EmNudge

EmNudge

@emnudge.dev

Code frog https://emnudge.dev/notes/ eng @ramp.com | prev @ tiktok, lyft

82 followers115 following221 posts

EmNudge
@emnudge.dev

Realizing a lot of edutainment content I used to enjoy spends most of its runtime avoiding making an argument. You don’t learn much other than that the author also doesn’t know and decided to waste your time anyway.

EmNudge
@emnudge.dev

this may sound kind of ridiculous, but one of the things I'm most excited for in the next evolution of TV/Movie writing is (hopefully) much increased accuracy of the material with LLM research. bsky.app/profile/adap...

Ada Palmer@adapalmer.bsky.social

Was watching Buffy’s Anya backstory episode. Friend asked why I sighed. “She’s not of a sufficiently high social class to have domesticated rabbits that far north in Europe in that century, but the writers couldn’t know, the paper was only presented at a conference yesterday.” Accuracy is fleeting.

EmNudge
@emnudge.dev

There's no easy way to write reasonably accurate historical media without intense consultation from numerous experts. This is annoying because it creates a waterfall approach to writing where the literary experts and the informational experts are 2 separate collaborative audiences.

EmNudge
@emnudge.dev

i.e. it's hard to write good stories if you need to go back and forth like this. Furthermore, it's hard to know who you need to consult or to get the access to such experts. Authentic writing is deeply difficult. Knowledge work made easier enhances our ability to better in this domain.

EmNudge
@emnudge.dev

Getting a name for something is often satisfying. Oh, I was just suffering from Aegrotare this whole time! And then you learn that aegrotare is latin for "not feeling well" and it becomes tremendously unsatisfying. Lots of terms are like this.

EmNudge
@emnudge.dev

I don’t really buy the second half of this at all. It obviously doesn’t generalize very well. Is it bad if there are fewer coal miners in the future as a result of modernization? bsky.app/profile/euge...

Eugene Vinitsky 🍒@eugenevinitsky.bsky.social

To take an example, it is unequivocally a good thing that we now have 700+ new proofs of interesting things. It is bad if there are fewer mathematicians in the future as a consequence (though of course there might be more instead somehow)

EmNudge
@emnudge.dev

to be clear - I don’t think it’s likely or even good, but it’s certainly not obvious if it’s correct. Why is it valuable to preserve an occupation outside of its merits? Is it just longing for an experience made unavailable? Like seeing the extinction of particular merchant or craftsman?

EmNudge
@emnudge.dev

The biggest harm (imo) is in how quickly these things happen. The biggest risk to a miner losing their job is loss of income and purpose. If they immediately got a fulfilling well paid job, I doubt there’d be much backlash

EmNudge
@emnudge.dev

Had a neat idea. I have a list of articles I've read that I like on my website. I stopped updating it a while back because this is annoying. I read a lot and merging the gold nuggets into my repo is a lot. Atproto and BlueSky might solve this? I can tweet articles with a tag and sync this on build

EmNudge
@emnudge.dev

fwiw this is pretty much untrue for 99% of the population. Talking is almost certainly faster than typing for everything but rigid and concise programmatic syntax. You can do a test yourself. Absent a speech impediment and impeccable typing speed, you'll find ... bsky.app/profile/stev...

Steve Klabnik@steveklabnik.com

I’ve been thinking about something related: typing is often faster than talking because in the past inputs have to be ridgidly correct. But you don’t need to have perfect grammar or punctuation or anything to talk to an LLM. It’s fine.

EmNudge
@emnudge.dev

you talk (on average) at your preferred rate of information exchange, which is an upper bound. That is, people talk at least as fast (often much faster) than they type. Average talking speed is 120-160 WPM and typing is ~50 WPM. I type at 110 WPM when I really try and 70 WPM when I don't.

EmNudge
@emnudge.dev

Seeing a bunch of "I asked AI to make this in raw WASM" posts around. Should note that this actually doesn't help much. This is not a productivity unlock. I spent a while investigating this with github.com/EmNudge/wat-... - trying to answer: will writing raw WAT/WASM make things faster? It didn't

EmNudge
@emnudge.dev

Turns out compilers are really good and powerful. The biggest issue is binary bloat, not performance. If you strip symbols aggressively enough, you can get close in size as well. LLMs are smart and will do good work with WAT, but not substantially better than C or Rust directly.

EmNudge
@emnudge.dev

You can take a look at the commit history. I really tried here. And actually made something interesting! But it wasn't substantial - the experiment failed. In truth, something off the shelf and written in C (but compiled to WASM) would have been sufficient.

EmNudge
@emnudge.dev

This is also legitimate. I highly recommend taking 15m to learn the secure shoelace knot. It's really easy and a permanent life improvement!

EmNudge
@emnudge.dev

prediction: The public is going to see progress not at all and then very quickly. A lot of "AI is useful, actually" arguments fall on deaf ears for those whose work seems to see little improved by any recent new model announcements. I kinda get it.

EmNudge
@emnudge.dev

The jump likely comes from hardware releases, which we're still meaningfully far away from realizing. Software improvements trickle down to hardware construction, but it's not an immediate step change. It will take some time.

EmNudge
@emnudge.dev

Even starting with an AI written prose page and carefully editing the document still leaves you with an AI smell. It feels off and wrong and bad to read. I don't think this is a lost cause. I still really really don't enjoy reading AI writing. However, we're likely just missing a primitive.

EmNudge
@emnudge.dev

That is, there may be a collaborative writing primitive that lets you leverage its ability to type and cycle through ideas without incurring a massive mental debt or style shift.

EmNudge
@emnudge.dev

Turns out LLMs that produce answers about poker hands without writing code are pretty terrible. Leads me to believe that the training data is just lots of wrong answers to basic questions. Quick pattern matching models get number crunching wrong. Pro models (even without code running) fare better.

EmNudge
@emnudge.dev

Models will train on this in no time if we make this a public benchmark. I think it might already be one. Still, interesting to see at present since this is presumably the kind of question someone might ask that isn't terribly contrived.

EmNudge
@emnudge.dev

It's easier if you propose the solution in advance to the LLM (tell me why this false statement is correct). Gemini gave me a script to trick itself that works consistently on flash models. This is cool, but also bad. Gemini will be used by many people. It should know to switch models for these.

EmNudge
@emnudge.dev

I like Voice AI conceptually. Hands-off way to learn new things. However, you cannot just take the text output from an LLM and read it aloud. It would take forever and is not at all how people communicate. It's insufferable and the source of lots of satire.

EmNudge
@emnudge.dev

Q: Why do outlets have more than one prong? A person: Wouldn't work otherwise. You need at least 2. One to receive and one to send back. A third for safety, sometimes. an LLM:

EmNudge
@emnudge.dev

A lot is helped with custom instructions. I have a bunch that I use. Might be time to modify it with some examples so that my text output is also similar to this.

EmNudge
@emnudge.dev

A lot of my takes around bad things involve the visibility of bad things being good and thus our conversation around bad things indicates good things. Thus, things are less bad now than between the time of the bad thing and us discussing it.

View full thread
EmNudge
@emnudge.dev

Discussion about "how do you know this photo is authentic" makes me excited - not because I think the state of photo authenticity or misinformation hasn't gotten worse, but because it has always been an issue. There have always been those with the resources to fool you. Now we can discuss it.

EmNudge
@emnudge.dev

"Solving" misinformation does more than combat the direct aggressor of today, but fixes perhaps decades of information authentication ambiguity and perhaps many decades more into the future. Signed certificates of data origins move trust in a way that's probably better for everyone

EmNudge
@emnudge.dev

A lot of my takes around bad things involve the visibility of bad things being good and thus our conversation around bad things indicates good things. Thus, things are less bad now than between the time of the bad thing and us discussing it.

EmNudge
@emnudge.dev

More concretely, lots of things that are talked about are just bad things which are now severe enough to warrant discussion, but were in fact always present. The time between bad thing and us discussing bad thing is quite large, actually. Information authenticity was always an issue, for example

EmNudge
@emnudge.dev

updog? What's updog? Nothing much, just datadog's open platform degradation analysis tool that lets you know when GitHub is lying about its own uptime. updog.ai

EmNudge
@emnudge.dev

Turns out spending a long time building a clone of da vinci resolve is a good way to learn how da vinci resolve works.

EmNudge
@emnudge.dev

running into that "smarter models are actually cheaper because they get the task done faster" problem again. Let my claude subscription run out in favor of an OpenRouter credit limit with various cheaper models.

EmNudge
@emnudge.dev

Yup, I'm back to a Claude max plan. I long for the day I can self host this sort of thing. Or perhaps my use of frontier models means I'll pretty much always pay for frontier performance

EmNudge
@emnudge.dev

Obviously not a one-shot job, but super low effort. I think I'm just doing enough of these to know how to do this very quickly and effectively. And the models have gotten better. That helps. I also do know enough about color grading here to give good feedback. Also I built a custom feedback tool.

EmNudge
@emnudge.dev

Claude, I know what we're gonna do today bsky.app/profile/emnu...

EmNudge@emnudge.dev

I don't understand why some very popular software is not both good and open source. Maybe the private plugin market is just so thick now that no one has an incentive to change the status quo. There should be a good free Davinci Resolve color grading competitor. There should be a good free capcut.