@mcc@mastodon.social
@mcc@mastodon.social avatar

mcc

@mcc@mastodon.social

glitch girl

This profile is from a federated server and may be incomplete. Browse more on the original instance.

rain, to random
@rain@hachyderm.io avatar

Recommend a keyboard shortcut that more people should know about! Any environment, any OS, any kind of software.

I've talked about M-. in shells before (last word of last command, absolutely incredible). But here's a completely different one in the most common programming environment on the planet:

In Excel, and Google Sheets, Ctrl-Shift-1 through 6 applies various display formats to cells. For example Ctrl-Shift-4 ($ on qwerty) formats a cell as currency.

mcc,
@mcc@mastodon.social avatar

@rain On the Macintosh, if you hold the Command (coverleaf) button and click the title of an active window, if that window corresponds to a file on the hard drive, it will show you (and allow you to open in the Finder) the enclosing folder.

(Holding command and clicking the titlebar of a non-active window, on the other hand, will let you move around the window without giving it focus.)

mcc, to random
@mcc@mastodon.social avatar

Would you let her sing to you https://mastodon.social/@mcc/112525923585865855

mcc, to random
@mcc@mastodon.social avatar

I had a dream that there was a Mastodon instance that could only be posted to from a particular building. When I was in college there was this one house just off campus that basically turned into a permanent LAN party, J.J. and his three roommates had set up a computer lab in the basement random people were playing WoW on 24/7. In my dream, a house like this had become a sort of maker space and they'd set up a Fediverse server to reject posts unless the request came from within the House's wifi.

mcc,
@mcc@mastodon.social avatar

The House's Mastodon server became a kind of chronicle of the various projects that were being created there, and the desire to get one's projects documented became an incentive to come to the hackerspace house to do them. In the dream I walked through the house and had to step carefully as the floors and couches were packed with college students intently typing on laptops

mcc,
@mcc@mastodon.social avatar

(It was unclear to me if the Mastodon server could be accessed for post reads from outside the house. I think it became a sort of Metafilter, allowing output but no input, the rest of the dream Fediverse watching the instance's traffic and the strange and cool things being made there in fascination, but barred from participating unless they could enter and become part of the physical community.)

mcc,
@mcc@mastodon.social avatar

Addendum: This is the second time I have typed this. When I found I'd awoken, I immediately thought, oh! What a strange and evocative dream! The Internet would like this! And I wrote up a long Cohost post describing all the the dream's details and the House's emergent culture. But no sooner had I finished than I discovered that in fact I was still dreaming, that this was not really Cohost but just another dream, and that all I'd written was to be lost within the outer dream forever

simon, to random
@simon@simonwillison.net avatar

I wrote about a common misconception I see people have about LLM tools like ChatGPT

Training is not the same as chatting: ChatGPT and other LLMs don’t remember everything you say

https://simonwillison.net/2024/May/29/training-not-chatting/

mcc,
@mcc@mastodon.social avatar

@simontatham I don't… I'm sorry, I clicked "Reply" on a post by a Simon Williamson and somehow this happened! I don't know what I did but I apologize _

mcc,
@mcc@mastodon.social avatar

@simon I find this article confusing. Is your point to address privacy concerns about queries to LLMs, or to provide helpful advice about how LLMs are used?

Your point is taken that yes incorporating inputs directly into the model is not how an LLM works (I'd quibble with your use of "stateless" in case of RNNs, but maybe that's standard lang in community) but if we're talking about privacy the model I always assumed would be the concern is "your inputs become training data for the next model".

mcc,
@mcc@mastodon.social avatar

@simon I think I'd summarize my concerns as:

  1. Cloud LLMs are totally opaque black boxes. You don't ever really know what they're doing with your data or how they're generating the results.

  2. OpenAI has demonstrated they're willing to ignore or creatively interpret copyright law when scraping training data. This means I must assume they will ignore or creatively interpret privacy law. If the privacy policy says my queries are not used to train, I cannot trust this.

mcc,
@mcc@mastodon.social avatar

@simon 3. The NSA is reading every piece of input data sent to OpenAI, be it a simple query or a file to be summarized. You have to assume this. If it's not true now it will be soon. OpenAI is too much of an attractive nuisance to not be a target for nation state intelligence. As OpenAI gets built into every application, the ability to monitor OpenAI queries becomes like the ability to selectively look over any computer user's shoulder, this is a useful ability, therefore the NSA will obtain it.

mcc,
@mcc@mastodon.social avatar

@simon And incidentally, (1) means that I think the emphasis you place on the limited history model for LLM "chatbot" systems is unduly strong even for basic usage. My limited understanding is the OpenAI developer AI documents context consisting only of the current session. But an OpenAI frontend could easily, for example, end each session by asking OpenAI to summarize everything up to that point, and then prepending that summary to the next session's input. The user wouldn't necessarily know.

mcc, (edited )
@mcc@mastodon.social avatar

@simon In fact OpenAI themselves could, and maybe should, do this when people are using the ChatGPT frontend. The only reason to assume OpenAI isn't doing this with ChatGPT is that would be useful and therefore if they were doing it they'd be bragging about it…! (EDIT: On a second read I see you link ChatGPT DOES have a feature such as this. Well then.)

That is a lot of replies, I apologize if this is badgering a bit.

mcc,
@mcc@mastodon.social avatar

@glyph @simon I agree but also I think this creates a problem with the conclusion— the article says the LLM doesn't inherently remember inputs, but admits in the text the company inherently may remember inputs, and if you are discussing policy as in the conclusion, isn't the company remembering inputs the same as the model remembering inputs with a slight delay? Like it's important nuance for a detailed discussion but I'm not sure it changes the 10,000 foot risk model at all.

mcc,
@mcc@mastodon.social avatar

@simon It is, and like… a problem with all cloud services, not just "AI", is that they can change without you knowing it? So like if ChatGPT or similar services one day introduced a degree of resistance to prompt leaking attacks would we necessarily know it…

mcc,
@mcc@mastodon.social avatar

@glyph @simon I'd also, and here I'm just being difficult, but I'd be curious "how much would you have to change the way LLM AI works to make it work like the 'magic' system Simon denounces?". Like, I've used markov chain systems that instantly incorporated every line of query text into their model (MegaHAL). LLMs are a lot more complicated, but a learn-as-you-go cloud neural network model sounds to me like an interesting engineering problem not an impossibility.

mcc,
@mcc@mastodon.social avatar

@glyph @simon Again the only reason to assume such a thing doesn't exist is if a company had solved that problem they'd probably be bragging about it. Which also means one or more companies working on the tech might have solved the problem already and not yet productized it (and therefore not publicized it)

mcc,
@mcc@mastodon.social avatar

@glyph @simon Like what if the misconception you're correcting is actually not an observation about AI but rather a situational, temporary quirk of how the products happen to work in 2024, which may not still be accurate in 2026, and therefore it turns out to not be useful to [the inherently slow process of] corporate or government policymaking in 2024? Again I know I'm being difficult here, but…

aeva, to random
@aeva@mastodon.gamedev.place avatar

I created something perfect on accident yesterday, and I cannot show any of you.

Not because it is a secret or anything, just because I can't record it without it turning into total garbage.

It is beautiful and perfect I want to make a whole game out of it and also Gnome's screen recording thing just completely shits itself when I try to record it, so you'll just have to pretend you understand just how good this all is.

mcc,
@mcc@mastodon.social avatar

@aeva hm can you compile it to webgl

mcc, to random
@mcc@mastodon.social avatar

Sorry, you want me to what

mcc,
@mcc@mastodon.social avatar

@natsume_shokogami There is more than one red flag here

mcc,
@mcc@mastodon.social avatar

@natsume_shokogami Yeah, these are some of the red flags I was thinking of

mhoye, to random
@mhoye@mastodon.social avatar

This is a remarkable graph.

You might have heard that "EV sales are slumping", "people are starting to avoid EVs", etc.

That's not what's happening.

What's happening is "Tesla is cratering so hard that it's skewing the aggregate market data."

mcc,
@mcc@mastodon.social avatar

@mhoye Hasn't the US gov been selectively issuing subsidies that give Tesla a boost on price while shutting out some of the offshore companies? I wonder if the numbers would be even more pronounced if not for that.

glyph, to random
@glyph@mastodon.social avatar

@mcc me thinking about using systemd: "I don't understand why people are so mad about this, launchd is better than sysv init, this is just launchd for linux, I guess they couldn't just port it because some low-level stuff is different but I'm sure it's basically the same"

me actually using systemd: "fuck fuck what is this shit what the fuck why didn't they just actually port launchd to linux, there's no excuse for this"

mcc,
@mcc@mastodon.social avatar

@vt52 @glyph Systemd was supposed to be the improvement on init.d! That was the pitch! It should have improved!

mcc, to random
@mcc@mastodon.social avatar

Tomorrow (Sunday the 26th) at 2 PM ET (11 AM PT) I am going to be streaming on Twitch the beginning of "Animal Well", a game that I know nothing about but everyone says is super cool and full of secrets. Watch at: https://www.twitch.tv/mcc111

(I usually try to announce streams on Friday if they're gonna happen but this time I didn't. Sorry. This may be a short stream anyway.)

mcc,
@mcc@mastodon.social avatar

@hook I'm still in the first playthrough which I think is when people find the "basic" secrets and I'm finding even the first tier of secrets highly compelling. This game is so thoughtful and it responds so well to prodding.

"Less brutal rain world" is actually not a bad comparison, though your protagonist character is less Fluffy. The art style is simple but gorgeous.

mcc,
@mcc@mastodon.social avatar

@hook The most likely thing I'd do is write a Cohost post but I don't have so much spare time for "blogging" lately…

mcc, to random
@mcc@mastodon.social avatar

For the last two years I've been semi-daily posting "What I'm Listening to Today" links here. Mastodon has some problems with threads containing hundreds of posts, so I re-create the thread once a year.

If you'd like to see, here's my "year two" thread: https://mastodon.social/@mcc/110266770603341546

Or, alternately, every song from year two in the least practical format possible: A 301-song, 38-hour YouTube playlist (note: video #1 contains flashing):

https://www.youtube.com/playlist?list=PLLIjft6ja7DM_kacOW8zo2vtr-aWpTNX6

And here's the thread for "year three":

mcc,
@mcc@mastodon.social avatar

What I'm listening to today: "Battery Driven", quadratschulz

This is a quirky little hiphop jam on a small collection of handheld/toy synthesizer equipment. Featuring 808 cowbells, Chase Bliss pedal mangling, and extended vocals by Miku Hatsune. For serious. That stylophone looking thing is the "Gakken Otona no Kagaku NSX-39 Pocket Miku Singing Keyboard", an officially licensed Vocaloid product. Skranky

https://www.youtube.com/watch?v=FeSrYpB53JU

  • All
  • Subscribed
  • Moderated
  • Favorites
  • •
  • megavids
  • rosin
  • Durango
  • osvaldo12
  • khanakhh
  • mdbf
  • DreamBathrooms
  • ngwrru68w68
  • magazineikmin
  • thenastyranch
  • cubers
  • Youngstown
  • slotface
  • kavyap
  • JUstTest
  • GTA5RPClips
  • anitta
  • tacticalgear
  • modclub
  • ethstaker
  • cisconetworking
  • InstantRegret
  • everett
  • provamag3
  • normalnudes
  • Leos
  • tester
  • lostlight
  • All magazines
    •