froztbyte

starting out[0] with "I was surprised by the following results" and it just goes further down almost-but-not-quite Getting It Avenue

close, but certainly no cigar

choice quotes:

Why is it impressive that a model trained on internet text full of random facts happens to have a lot of random facts memorized? … why does that in any way indicate intelligence or creativity?

That’s a good point.

you don't fucking say

I have a website (TrackingAI.org) that already administers a political survey to AIs every day. So I could easily give the AIs a real intelligence test, and track that over time, too.

really, how?

As I started manually giving AIs IQ tests

oh.

Then it proceeds to mis-identify every single one of the 6 answer options, leading it to pick the wrong answer. There seems to be little rhyme or reason to its misidentifications

if this fuckwit had even the slightest fucking understanding of how these things work, it would be glaringly obvious

there's plenty more, so remember to practice stretching before you start your eyerolls

One for the sidebar, in the spirit of incident-day-free counters[0], tracking how many days since the last time there was a dipshitted thing from the tescrealtors

Could do it with flap-counter or nixie-clock numbers for a bit of feel?

[0] - is this an insensitive idea? I know the counters tend to form part of safety culture and their reset indicates harm, but those clowns are exactly dangerous, so..

[open scene]

background: a brightly lit airy Social Gathering space with multicoloured furniture (meeting rooms are so 2010). people have been arriving in clumps of 2~5 for over 30 minutes, and the presentation can start soon

sundar: I want to thank you all for coming. this one should be quick today.

* sundar briefly sweeps his eyes across the room before continuing *

sundar: guys! GUYS! we made the prompt VIDEO CAPABLE! it can follow A STREAMING SEQUENCE OF IMAGES!! you can immediately start testing this from your corporate account (whispers if you're in the right orgs). for the public scoff, we'll start with Ask Us pricing in a few months, and we'll force it on the usual product avenues. the office and mail suites stand ready to roll out the integration updates before anyone can ask. you know how the riffraff gets....

* some motion and noise in the back *

sundar: ... sorry melanie, what's that? speak up melanie I can't hear your question. you know how much that mask muffles your voice...

* a game of broken telephone for moving a handheld microphone to the back of the room ensues *

melanie: hi sundar, congratulations to the team for their achievement. I wanted to ask: how does gemini pro solve the issues other models have faced? what new innovations have been accomplished? how is it dealing with the usual issues of correctness, energy consumption, cultural contexts? how is it trained on areas where no datasets exist? were any results sourced from cooperation with the AI ethics and responsibility workgroups that have found so many holes in our previous models?

sundar: * smiles brightly, stares directly into middle of crowd. moves hand to the electronic shutter control, and starts pressing the increase button multiple times until shutter is entirely opaque *

[sundar walks off into the fake sunset, breaks open the boardroom whiskey]

[inside the private exec room]

sundar: FUCK! that was too close. didn't we fire those types already in the last layoffs...? someone get me HR, we need to do something

[end scene]

this is in part because it's for (yet another) post I'm working on, but I figured I'd pop some things here and see if others have contributions too. the post will be completed (and include examples, usecases, etc), but, yeah.

I've always taken a fairly strong interest in the tooling I use, for QoL and dtrt reasons usually (but also sometimes tool capability). conversely, I also have things I absolutely loathe using

1. wireguard. a far better vpn software and protocol than most others (and I have slung tunnels with many a vpn protocol). been using this a few years already, even before the ios app beta came around. good shit, take a look if you haven't before
2. smallstep cli. it's one of two pieces of Go software I actually like. smallstep is trying to build its own ecosystem of CA tools and solutions (and that's usable in its own right, albeit by default focused to containershit), but the cli is great for what you typically want with certificate handling. compare step certificate inspect file and step certificate inspect --insecure https://totallyreal.froztbyte.net/ to the bullshit you need with openssl. check it out
3. restic. the other of the two Go-softwares I like. I posted about it here previously
4. rust cli things! oh damn there's so many, I'm going to put them on their own list below
5. zsh, extremely lazily configured, with my own little module and scoping system and no oh-my-zsh. fish has been a thing I've seen people be happy about but I'm just an extremely lazy computerer so zsh it stays. zsh's complexity is extremely nonzero and it definitely has sharp edges, but it does work well. sunk cost, I guess. bonus round: race your zsh, check your times:

% hyperfine -m 50 'zsh -i -c echo'
Benchmark 1: zsh -i -c echo
  Time (mean ± σ):      69.1 ms ±   2.8 ms    [User: 35.1 ms, System: 28.6 ms]
  Range (min … max):    67.0 ms …  86.2 ms    50 runs

6. magic-wormhole. this is a really, really neat little bit of software for just fucking sending files to someone. wormhole send filename one side, wormhole receive the-code-it-gives the other side, bam! it uses SPAKE2 (disclaimer: I did help review that post, it's still good) for session-tied keying, and it's just generally good software
7. [macos specifically] alfred. I gotta say, I barely use this to its full potential, and even so it is a great bit of assistive stuff. more capable than spotlight, has a variety of extensibility, and generally snappy as hell.
8. [macos specifically] choosy. I use this to control link-routing and link-opening on my workstation to a fairly wide degree (because a lot of other software irks me, and does the wrong thing by default). this will be a fuller post on its own, too
9. [macos specifically] little snitch. application-level per-connection highly granular-capable firewalling. with profiles. their site does a decent explanation of it. the first few days of setup tends to be Quite Involved with how many rules you need to add (and you'll probably be surprised at just how many things try to make various kinds of metrics etc connections), but well worth it. one of the ways to make modern software less intolerable. (honorary extra mention: obdev makes a number of handy pieces of mac software, check their site out)
10. [macos specifically] soundsource. highly capable per-application per-sink audio control software. with the ability to pop in VSTs and AUs at multiple points. extremely helpful for a lot of things (such as perma-muting discord, which never shuts up, even in system dnd mode)

rust tools:

1. b3sum. file checksum thing, but using blake3. fast!. worth checking out. probably still niche, might catch on eventually
2. hyperfine. does what it says on the tin. see example use above.
3. dust. like du, but better, and way faster. oh dear god it is so much faster. I deal with a lot of pets, and this thing is one of the invaluables in dealing with those.
4. ripgrep. the one on this list that people are most likely to know. grep, but better, and faster.
5. fd. again, find but better and faster.
6. tokei. sloccount but not shit. handy for if you quickly want to assess a codebase/repo.
7. bottom. down the evolutionary chain from top and htop, has more feature modes and a number of neat interactive view functions/helpers

honorary mentions (things I know of but don't use that much):

1. mrh. not doing as much consulting as I used to, using it less. quickly checks all git(?) repos in a path for uncommitted changes
2. fzf. still haven't really gotten to integrating it into my usage
3. just. need to get to using it more.
4. jql. I ... tend to avoid jq? my "this should be in a program. with safety rails." reflex often kicks in when I see jq things. haven't really explored this
5. rtx. their tagline is "a better asdf". I like the idea of it because asdf is a miserable little pile of shell scripts and fuck that, but I still haven't really gotten to using it in anger myself. I have my own wrapper methods for keeping pyenv/nvm/etc out of my shell unless needed
6. pomsky. previously rulex. regex creation tool and language. been using it a little bit. not enough to comment in detail yet

archive

"There's absolutely no probability that you're going to see this so-called AGI, where computers are more powerful than people, in the next 12 months. It's going to take years, if not many decades, but I still think the time to focus on safety is now," he said.

just days after poor lil sammyboi and co went out and ran their mouths! the horror!

Sources told Reuters that the warning to OpenAI's board was one factor among a longer list of grievances that led to Altman's firing, as well as concerns over commercializing advances before assessing their risks.

Asked if such a discovery contributed..., but it wasn't fundamentally about a concern like that.

god I want to see the boardroom leaks so bad. STOP TEASING!

“What we really need are safety brakes. Just like you have a safety break in an elevator, a circuit breaker for electricity, an emergency brake for a bus – there ought to be safety breaks in AI systems that control critical infrastructure, so that they always remain under human control,” Smith added.

this appears to be a vaguely good statement, but I'm gonna (cynically) guess that it's more steered by the fact that MS now repeatedly burned their fingers on human-interaction AI shit, and is reaaaaal reticent about the impending exposure

wonder if they'll release a business policy update about usage suitability for *GPT and friends

archive (e: twitter [archive] too, archive for nitter seems a bit funky)

it'd be nice if these dipshits, like, came off a factory line somewhere. then you could bin them right at the QC failure

Mr. Altman’s departure follows a deliberative review process [by the board]

"god, he's really cost us... how much can we get back?"

which concluded that he was not consistently candid in his communications with the board

not only with the board, kids

hindering its ability to exercise its responsibilities

you and me both, brother

I don't really know enough about the C64 to say anything one way or the other, but this comment on youtube did okay:

@eightbitguru
1 year ago
2021: We have definitely seen everything the C64 can do now.
2022: My beer. Hold it.

and I'm posting this without even having seen the whole thing yet

nitter archive

just in case you haven't done your daily eye stretches yet, here's a workout challenge! remember to count your reps, and to take a break between paragraphs! duet your score!

oh and, uh.. you may want to hide any loose keyboards before you read this. because you may find yourself wanting to throw something.

will this sure is gonna go well :sarcmark:

it almost feels like when Google+ got shoved into every google product because someone had a bee in their bonnet

flipside, I guess, is that we'll soon (at scale!) get to start seeing just how far those ideas can and can't scale

archive.org | and .is

this is almost a NSFW? some choice snippets:

more than 1.5 million people have used it and it is helping build nearly half of Copilot users’ code

Individuals pay $10 a month for the AI assistant. In the first few months of this year, the company was losing on average more than $20 a month per user, according to a person familiar with the figures, who said some users were costing the company as much as $80 a month.

good thing it's so good that everyone will use it amirite

starting around $13 for the basic Microsoft 365 office-software suite for business customers—the company will charge an additional $30 a month for the AI-infused version.

Google, ..., will also be charging $30 a month on top of the regular subscription fee, which starts at $6 a month

I wonder how long they'll try that, until they try forcing it on everyone (and raise all prices by some n%)

The Mistral 7B Instruct model is a quick demonstration that the base model can be easily fine-tuned to achieve compelling performance. It does not have any moderation mechanism. We’re looking forward to engaging with the community on ways to make the model finely respect guardrails, allowing for deployment in environments requiring moderated outputs.

“Whoops, it’s done now, oh well, guess we’ll have to do it later”

Go fucking directly to jail