Adam T. Croft
27 posts

Adam T. Croft
@adamtcroft
Technical Audio Lead @Bungie. I help demystify bleeding-edge audio and AI tech for sound designers, musicians, and game developers.
Katılım Ekim 2023
47 Takip Edilen2 Takipçiler
Adam T. Croft retweetledi

This is utterly fascinating. We’ve known code written by AI is harder to untangle. It appears this is the case with writing as well.
Tricia is an editor and she says that when an author submits work that is written by AI, she has a much harder time editing it. It’s all one interconnected black-box piece of writing that is not amenable to change. Whereas she finds that human writing, while seemingly messier, is actually much more structurally straightforward.
My theory as to why this is is that LLMs think one token at a time. And after every token, they essentially look back and ask, “have I said the thing the prompt wants me to say?” If not, it keeps elucidating.
The result is tight chain of thought writing that requires each preceding token to make sense of the next.
Whereas human writing starts from a pre-language idea in the author’s head, and looks forward many sentences and paragraphs ahead to approximate the author’s intent.
It’s somewhat fuzzy. But I think LLMs fundamentally “think” in a much different way than humans. They are certainly not useless. But I think it’s a grave mistake to equate them with human intelligence.
Tricia Dearborn@TriciaDearborn
If you're thinking about using gen-AI to "write" books, this 🧵 is for you. I’m a highly experienced editor who’s been in the biz a long time. Recently I’ve had manuscripts come to me where the author has used gen-AI – not for writing, I’ve been assured, but for
English

@sonicbarber @MarathonTheGame Get Claude to write your presentation for you while you play Marathon (kidding, kidding)
English

Finish writing presentation on @MarathonTheGame or play @MarathonTheGame ? Hmmmmmm
English

I don't know what this says about me - but my current relaxation past time is blasting music (tonight is @Pitchshifter , last week was ODB) and playing Crime Scene Cleaner.
Other people read books, I digitally clean crime scenes 🤷
English

@wookash_podcast By the logic "licensed IP == stay away" I'd need an explanation on his stance for Goldeneye, NBA JAM, PS4/PS5 Spiderman, lots of Star Wars and LEGO games...
I didn't play the old HP titles, but still 🤔 at that take.
English

My dear friend has mocked me for liking old Harry Potter games (1 & 2, though 3 is ok as well). Were these games any good, or is it just nostalgia?
His stance is "if it wasn't well-known IP you wouldn't like it". He generally stays away from games that licensed IP - HP, LOTR, Warhammer, etc. Cause he wants to play games for **their core game ideas**, not to play "reskins" of popular game mechanics with known characters.
But I'm like, bro, Harry Potter 2, with its all secrets, silver cards, quidditch, was crazy good! Had so much mystery, the magic was basic, yet it was exactly what 7/8/15 year old me wanted.
Does anyone remember these games, and being older at that time, can wage in, and offer their take - were they plain games with "stolen" / "adapted" mechanics from other works, or were there original games, that happened to license popular IP?
English


Marathon, our new extraction shooter @bungie is free to play for this weekend’s server slam starting Thursday 💪
More details: bungie.net/7/en/News/Arti…
English

How is it that every @Convergecult album becomes my new favorite Converge album?. “Bad Faith” is such a sick song.
English

@jamonholmgren Yes, and also from anecdotal personal experience - a LOT of modern AAA devs don’t code at all but implement their own content. Few want to learn C++, much less its eccentricities. Plus they gravitate towards node graphs.
English

It still surprises me how many game devs who are using Unreal engine will use blueprints in rather than code.
No snark intended, but rather a speculative comment: I think it might be because C++ really isn’t a good primary language for working in day to day.
It’s super fast and has broad support and is very powerful, but it really and truly sucks from a productivity standpoint, and is far too massive (and permissive) of a language to use for normal scripting, and too many non-obvious footguns.
So blueprints allow you to essentially have a narrower scoped scripting language on top of C++.
That’s my theory. As always, I could be wrong; often I’ll throw out ideas to get corrections from people who have deeper experience and I learn things from that.
English

I would say AI isn’t as helpful as I’d like it to be, but Codex just used @steipete’s gog CLI to download and categorize a bunch of receipts sitting in my email for taxes 😂
I guess I’ll have to actually have it sort through my email hell now…
English

I’ve started an experiment where I’m having Codex develop audio plugins for me overnight. Right now I’m having it create JSFX for REAPER, just based on some simple prompts stored in the repo.
Download plugins (only one right now) if you use REAPER: github.com/adamtcroft/nir…
English

@LowLevelTweets @ThePrimeagen Now I need this to be the new PrimaGames 😂
English

I AM LOOKING FOR HELP (paid position):
* I need help with Unity / Unreal -- I want to revamp and recreate / change my scenes and i want to create a whole universe here.
If you know anyone / you want to do this, let me know, this would be a longer term relationship.
IF THIS SOUNDS GOOD: Fill this in -- docs.google.com/forms/d/e/1FAI…
English

@jamonholmgren It’s… ironic. In my LLM experiments, I’m only more convinced that area skill and expertise are gaining in importance (versus the hyperbolic “all our jobs are gone!”)
LLMs output isn’t remotely deterministic - so it requires a capable driver.
Love your thoughts ❤️
English

Bit by bit, we are starting to see what the new AI assisted software development world is going to look like for the next several years.
My current (still evolving) take:
- Massive unleashing of experimental work, proofs of concept, rough drafts
This should lead to a huge boost in the amount and creativity of software products that come to market, at the cost of a sudden increase in noise, a veritable din
- Significant decline in average code quality
Some code gets better, a lot gets worse, and the limitations of the current technology and unlocking less experienced developers to create software will lead to a near crisis in poorly built products in the near term
- Large proliferation of tools
As we scramble to adapt, experimentation and perspectives will lead to a vast array of possible solutions, each with their own sets of tradeoffs. Reminds me of the early days of Web 2.0, where there was a new framework every week and they all sucked
- Some emerging best practices
Over the past 6 weeks I have talked to many very experienced developers (often 1 on 1 video calls), and we are now starting to circle some common threads for AI-assisted software dev best practices (these are off the cuff, so don’t expect perfection):
1. Slow down, learn the tools, figure out the tradeoffs
2. Quality still matters when it matters; often, the existing models and tools fall short of maintaining that quality on larger code bases
3. The developer is responsible for the code they ship
4. Documentation (via skills, tasks, or just markdown docs) is tremendously helpful, but you should also understand it, not just rely on the AI to
5. High level architecture is still an area where a human with a lot of experience can add a ton of value
6. AI is not a substitute for good taste (and caring about things)
7. Some techniques are locally productive and globally harmful (more on this below)
8. The bottleneck is in review, understanding, and higher level systems architecture more so than coding speed.
9. Some developers are more adept than others at various parts of this new value chain. Current teams are full of developers vetted for and hired to do one job who are facing a significantly different way of doing it.
10. Coding itself might get done a different way, but the fundamental engineering patterns are often still extremely important. Not in cases where it’s just about satisfying some developer love of symmetry, but definitely in domains like data modeling and the like.
More on local optimization vs global concerns: agent chains like Ralph and aggressive code gen can feel incredibly fast, but tend to accumulate inconsistencies and tech debt over time. Speed at a file/feature level does not guarantee speed at the overall system level, and I have felt this personally when I’ve leaned too hard on such tools.
- We will learn more as time moves on. Be kind
We are adapting, evolving, playing with the tools, sharing, feeling the bruises when we get it wrong. There are educators trying to stay ahead of it and provide value. There are normal devs just trying to make a living, and stay relevant. None of this is as unique to you as you might think — I hear from others, and they’re feeling the same pressures. Be kind to each other
English



