 positive outcomes:

1: It gives legal leverage against vectors of misinformation. Misinformation is tricky to combat legally and this is a first obvious measure that can be taken.

2: Reputation will become more important in the future, and if you are found to have repeatedly posted AI content claiming it is not it will damage you. Not to mention that users can point out the obvious discrepancy.

I don't see how these are contradictory. The battle for reality will only worsen and you likely need multiple failsaves.

The non thinking ones have voting power too and they will burn this shithole down because they think some powerful people in nice suits act on good faith and should be given all the decisionmaking and resources.

This way of thinking is a very efficient way of destroying your culture and civilisation.

There's creativity and there's the creative process. I like to dictate documentation in long form monologue, and then give instructions how to structure it. Then I give it a final pass with manual adjustments and the end-result is exactly how it would have looked if I had written it from scratch but without the bsywork.

Don't be stupid enough to let machines do your own thinking, treat them as a dishwasher.

An obvious start would be that any such technology needs to have its own clearly visible watermark and removing it constitutes a crime.

I'm surprised by the sentiment.

With enough money you can already buy attention and presence. Now you can even facilitate a fake discourse to claim legitimacy and steer manipulation into entirely new realms. If this is not counteracted not only will rogue nations destroy democracies even faster, but powerful corporations and rich individuals will also use this to increase their power further and damage simple consumers.

Nothing about this concerns any creative process.

> It won't work.

Maybe. But the rest of your post doesn't have anything to back this up.

> I'm also curious, what about AI assisted media? If a company writes a full article, then uses AI to edit it a bit, does that need to be labeled?

Why would it? The threat is mass generation of articles without any human oversights, and realistic looking images claiming occurrence of events that never hapened. Nobody cares about an AI typo and stylecheck.