this is just constrained generation? I thought the latest crop of decision models (inspired by Jev) do something fundamentally different in the architecture; they're not simply off-the-shelf models with a token mask
6,051 karma · joined October 14, 2015
I feel like lately this is becoming more common knowledge - but still something most people don't realize.
Part of it is probably the fact that it's impossible to depict "real" CPR in popular culture (movies, TV shows, etc) unless the production goes to extreme lengths to use a fake dummy. Even on The Pitt (which seems to make a point of being hyper realistic) I've seen them do "fake" CPR with shallow compressions.
As such, most of the fixes for Podman/Docker incompatibilities is just addressing that assumption with a few extra flags on the Podman commands to change how the user namespace maps between the container and the host, etc etc.
Adding lanes makes it far more confusing. I consider myself, you know, pretty smart. Not stupid, at least.
But I almost sideswiped someone in this roundabout the other day. Years of driving experience gave me an intuition that the middle lane would not cross over the outer lane. E.x. a car in the inner lane would not pass through the outer lane (except at the very end). So when I saw an oncoming car in the inner lane I thoguht I was safe to enter the outer lane. Not so. The inner lane car was actually lane-changing to the outer lane (at the exact point I was about to enter the roundabout) in order to exit.
But… does this escape the container? If not (the author seems to indicate it does not) then does it matter if you are in Docker or rootless Podman, right, since the end result is always: you have elevated to root within the container. If the rest of the container filesystem isolation does its job, the end result is the same? Though I guess another chained exploit to escape the container would be worse in Docker? Do I have that right?
Which implies there was some regression, some issue, some incorrect behavior or negative impact. One has to wonder… what could it have been? What could the issue with having a more accurate clickbox for the corner of the window possibly be?
My husbands TickTock feed is full of things like “10 things Americans do that Chinese think are weird”, “10 reasons Chinese cities are in the 22nd century” etc etc.
I personally don’t think that’s propaganda - most of it is factually true and would be pushed to the front of any fair algorithm because it is engaging. But I can kinda see the concern, even though I disagree with the outcome.
Lots of ifs there, though. I do trust Moxie in terms of execution though. Doesn’t seem like the type of person to take half measures.
Jokes aside, I'm curious how this is even possible over decades-old cable. I get there's a new DOCSIS standard, but I'm less interested in the protocol and more interested in the simple physics of it. How can a simple coaxial cable cram so much bandwidth?
Personally I've never seen a really strong source for that story, only anecdotes. I think it's an oversimplification to say "Strawberry Fields" made Brian Wilson insane. Instead, he was in a mental decline already. The pressure of "Brian Wilson is a genius" was getting to him:
https://en.wikipedia.org/wiki/Brian_Wilson_is_a_genius
There's a similar story with stronger sources, though. If you want to know about Brian's state of mind around that time, listen to his song Heroes and Villains. Basically, Brian worked on this song like it was his magnum opus, trying to reach the level of Sgt. Pepper. Quoting from Wikipedia (sue me):
> For Wilson, the single's failure came to serve as a pivotal point in his psychological decline, and he adopted the song title as a term for his auditory hallucinations.
> In the September/October 1967 issue of Crawdaddy!, journalist and magazine founder Paul Williams wrote that the song "originally had a chorus of dogs barking, cropped when Brian heard Sergeant Pepper, and was in many ways - the bicycle rider - a far different song."[39]
> Wilson held onto the final mix of the song for about a month. On the evening of July 11, 1967, he was told by his astrologer (a woman named Genevelyn) that the time was right for the record to be heard by the public. Without informing Capitol, Wilson called his bandmates and, accompanied by producer Terry Melcher, traveled by limo to personally deliver a vinyl cut of the record to KHJ Radio.[72] According to Melcher, as Wilson excitedly offered the record for radio play, the DJ refused, citing program directing protocols.[77][78] Melcher recalled: "Brian almost fainted! It was all over. He'd been holding onto the record [and] had astrologers figuring out the correct moment. It really killed him. Finally they played it, but only after a few calls to the program director or someone, who screamed, 'Put it on, you idiot!' But the damage to Brian had already been done."[79]
And this is all the tip of the iceberg. To have an even better understanding, you'd need to listen to the Smile! sessions, and the eventual 2004 "completed" recording of Smile!.
Personally, I think Brian was a genius (well, is; he's still alive, though not looking too good these days, sadly). But unlike The Beatles, who were four friends with an unbelievably tight bond (even after their breakup), Brian had no one else in the Beach Boys who could match him. And I think it was a weight on his shoulders, and that combined with the drug use (and likely a stroke at some point, which is obvious if you ever hear him speak post ~1968) brought his downfall.
I remember reading somewhere that literally most of the satellites orbiting earth are Starlink satellites. As in, more than half of all satellites are Starlink.
Obviously that statistic does not mean they have a successful business, that there's enough of a market, etc etc.
But one can imagine the types of services you can supply in just a few years to customers when you have by far the largest satellite constellation in the sky, with global coverage, and the ability to launch dozens of new satellites at low cost.
There's a lot of low-hanging fruit there. Particularly because the device has a USB port and, by design, exposes a user partition that you can read/write to (so you can upload files and documents and ebooks to the device).
There's definitely been an effort by Amazon to lock them down, but just taking your reverse-engineering tool of choice and decompiling their firmware binary will give you tons of readable code to dig through. They use a mix of java, native c, and javascript.
Fun fact, at startup the Kindle looks for certain files in the user partition, with certain naming patterns. You can, for example, disable the screensaver by dropping a file with a special name there. They patched this once, but after doing a grep for the user-partition mount location (to see all the places in their code where they read from user partition files) I was pretty quickly able to find another way to do this. It's fun stuff.
However, I'm not very disciplined with it. I don't have a "process". It's just my scratchpad with daily notes that end up being disorganized and gibberish.
Does anyone have a good process they follow with Obsidian? A good template, a good, disciplined strategy to stay organized? I'd love to hear it, please share :)
It was great! I loved it. I really felt like the narrator was knowledgeable and had something to teach me. His explanations for the process actually made sense, and weren't woo-woo mumbo jumbo. Perhaps it is all placebo effect, but even if so, the effect worked for me perfectly.
The problem is, as the industry exploded, it became harder and harder to find meditation guides that have that quality.
Even on Headspace, which I used to love, there's only so much they could provide, before the demand necessitated putting out meditation guides that are more and more ridiculous ("meditation for doing the dishes", "meditation for walking the dog", etc).
And I haven't found any guides that help me to the degree that the original Headspace ones did. Would love it if anyone has any recommendations.
(of course, you might say "you don't need a guide once you know how to do it", but personally I find it really hard without some recorded guide helping me along)
this is the part that blows my mind. You don't have to do this! You don't have to sample the entire output, and then validate after the fact.
You're not required to greedily pick the token with the highest score. You get the scores of all tokens, on every forward pass! So why even waste time picking invalid tokens if you're just going to validate and retry later on??
(note: when I say "you" here, I mean whoever is hosting the model. It is true that OpenAI does not expose all token scores, it only gives you back the highest-scoring one. So a client-side library is not able to perform this grammar-based sampling.
BUT, OpenAI themselves host host the model, and they see all token outputs, with all scores. And in the same API request, they allow you to pass the "function definition" as a JSON schema. So why not simply apply that function definition as a mask on the token outputs? They could do this without exposing all token scores to you, which they seem very opposed to for some reason.)
Guidance (and this project?): Let's not even bother with trying to convince the model; instead, we'll only sample from the set of tokens that are guaranteed to be correct for the grammar we want to emit.
When this is, really, a solved problem. I've been using github.com/microsoft/guidance for weeks, and it genuinely, truly guarantees correct output, because it simply does not sample from tokens that would be invalid.
It just seems so obvious, I still have no clue why OpenAI does not do this. Like, why fuss around with validating JSON after the fact, when you can simply guarantee it is correct in the first place, by only sampling tokens if they conform to the grammar you are trying to emit?
Have you ever seen a Youtube video about someone reviewing a Tesla or comparing it to another car? The comments are always full of hostile and vitriolic remarks by people who are personally offended if the video suggests the cup holders on a Model 3 are less than perfection. For some reason, Tesla is surrounded by a cult of personality where it's not just a car, it's a lifestyle.
And bizarrely, something similar is happening with this funny floating rocks. Here we are, on HN, and people in this very thread are calling Nature (Nature!) an "online sensational clickbait magazine" because they want to believe the hype that the rock has properties that they only learned about from Wikipedia a few days prior (and only understood 5% of it, at that)
Is there reason to be excited? Hell yeah. Are all the different replication attempts super fascinating? Hell yeah. Could it be the real deal? It could!
But this has become some weird spectator sport, where you're either a believer or a skeptic, and if you're on a different side than I am then screw you, even if you are Nature.
But during inference, there's no trick. The system message remains once at the top.
The base Llama-Chat models use something called "ghost attention" (they describe it in their paper). No clue how it works, but the result is, the model sticks to the system prompt extremely well. If you tell it that it's Marvin the Paranoid Android in the system prompt, it will stick 100% to that.
In llama-derived models like Vicuna, if you tell it to act as Marvin the Paranoid Android, eventually the regular "assistant" voice starts to bleed through, after only a few chat turns.
Doesn't sound like a big deal, but in cases where you have strict rules you want the model to follow, then the base llama-2-chat models are far better than any derived ones that do not implement ghost attention.
Are the superconductivity properties "all or nothing"? Meaning, if you get the synthesis wrong of the material, would you expect it to be "kind of" superconductive? (I guess, a very low, but non-0, resistance)?
Or is it something where it has normal resistance in all cases unless you get it just right, and then it becomes superconducting?
One more question - in physics, it seems like something having a value "0" is never really possible. Like, you can have something with low friction, but not zero friction. We pretend things have zero friction for high school physics problems, but they don't really. Is superconductivity like that? Is it actually, genuinely, exactly 0.0000(...)? Or is it just very close to zero, so close that it can behave like something with zero resistance (but in reality it's resistive, but just barely)?
There are so many garbage, lazily written product reviews, by websites that only exist to get people to click affiliate links. These sites only have one goal, which is to get you to click an affiliate link and make a purchase. So it is not in their best interest to say "You shouldn't buy this."
Rather, they make a list of "top X Foobars", they start with a really expensive one, then they follow with a more reasonably-priced one, and give it a very positive review. It leads to clicks and purchases.
Given this, it's not surprising to me that even the best LLMs carry pieces of this with them. Ask it to predict text describing some tech product on a sales page, and of course parts of that low-quality data will bleed through.
From a dissenting opinion: "Today, the court, for the first time in its history, grants a business open to the public a constitutional right to refuse to serve members of a protected class."
Remember, the job of the Supreme Court is not to decide what is morally correct or just, but to interpret the Constitution.
then again, not sure why that messaging would be any better for them than admitting to a DDOS...
denoising in latent space certainly seems like the "correct" path. My (amateur) thinking is, the more you can do in latent space, the better.