Finance and investment banking is about resource allocation. To get the stuff like agriculture innovation etc done, you need money. You need someone to look and say, this is worth investing money in. This is far from trivial. The investment firms try to predict what technology or company is going to make something profitable. This allocation task is very important. There are many ways to fail at producing anything useful. Intent purity is not sufficient.
Of course the suspicion towards this is eternal. People always hated the traders as opposed to the farmers. But trade is crucial. And it relies on estimating what value things have and rewards correcting over the incorrect beliefs of uninformed people. This kind of information based knowledge work was always disliked by most people as it seems lazy. And for sure there are zero sum and rent seeking aspects or insider trading etc. But it's not so simple as to say that all investment and finance jobs are negative and working on farm efficiency is always better.
I don’t hate traders, I just wonder about Jeffrey Skilling graduating Harvard at the top of his class. In aggregate, is it too risky to need so many like him in finance?
To what extent were they correct and to what extent not? Is their correctness also linked to the correctness of the similar argument today or you're just noting the analogy?
I guess some of the appeal of those sparse photos is the element of fantasy and imagination. Wondering what it could have been. Looking at a low quality yellowing wedding photo of your grandma... It allows you to think and wonder. Seeing it in 4K video or a volumetric 4D gaussian splat in VR robs you of all that sentimental mystery.
Nostalgia and idealization of the past is also harder when you have a more representative cross section of past moments.
They are heavily dogfooding. Coding is needed to orchestrate the training of the next Claude model, data processing, RL environments, evals, scaffolding, UI, APIs, automated experiments, cluster management, etc etc. This allows them to get the next model faster and then get the next one etc.
Making a model that's great at other kinds of knowledge/office work is coincidental, it doesn't feed back directly into improving the model.
Name recognition has big value. People remember what an advancement the first banana was. Nowadays it's no longer so unique, ChatGPT's and Grok's image editors are also strong.
All things traditionally done by artists or artist adjacent roles. I can understand at an individual level, say for a solo gamedev who wasn’t going to pay an artist anyway. That’s not at scale though.
Larian Studios most recently was under fire for this [1]. Like I can see a director going “what would X look like?” and then speeding over to the concept artists for a proper rendition if they liked it. I don’t think this is at scale though. Any large business is just going to get rid of the concept artists.
There are many places in general office work where you need some kind of graphics. Slides, reports, info graphics, dataviz. Or academic papers. Some are just illustrations, like a fancy clipart or stock photos, some are drafts for a proper tikz or svg or something that you then redo in draw.io etc. There is much more use for graphics than the use cases where people would ever even consider hiring an actual artist. I've seen good results for iterating on eg model architecture figures quickly between PhD students and supervisors, faster than dragging boxes around and fiddling with tikz. Obviously you don't simply paste the result into the paper. You redo it but it's a good discussion basis. That's for info graphics stuff. But the same can apply to creative stuff, like an event poster, an invitation card to your wedding, storyboards, mood boards, DIY interior design, outfit planning etc etc
They initially wanted to call it just "Gemini 2.5 Flash Image (preview)" but the Internet stuck with the anonymous codename Nano-banana from LMArena because it's interesting and quirky. Google didn't officially adopt it until several days after the public release, exactly because of what you say. Eventually, not using it in their comms got more confusing because regular people were asking how they can find this Nano banana thing everyone is hyped about.
I think they are betting that any of this code is transient and not worth too much effort because once Opus 5 is traimed, they can just ask it to refactor and fix everything and improve code quality enough so that things don't fall apart while adding more features, and when opus 5.5 comes out it will be able to clean up after opus 5. And so on. They don't expect these codebase to be long lived and worth the time investment.
I see it on myself too. It feels too irresistible to start adding more features to software you develop with LLM agents. Everything feels like just a few prompts and will be done in half an hour. Why not add this too? Just another sentence in the prompt. Next thing you know you have more features than you remember and the AI starts to have a really hard job keeping it all functional.
Coding with AI requires immense restraint and strong scope limits.
Of course the suspicion towards this is eternal. People always hated the traders as opposed to the farmers. But trade is crucial. And it relies on estimating what value things have and rewards correcting over the incorrect beliefs of uninformed people. This kind of information based knowledge work was always disliked by most people as it seems lazy. And for sure there are zero sum and rent seeking aspects or insider trading etc. But it's not so simple as to say that all investment and finance jobs are negative and working on farm efficiency is always better.
reply