Ivan Tanev

I write code for fun and profit.

[Mitchell Hashimoto @mitchellh]

The problem with the "if it works who cares what the code looks like" mindset for agentic work is that it assumes the agent has a perfect understanding of "works." Realistically, things are underspecified, agents make bad assumptions, etc.

To be fair, agents are pretty good at unit test coverage. They're pretty bad at designing human experiences (API, CLI flags, etc.), especially cohesive ones for future roadmap plans they may not have visibility into (unless your backlog is perfect and vision fully laid out, which I doubt). They're bad at knowing where performance matters and what type (CPU vs memory tradeoffs). They're bad at where compatibility matters and where it doesn't (and tend to err on the side of preserving it without further guidance). Etc.

Unless you have this ALL specified, you can't possibly claim "it works" without taking a look and thinking about it.

[Ivan Tanev @VanTanev]

@davis7 What does Effect add here? You're paying the full effect runtime cost to run a single promise, without augmenting the error channel in any way, so it's equivalent to a normal promise rejection.

Removing Effect.runPromise() from that code is strictly better.

[Ivan Tanev @VanTanev]

I think it really depends on your engagement with the art community. In different fandom or niche art circles, I see AI art being constantly used for, frankly, scamming people. Mostly in the form of taking commissions or creating patreons.

The algorithm really rewards consistent posting, and most real artists cannot create a new piece every day, while that's mostly trivial for AI, even with advanced prompting. So you get a loop where the algorithm promotes new consistent artists that have a """good""" style and do not disclose AI.

It is a real menace for small semi-pro and professional artists that rely on commissions. For small artists, the only way to build a following is through social media, and it's being crowded out by scammers.

Back to my original point about your examples - if you label your stuff as AI, than I think it's mostly fine! Then at least the audience can make an informed choice that aligns with their morality about the issue. You can even make the argument of letting your limited artist resource focus on the most important parts.

However, that argument only works if you're truthful about what you used.

[Ivan Tanev @VanTanev]

Leaving aside the intellectual theft angle for a moment, it's also a question of assigned value. There is a good reason that AI 'artists' lie about using AI - because they want recognition equal to someone who put in the effort to achieve mastery without putting that effort in.

For creative fields, the scorn for anything AI (that is not explicitly labeled as such) is deserved. With your collage example, usually it's clear from the style that it's a collage, and you can appreciate the artistry on that level. Selecting the images yourself and arranging them in a way that gives new meaning. Meanwhile, with AI art, the output is usually explicitly prompted in a way that imitates genuine human output.

There is a gray zone where people mix the two - some human art with eg generated backgrounds, or ai enhanced shading or something of the sort. That can be more defensible as being closer to collage art, though labeling probably should still be applied.

[Ivan Tanev @VanTanev]

@realChakrawarti @kettanaito @joelhooks Thank you, that is genuinely encouraging to hear!

[Ivan Tanev @VanTanev]

@joelhooks @kettanaito But I appreciate the link still seems like a book worth reading!

[Ivan Tanev @VanTanev]

@joelhooks @kettanaito Plot twist: It's fiction :D

[Ivan Tanev @VanTanev]

@kettanaito @joelhooks I will release both, and whatever happens, happens.

[Ivan Tanev @VanTanev]

@kettanaito @joelhooks I have recorded a 5 hour video course on monorepos and typescript. I'm kinda scared to release it because it'll likely be ignored. Two years of work.

Now I'm a year into writing a book. I don't know why I do this to myself, but ain't stopping 😂

[Ivan Tanev @VanTanev]

@kalomaze Where does deepseek v4 pro stand?

[Ivan Tanev @VanTanev]

@RichardHanania Would you like to put a wager on if everything will be fine in 2 months if Hormuz remains closed?

[Ivan Tanev @VanTanev]

@nordmarj129692 @hecubian_devil Microsoft is a public company that could disclose this revenue breakdown but intentionally completely hides all AI revenue under "Cloud."

Make of that what you will.

[Ivan Tanev @VanTanev]

@libpol_org @hecubian_devil Don't necessarily disagree with the rest of what you wrote, but 5x demand and revenue leaves them squarely bankrupt.

[Ivan Tanev @VanTanev]

@MattBGilliland @hecubian_devil All true.
How do you square that with none of the big public companies making their AI revenues public?

[Ivan Tanev @VanTanev]

Wrong. Microsoft, Google, and Amazon are public companies, but they deliberately obfuscate AI revenue.

Microsoft in particular is in a position where they both have a license for reselling OpenAI models and they provide inference, but for some weird reasons omit a breakdown of AI-related revenue in their earnings.

[Ivan Tanev @VanTanev]

@Raikaru00 @hecubian_devil The vast majority of organizations already have a github enterprise subscription.

[Ivan Tanev @VanTanev]

@Raikaru00 @hecubian_devil There are many organizations that use github Copilot for agentic work, because it's easier to fold spend into an existing subscription rather than setup a new one. If you've worked in large organizations you should know that.

[Ivan Tanev @VanTanev]

@Raikaru00 @hecubian_devil Github Copilot had access to opus and gpt coding agents on "number of messages" basis instead of number of tokens, for enterprise coding uses.

Seems to me it's you who might not know how this works?

[Ivan Tanev @VanTanev]

That's the neat thing - for some reason, no one wants to report the actual numbers. So we don't know what the breakdown is.

But github's Copilot billing is now changing significantly because they did subsidize for enterprise uses, to the tune of about 9x. So that's one data point.

[Ivan Tanev @VanTanev]

@Raikaru00 @hecubian_devil Of course they do 😂
This is why github is changing their entire billing model for LLM usage through Copilot come July.