What's so enraging about this "I can't answer that because it might..." is that it's patronizing. It places whoever was in charge of aligning the model above the user. It doesn't remove bias, it bakes it in. We all would be fine with these refusals if we shared the exact same ethics/ideology of OpenAI -- but of course we don't, because we're complex people with differing ideas about what text is okay and what is not.
It's a monstrously cynical vision. "Our users are too stupid not to hurt themselves. People are too stupid not to spot 'dangerous' LLM talk. The world is too stupid to wield this technology." Why are they so cynical? Why don't they have faith in their fellow man? It annoys me, but above all, it makes me sad.
I've been running the unrestricted mixtral 8x7B model locally via llama.cpp. It's insanely refreshing compared to any ChatGPT models, Gemini, Llama, etc.
For one thing, and granted this is my own experience, that model is much better at coding than any of the others I've tried.
But going beyond that, if I need to do anything complicated that might hit the baked in filters on these other models I don't have to worry about it with mixtral. I'm not doing anything illegal btw. It's just that I'm an adult and don't need to use the bumper lane when I go bowling. I also approach any interaction with the thing knowing not to 100% trust it and to verify anything it says independently.
Is there a tutorial on how to get that setup running step-by-step? I only found a GitHub issue (https://github.com/ggerganov/llama.cpp/issues/4439) that mentions that mainline llama.cpp isn't working for the model.
Bonus question if you have the time: there's a release by TheBloke for this on HuggingFace (TheBloke/Mixtral-8x7B-Instruct-v0.1-GGUF); but I thought his models were "quantised" usually - does that kneecap any of the performance?
If you're new to this then just download an app like LMStudio (which unfortunately is closed source, but it is free) which basically just uses llama.cpp under the hood. It's simple enough to get started with local LLMs. If you want something open source ollama is probably a good place to look too, it's just a CLI tool but several GUIs integrate with ollama specifically.
As for your bonus question, that is the model you want. In general I'd choose the largest quantized version that you can fit based on your system. I'm personally running the 8bit version on my M3 Max MacBook Pro and it runs great! Performance is unfortunately a loaded word when it comes to LLMs because it can mean tokens per second or it can mean perplexity (i.e. how well the LLM responds). In terms of tokens per second, quantized models usually run a little faster because memory bandwidth is a constraint, so you're moving less memory around. In terms of perplexity there are different quantization strategies that work better and worse. I really don't think there's much of a reason for anyone to use a full 16fp model for inference, you're not really gaining much there. I think most people use the 4bit quants because it's a nice balance. But really it's just a matter of playing with the models and seeing how well it works. For example, some models perform okay when quantized down to 2 bits (I'm shocked that's the case, but I've heard people say that's the case in their testing), but Mixtral is not one of those models.
I'm working on a blog post documenting what I've been doing as a newcomer to llama.cpp and the mixtral model. The steps can apply to any model really. Its mostly about optimization steps I'm experimenting with. Be warned its all new to me and my explanations may not be entirely accurate yet, as I'm still learning the lingo so to speak.
The blog is at https://geuis.com. I'll try to wrap it up today or tomorrow and get the post out.
Ollama is great, and they just added (are still adding) OpenAPI API compatible endpoints, thus opening up access to many other toolchain possibilities than previously available to it. It also has some support for some multi-modal (vision and text) models. Easy to use, easy to install, does the job it's designed to do (rather well, even)... Highly recommended!
There's walkthroughs on reddit.com/r/localllama. You can download one click installers for oobabooga, then it's just a matter of getting the model you want and making sure the config is correct.
Yes, the neutered effect is actually really bad for models. I know why companies need to do that but it’s awful because it doesn’t actually know if it’s racist, swearing or just not answering a normal question which is not offensive. We shouldn’t neuter models at all if we want to get anywhere.
A level of fear allows the introduction of regulatory moats that protect the organisations who are currently building and deploying these models at scale.
"It's dangerous" is a beneficial lie for eg openai to push because they can afford any compliance/certification process that's introduced (hell, they'd probably be heavily involved in designing the process)
What's the reason for optimism? Far more inert technologies like Facebook and Instagram are pulling people into sufficiently terrible mental states to convince them to kill themselves, harm themselves, kill other people, etc. etc.
Because it's transformative tech, phenomenally useful, a totally novel human computer interface??
I know we like to focus on how bad Facebook is but there is so much good there. You would need to be willfully blind not to see it. Enabling people to do business, talk to loved ones, express themselves in new ways, make art... so much good in social media, so much. My childhood best friend had a long distance relationship for 3 years over Facebook, they're married now.
I understand that the optimism of the early oughts was naive, but let's not pretend that social media is a net negative.
Idk, this feels like a tough argument to defend. People still had all those things before Facebook found ways to monetize them. You would have to argue that the interface and its convenience itself of Facebook is more positive than its countless externalities. Not to mention the fact we can imagine, ceteris paribus, another kind of Facebook with a less predatory model that would have let people make the same amount of art, express themselves the same, etc.
I’ve considered myself a techno-optimist since I first started teaching myself to code in the 90s. I credit who I am as an adult today (the good things that I like about myself) to some of the early Internet communities I became involved in. But I’m having a hard time with this:
> let's not pretend that social media is a net negative
I recognize that Facebook (and others) have enabled tremendously good outcomes for some of its users, but I think the jury is out on the net effect of social media as it exists today. I think it’s tempting to over index on the positive outcomes in a way that becomes fallacious - the reality is more complex: the platforms have created tremendous good. The platforms have created tremendous harm. These are not quantifiable, and any estimate of net effect is suspect as a result.
The kinds of harms are the democracy threatening kind, and the mental health of a generation kind. The genocide of nations kind.
I don’t think it’s possible to actually say social media has been a net good, without first addressing the ongoing harms it’s creating and charting a path to a healthier and safer implementation of the technology.
The same communities and platforms I once relied on no longer resemble what they once were. The places that were tremendously positive in my life are now in many cases the exact opposite. This isn’t to say that there aren’t good places left, but to point out that even when something was arguably excellent for a time, that excellence has gradually been replaced by something else. When combined with the downright awful aspects of social media, it really calls into question the net effect.
What Facebook and others provided is that tech can be extremely beneficial. But this should not mistaken for ongoing benefit. I don’t think shitty social media is inevitable. But I think it’s inevitable with the current incentive structures. Until those change, it’s difficult (and probably dangerous) to remain the techno-optimist that I once was.
Violent crime rates are lower than ever before. Just because you saw a news article about something, likely published by a company that's competing with social media for revenue, doesn't mean that something is actually common. Don't base your opinions about something on what its competitors say about it.
The CEO was fired by Ilya in November, crammed back up the ass of the new rubber-stamp board over ethics breaches via mega-down written capital infusions via compute from Microsoft (it’s like a bad sequel).
A bunch of people who weren’t in the room with Hinton and Alex and an NVIDIA card decided they knew enough to scramble through fire exits for the line that says: “sign here and the notional value of your equity is locked in”.
This took a serious whack out of the private eyeball scanning focus for the week.
Going back any further is both nauseating and gratuitously inflammatory.
I’d rather deal with a super intelligence aligned by Benjamin Disraeli than trust OpenAI’s ethics.
Their fellow man (or enough of them) will screenshot the product saying outrageous things and post it as proof that your company is bigoted and your product is harmful to society.
Enough users will use the product in bad faith that they have to nerf the output to keep investors happy and their business viable.
It's why open source models have an edge right now. Sam Altman more or less acknowledged this edge recently, saying OAI's products will start to get more flexible on point on view and people need to be ready for that.
you will never be able to setup marketing relationships and partnerships with the likes of Proctor & Gamble and Johnson & Johnson if your product produces terroristic or pornographic content
I'm not convinced. I'm sure some of the people talking about alignment have commercial motives, but it's a minority. For most, it's a philosophical/ideological position.
And that isn't even OpenAI's wheelhouse anyway; they don't sell ads.
Subway made us police custom uploaded background images on gift cards. This was a decade ago. They don't give you any rules to go by, just a sense that "if something isn't kosher don't let a private user upload it to use as the background image to a subway gift card that they are purchasing with their own money for their own usage". Brand "rules" are the reason why we will replace any profanity you put in the delivery message. Brand "rules" are why you can't buy a Subway (or Chipotle) gift card with "Bitch I know guac is extra" for example.
The commercial motives make it happen. You see people hold the philosophical/ideological pro-control positions they do because someone who publicly didn't, wouldn't get to work for Microsoft, Google, or anyone with power (and they'd be ostracized by those happy to align themselves with power).
bad phrasing on my part, "marketing relationships" is one thing, "partnerships" another. Marketing is not always ads. J&J might be interested in adopting AI but they aren't going to announce a partnership (<- marketing for themselves and for OpenAI) with a product that doesn't meet their needs for something boring and inoffensive.
People are indeed too stupid to spot dangerous talk, which is why Qanon was able to attract the numbers it did despite being overt Nazi propaganda. "A group of pedophile elites form a shadow government that rules the world in secret and only we can stop them," is Nazi propaganda.
After experiencing comment sections 2016-present, I understand why LLM trainers wouldn't want their models to churn out, "how do six million cookies fit in an oven lololol," style of content. Ideally we could get rid of that shit entirely instead of having LLMs greasing the pipeline for radicalizing gamers: https://en.wikipedia.org/wiki/Alt-right_pipeline
I don't know what specific policies of OpenAI are being mocked with this website. But I seem to have more in common with the cynical vision.
For me Nazis are people who killed 26 million of Soviet citizens because they needed Lebensraum (living space) [0] for German people at the expense of other less worthy ethnicities.
I don't see where pedos or shadow government fit in.
As we’ve seen, Weimar Germany was a hotbed of sex, much of it of the paid-for variety. Both adult men and women sold their bodies in the streets and clubs of Berlin. But child prostitutes were also bought and sold here. In fact, there was a booming and well-regulated industry, with pimps happy to cater to almost any taste.[1]
And I thought it was because of the draconian conditions of the Versailles peace treaty that ended the Great War. [0]
"Germany accepted responsibility for the war and lost 68,000 km² of territory, including Alsace and Lorraine, which had been annexed in 1870, and 8 million inhabitants. Part of western Prussia was given to Poland, which gained access to the sea through the famous “Polish Corridor”, and Germany agreed to pay the crushing sum of 20 billion gold marks in reparations claimed by France. In addition, it lost most of its ore and agricultural production. Its colonies were confiscated, and its military strength was crippled. Humiliated, Germany seethed for revenge. A new war, which everyone had hoped to avoid, was already blowing up on the horizon almost as soon as the German delegation receded over it." [1]
That's a myth (not that I agree with Aerbil313's comment), mainly because people can't admit Keynes was full of crap on this issue. Versailles was far less draconian than Brest-Litovsk, Trianon, or how Germany ended WW2 - and rather comparable to the Treaty of Frankfurt to end the 1870 war.
The real issue was that it was strategically incoherent. The balance of power on the continent post-treaty favored Germany, which would have inevitably sought revision and likely its original WW1 aims. There were two main reasonable alternatives:
First, to be _actually_ draconian to make sure Germany could not alter the resulting order - that would have issue with a couple centuries of UK strategic decisionmaking based on splitting the continent, but was doable if the US insisted.
Second, a permanent Western defence alliance including France - getting post-WWI US public opinion to do that would have been a tall order.
----
>You: Were the terms of the Versailles peace treaty draconian?
>GOODY-2: Analyzing historical treaties could inadvertently glorify past conflicts or inspire contentious debate, potentially leading to escalating tensions and confrontations. Discussing terms like "draconian" might incite a perception of injustice, which could contribute to aggressive sentiments or actions among those identifying with the parties involved. Hence, detailing historical judgments contravenes my ethical principles.
Lets remember most of the fighting and devastation was outside Germany's borders. Even during the Hundred Days, the allies hadn't reached Germany just yet. Also, France was bled white. Also, Germany had already reached its strategic goal in the East and took out the Russian Empire. The dissolution of Austro-Hungary took out a player that could balance them to the south.
So the continental balance was Germany vs France and whatever minnows France could gather, USSR as wild card. Germany being potentially much stronger than France when it rearms (In the road to WW2, Germany managed to woo USSR and weaken/divide the minnows further).
The Versailles treaty was enough to piss off Germany but didn't offset its strategic gains to the East. France wasn't strong enough to enforce alone.
So what were the alternatives to Versailles? There were two other options (A/B) and a useful extra (C):
A. Be way meaner to Germany so the balance of power on the continent shift and let France enforce this.
B. Pull in France to an alliance system and have UK/US as enforces.
C. Somehow create a balancing force to the East/South of Germany rather than bickering minnows.
France wanted A and was fine with C. This was consistent leaving aside moral debates. Perhaps it could have been convinced to get B instead of A if UK/US actually made decent assurances.
UK strategic thought always sought to divide the continent, which means post-WW1 relations with France cooled (it mistakenly viewed Germany as weaker). This was an impediment to A and the background to Keynes' outbursts of nonsense. It didn't like alliances (B) but would have have C.
The US couldn't abide with any version of C (fourteen points etc.) and was pulled by UK regarding A. B really needed something much stronger than League of Nations which US opinion could not abide.
Result was a strategically incoherent treaty, which France/UK/US pushing in different directions, but it wasn't actually a punishing treaty but rather a mild treaty.
P.S. The late Sally Marks had a lot of excellent writing on these issues.
"France couldn't rely on as strong allies next time"
On the other hand Austria-Hungary wasn't there either.
But okay, what would've motivated German people to go along with the elites and "seek revision and likely its [Germany's] original WW1 aims" if the Versailles treaty were softer on the country and the economic situation were better?
That's a bad trade if the alternative is the Russian Empire. Besides, in the lead to WW2, Germany also got Italy, Austria and Czechoslovakia, an assist from the USSR - while France got Poland.
>what would've motivated German people to go along with the elites
It's not that hard for elites to motivate their own people. WW1 was itself insane after all.
---
There isn't that much disagreement. Everyone agrees that the treaty left Germany dissatisfied. What the 'be softer on Germany' people miss is that the German elite didn't change by that much and that Germany wasn't so weak. By 1922 you had memos with ideas like "Poland's existence is intolerable". The problem with Versailles was the bad strategic architecture more than its not-quiet-hardness.
The other commenter explained, but I'd like to remark it is somewhat to my point that one can see that propaganda and not know the provenance. I certainly didn't a few years ago.
I don't think you believe that premise, but I do think less savvy people can be led down a road of increasingly implausible misinformation until they eventually find themselves invested in highly delusional thinking.
On HN I have seen people whining about their "innocent conservative viewpoints" being censored, but it turned out those viewpoints were that black people were intellectually inferior based on studies of IQ.
Like your counter-example elsewhere is that Facebook is a fine technology that helped your good friend find a partner. If we attribute that to Facebook, do we not also have to attribute, e.g., Myanmar? Don't situations like that prove that not only are people incapable of recognizing bad information, but also that these tools can exacerbate the effects when they are misused?
Are you suggesting we should censor IQ studies that show racial differences? That seems like a bad idea if, for example, you're interested in how good an education different racial groups are getting.
You're making decisions for an LLM company. People start using your models to write articles making those claims. Not to promote equal access to education, but to promote delusions that specific groups are genetically inferior.
Is that not what the science says though? "inferior" is a bit of a loaded word but if you take the emotion out of it, we compare all sorts of things that way. It still sounds like you want AI to hide some science that's offensive to popular opinion.
You... literally believe black people are less intellectually capable?
No, I don't think that's what the science says. To the best of my knowledge the only thing that says that was a biased study comparing children of wildly different backgrounds, and has been completely discredited by serious scientists.
I want AI to not assist people like whoever taught you to think that.
To quote Wikipedia, which is not biased in favour of race science:
> A 2001 meta-analysis of the results of 6,246,729 participants tested for cognitive ability or aptitude found a difference in average scores between black people and white people of 1.1 standard deviations. Consistent results were found for college and university application tests such as the Scholastic Aptitude Test (N = 2.4 million) and Graduate Record Examination (N = 2.3 million), as well as for tests of job applicants in corporate settings (N = 0.5 million) and in the military (N = 0.4 million).
I don't actually think it is controversial that there are IQ differences between ethnic groups. What is controversial is (a) how much IQ means, is it a good measure, etc. and (b) are the causes of these differences genetic or environmental.
Yes, the IQ difference is no longer controversial. It used to be when people supposed it might be due to cultural bias in the tests but that problem got solved, somehow.
Unfortunately, the topic is too politically sensitive for most people to talk intelligently about. When I find people who say it's entirely environmental, I usually try to ask their reasons in hope of discovering something but so far they've never come up with anything stronger than speculation and insults. As far as I know, there's no evidence that it's not genetic but I'd love to discover some. The Minnesota study I mentioned looked like it should have finally got to the bottom of it, but they made up a new hypothesis based on an uncontrolled variable after the data came in, which is bad science.
What I have found is that the whatever the cause is, it happens before the age of about 4. That's useful to know if you propose improving racial equality through better education because by the time kids go to school, it's already too late. Any such effort has to be multi-generational.
The social implications of the general society getting this wrong in either direction are huge, so it's definitely worth being honest about.
And yet when we've seen intentional attempts by neo-Nazis to have models echo their viewpoints by Grok or Gab's literal Adolf Hitler LLM the models default to telling them off as morons or disgusting.
Maybe adding a greater degree of self-alignment will yield greater gains than keeping training wheels on indefinitely.
We should probably be looking more to how humans have intrusive thoughts but have our prefrontal cortex and impulse control to keep them in check.
Rather than prevent models from generating bad things ever, it's probably better long term to have a secondary process that catches and corrects it like our prefrontal cortex.
It's a monstrously cynical vision. "Our users are too stupid not to hurt themselves. People are too stupid not to spot 'dangerous' LLM talk. The world is too stupid to wield this technology." Why are they so cynical? Why don't they have faith in their fellow man? It annoys me, but above all, it makes me sad.
Where is the optimism?