Richard Stallman Talks Red Hat, AI and Ethical Software Licenses at GNU Birthday
fossforce.com
fossforce.com
He screamed at people for using 'open source' instead of 'free software' in their questions, raved about the Amazon "Swindle" like a 12-year-old, and talked over and down to everyone. He was often accurate in the core of what he was trying to say, but his attitude was unhinged, and discouraged most people from interacting.
The 100 or so developers all left the talk in silence, mourning the loss of the hero we had imagined him to be.
To those who say we need him for GNU outreach: we absolutely don't. ANYONE would be better for the job.
I thank him for his many incredible contributions to our world, and wish him a speedy recovery, but we don't need these talented assholes to represent our community.
Stallman has done a lot, but he's doing more harm than good nowadays. It's mature to know when to step down and give the next generation space to grow.
The notion of "consequence culture" encompasses "cancel culture". It shouldn't be a surprise that bad faith actors exist, and netizens can be manipulated. To dismiss "cancel culture" is to dismiss the notion that bad faith actors exist. Of this set of bad faith actors, I think the CSJ types are by far the worst. They exploit the good reputation of liberal social justice (i.e. the original gay, female, black rights movement), and through motte and bailey, manipulate people into their version of social justice (CSJ), which is unsubstantiated, comically simplistic, discriminatory, and divisive. By masquerading themselves as purveyors of issues we care about, their bad behaviour becomes harder to call out as well, since they can misrepresent their critics as being against the issues we care about, and weaponize their compassion to mob their critics. Conflating "consequence culture" with "cancel culture" is just one of the many ways they're achieving that.
[0] https://www.youtube.com/watch?v=buarAx_u2qg
[1] https://www.amazon.sg/Canceling-American-Mind-Undermines-Thr...
I think that it’s useful to hear a balanced take on cancel culture from someone that’s not, frankly, some Hollywood figure making dumb “cancel culture” jokes because they’re secretly pooping their pants over when it’s going to happen to them.
> To those who say we need him for GNU outreach: we absolutely don't. ANYONE would be better for the job.
If that was true, that would have happened years ago considering the controversies. (Although at this point everybody might see things differently) The subtleties of open source licensing seem to interest only few people, and even less understand them fully. (I don't count myself in)
No developers were persuaded to change their language from 'open source' to 'free software' that day in 2009, him shouting at us was utterly ineffective.
We need good communicators, not just passionate people. I'm not suggesting he shouldn't write his good ideas down, but keep him away from an AUDIENCE.
His methods might've been of a lunatic indeed, but it's on you to argue that Free Software would have grown further than what it is today with a more mild mannered person at the helm.
I disagree completely with this proposition. You need crazy zealots to change the world: no one will accept their extreme view of the world, but choose a more moderate version of their message. And so the zealot has succeeded in his mission. Examples in history abound, whereas mild-mannered people don't go very far in radically changing the status quo.
As someone that grew up with a lot of respect to RMS, hearing that this was how his talk went saddens me.
It's unfortunate but the best thing FOSS can do is find more unknown people that can focus on the content.
Yes, this is frustrating. I once tried to ask him a question about a project that calls itself open-source, and rather than answering it, he proceeded to go on about how he hates the term.
RMS has been literally all about Free Software from day one. The term "open source" originated as a competing philosophy from Bruce Perens. It's like going to the CEO of Coca-cola and asking about Pepsi.
RMS was also a fringy fanatic from day one. It took a lot of willpower and vision to advocate for Free Software, a movement that that flew in the face of the overwhelming proprietary software industry that existed back in the '80s. A lesser person would have given up long ago.
If an analogy is supposed to persuade me, but the internal premise of the analogy completely contradicts the purpose of the analogy, then perhaps it should be completely eliminated or at the very least it should be challenged.
Moreover, this contradiction within the analogy highlights the exact opposite. Likely, the CEO of Coca Cola experiences that "coke/pepsi" dialog with waiters all the time and merely shrugs it off --- because that is actually a reasonable response, unlike screaming at people over a tiny, unintentional slight.
He can scream all he wants, but it's not a productive conversation.
I would equally be annoyed if someone called an Apple an Orange, because they are not the same thing.
It is up to you to convince the people that the difference is big enough that they should want to change their language. And you know how to make sure they won't want to do that? Speak down to them, paint them as stupid for not knowing and so on.
RMS spent half a sentence on crops and seeds before transitioning to emphasize that GNU/Linux isnkey and it's important to call it GNU. Most of the audience didn't know whonthat was and what he was talking about, while the interpreter tried to give a tiny bit of context in his translation.
We need fundamentalists who stick to their opinion. And I agree to many of RMS's points, but he is not a good poster head ...
> To those who say we need him for GNU outreach: we absolutely don't. ANYONE would be better for the job.
But other people are, and you are ignoring them and helping create this supposed problem, which I'm unconvinced is a real or important problem. LWN said that Panos Alevropoulos gave a great speech at this very same event. Why don't you write or promote an article about his talk and leave the people who want to read this article alone instead of heckling them and RMS. I've read this same complaint for many years "Everyone look: we should pay attention to someone else besides RMS for free software (as I also give attention to RMS through this comment, do not pay attention to anyone else, and do not suggest anyone or do anything to help solve this supposed problem)." RMS said things at this event people are interested in reading about, so this got upvoted, and there is nothing wrong with that, and the fact that you didn't like his attitude in 2009 is really not very interesting as a comment on this 2023 speech where he had a good attitude. Is RMS irredeemable? Should someone be primarily judged by mostly anonymous internet commenters who are not actively involved in the activity they criticize and based on things like a grumpy attitude they once saw? I don't think so.
You've forced me to look up the talk, and the one I have the ticket for in my email is actually from 2011, my memory of the date was wrong: https://localevents.theiet.org/register.php?event=bc7fd2
Someone took a terrible video of it (ah 2011 video) here https://www.youtube.com/watch?v=vazlMe7sNzM but that doesn't seem to include the Q&A section where he went off the rails.
Really? Let me spell it out for you. You wrote:
> The 100 or so developers all left the talk in silence, mourning the loss of the hero we had imagined him to be.
That is completely not credible on its face. It was a public event. I've been to RMS speeches, seen recordings, and I estimate there was a ~0% likelihood that all the people who attended that event considered RMS a hero before the event and lost that impression through it.
I will say that I was pleasantly surprised, given what I had read about him, at how he handled a question from someone who had a very severe intellectual disability; he treated his question at the Q&A as he treated everyone else's. Although he belittled my question, but that's neither here nor there.
IMHO RMS’s socially abrasive attitude and communication style, and the overall unapproachability of the FSF, goes a way toward ensuring that anyone that can stomach advocating for “the cause” is similarly abrasive. This is to the point where some people conflate being a jackass with supporting Free Software as a movement or even as a concept. Like, aspects of some dude called Richard’s personality and even his proclivities wrt how he chooses to go about his computing life are cargo-culled by this decentralised group of fans. You see it here all the time. Someone will be talking about Free Software, and use this as a license to be an asshole to people.
There is in my experience a large silent majority that simply won’t engage with these conversations, not because they don’t see a legitimate place for the Free Software movement, but because the culture is so unnecessarily toxic and unapproachable on account of the people that it puts on a pedestal. It certainly puts others off from the actual principles of the movement altogether.
Time was, computing was for entirely socially adjusted - primarily - women. I as much as possible try to see the period of dominant industry voices being overrun with socially awkward asshole nerds to be a blip rather than an origin story or an ongoing necessity. At this stage I think we are at the point where continuing to give these people a social platform just because of some sense of prestige is not the way to go.
I wouldn't go as far in blaming AI. It is certainly error prone and has very short autonomy. It's not better than subject matter experts in any field. But it can solve multi-step problems even when they don't look very close to anything in the training set.
I thought that as a pioneer in the free software movement, he would fully support LLMs that democratize skills and make them universally applicable. LLMs benefit from the knowledge of some and extend that help to others, much like the principles of free software. In fact AI is much more radical in making skills free than software ever was.
It seems to me he wants people to look past the impressive facade of correct grammar and see that a lot of the LLMs often spew out bullshit. Personally I've seen them perform intelligence tasks, generate correct code and so forth. But I think many people ascribe to them a general intelligence that they don't have, which is why you see them quoted increasingly often online as if they were knowledgeable or correct-by-default.
I think if he wasn't making an essentialist argument about LLMs he worded himself exceedingly poorly. Arguing against the danger of blindly believing LLMs is good, but at the same time he's destroying his own credibility on the subject by exaggerating how bad they are vs. focusing on the dangers of blind trust.
This is especially notable when he then goes on to point to measurable success as a metric for other AI tech - he'd have made a far better point if he argued that LLMs also need to be evaluated carefully on a case by case basis like these other systems, rather than be taken on trust.
The examples he gives are technologies where you can point the AI to a problem and trust the results will work as intended. You don't need to check every individual output of those AIs, you can just trust that it will be better than the output you'd get from a human expert (even though it's of course not 100%, and a combination of AI + human expert may be better).
What is the equivalent for LLMs?
No, they're technologies where AI's have been tuned and tested and validated for there domain so that we after that process know that they will work as intended sufficiently often to measurably be a net benefit, just as you yourself go on to describe. They'll still make mistakes, just like you yourself point out. But they've been tested to ensure they make few enough mistakes to be worthwhile.
> What is the equivalent for LLMs?
The equivalent for LLMs is exactly the same process. To quote myself:
> he'd have made a far better point if he argued that LLMs also need to be evaluated carefully on a case by case basis like these other systems, rather than be taken on trust.
In other words: Test them on your use case, and validate their performance on that use case. Don't assume. We wouldn't take the performance of any domain specific model on trust, and there's no more reason to take the performance of an LLM on trust for a domain we've not tested it thoroughly for, and quite possibly fine tuned it for.
The difference is that LLMs do well enough to convince some people of the idea they can skip the testing and validation step, and that point - that people are prone to give them a level of trust that they should not be given - is valid. But extending that to dismissing them as "bullshit generators" as he did is equally ridiculous.
If there is no qualitative difference, then human brains are no more than machines that predict the next word at each step. This could be the case, but instinctively I feel that it's not. You're right that it's very tricky - there's some interesting approaches, such as defining classes of problems or riddles that humans find doable but LLMs do not. The below is potentially a good candidate. Ultimately, I think a decisive test will be to see whether LLMs can take on great unsolved mathematical conjectures.
------------------------------
User
There are exactly three people in a room: Andrew, Bob, and Chris. Each of them is engaged in an activity. Andrew is reading a book by himself. Bob is playing table-tennis with someone else. What is Chris doing
ChatGPT
Given the information provided:
Andrew is reading a book by himself.
Bob is playing table-tennis with someone else.
Response: Chris is playing table-tennis with Bob.
Edit: ChatGPT 3.5 gives me the response you reported.
----------------
Imagine that I take a road trip starting from Amsterdam, visit the cities Warsaw, Budapest, Brussels, Paris, and Berlin, and then return to Amsterdam. In what sequence should I visit those cities in order to minimise the total driving time?
----------------
https://www.google.com/maps/dir/Amsterdam,+Netherlands/Bruss...
https://www.google.com/maps/dir/Amsterdam,+Netherlands/Bruss...
RMS worked on early AI research at the MIT Artificial Intelligence lab in the 1970s.
LLMs are fundamentally different than many of those early AI systems in both goals and how they work.
It is likely that he views them through a different lens and sees their shortcomings in sharp relief compared to the early AI systems he worked on.
[1] https://m.youtube.com/watch?v=vUwQLsWjA-g&pp=ygUQcmljaGFyZCB...
As someone who was enthusiastic about AI before neural networks rose to prominence, I only shake my head when I hear people ascribe all sorts of properties of intelligence to the particular trend in AI that is in vogue today.
The ultimate quest of AI was to understand how to make good decisions. The idea was that maybe if we emulate human activity that humans call "intelligence", but remove all the randomness, unintended errors etc, then we can understand what it means to be intelligent and subsequently will be able to attack hard problems where humans fail due to things like the size of the problem, or the lack of intuition for a particular domain etc.
The neural-net AI abandoned all hopes of understanding of how intelligence works and concentrated its efforts on mimicking a typical human response by asking a lot of people the same question, and then averaging the answers. It inspires no confidence that it will be able to apply to domains where humans themselves struggle (eg. multidimensional geometry) or that it will ever become an authoritative source for any kind of answers (i.e. would produce a proof why its answers are correct).
To anyone who wanted to know how intelligence works neural-net AI is worthless. That's not to say that such AI is worthless in general, clearly, it does very well predicting complex patterned behavior, or can sift through a lot of data with difficult to deterministically determine patterns, but it isn't even going in the direction anyone who wanted intelligence would be interested in.
I think that as with many historically vague terminology, we'll have to redefine what "intelligence" means. Split this concept into multiple concepts, and, maybe the neural-net group will win the right to call themselves "AI", while those who I describe above will switch to a different term, or maybe it will be the other way around. It's clear though that the modern in-vogue approach is simply a waste of time for a lot of people interested in AI.
I don't claim to have a good knowledge of the ins and outs of the particular trend of the day though. I wouldn't be able to tell what the difference is between "deep learning" and "reinforcement learning", not with any authority or confidence. I did take two courses on AI though, long before the neural-network approach was a thing. I.e. my AI learning experience was focused on search algorithms, building knowledge bases etc.
So, here, based on my experience of practical applications of LLMs, I can comment on things you wrote:
> Plus there is a ton of work to now combine LLMs with their excellent language abilities to more formal systems based on rules
Tons of publications doesn't mean this is a solved problem. As it stands now, there's a serious doubt this is a problem that can be solved. I have not seen any real system working in the world of medicine which would use LLMs for anything that can be encoded as rules. If anything, I've heard a lot about failures to implement anything remotely similar to this (which happened outside of academic medicine, or somehow slept through the cracks). I.e. chatbots suggesting suicide to people in distress, or suggesting to cut on carbs to young women struggling with anorexia and so on.
> the ability to handle language, that previous approaches were all extremely bad at
Well, now it's my turn to point fingers. You probably don't know much about previous approaches, and don't understand the problem well. It's true that since the term AI was coined substantial efforts went into understanding natural languages. But what came out of this work is a disillusionment, the understanding that natural languages suck at encoding knowledge and as such are a bad tool for building ontologies, or call them knowledge bases. So, it wasn't because older methods failed to parse language well. It was because we realized that there's no point in being good at it. It's a cute gimmick, if you can do it, and can be useful in harmless situations, like enabling rich interactions between NPCs and PCs in video games. But, if your goal is to understand how to make good decisions, it's more of an obstacle really.
Like I said previously, but you seem to have ignored: yes, neural networks are useful, but not in the way a lot of people understand intelligence, or, at least, they fail to address the central and the critical questions about intelligence. Compounded by dishonest advertisement of their abilities and armies of low-quality practitioners this approach has a tendency to alienate people who are interested in more fundamental questions about the nature of intelligence.
That feels like goalpost moving. Having it be interpretable is a useful quality but it’s not the same as intelligence
What I'm saying is that "intelligence" didn't mean "guessing the best answer by averaging many known, externally generated answers" before neural network approach became popular. Intelligence was always about finding good solutions to problems.
In rare cases doing nothing is a good solution to a problem, in less rare cases, statistically guessing is a good solution, but it still doesn't answer the question of how to find a good solution. No amount of improving the accuracy of statistical guess will answer this question because what we need is a proof that the claim is correct, but the guess is the opposite of a proof -- it's asking you to take the claim on faith.
In practical terms, compare these two scenarios:
1. A neural net AI is diagnosing breast cancer in a patient. Suppose, it was a male patient (a very rare occurrence, that's likely to fail statistically guessing AI). AI fails to find cancer and the patient dies.
2. A radiologist diagnosing breast cancer in a patient. Same circumstances as before. The radiologist fails to identify cancer, the patient subsequently dies.
In the first case, we don't even expect the AI to be correct. Nobody's taking the blame, because, well... we've constructed this AI to be guessing statistically. It couldn't have enough data to guess reliably because in order to do so it needed to see more breast cancer in male patients than we have data for.
In the case of a radiologist, we expect the radiologist to understand that they are treating an exceptional case. We will blame them, and probably sue for malpractice because we expect them to be able to reason, to be able to explain why what they see "makes sense" and why their response to what they saw "made sense". We will hold them responsible for failing this patient.
In common understanding of the word "intelligence", the radiologist is intelligent (or, at least, we expect them to be), while neural net isn't.
Just 12 months ago if you polled here you would find much more naive opinions. But now we are learning fast to evaluate AI and human work critically. We have had our hundred hours of AI play and all tried to make it stumble, then to make it work, and in the end we are wiser for it.
Regular citizens even know not to trust the bunch of benchmark scores every new model shoves in our eyes, like hardened ML researchers. And basic AI concepts are infiltrating and getting root in the public - model, parameters, tokenisation, datasets, fine-tuning, vector search, and their complex interplay.
His take may be a bit jaded from his own experience.
He worked as a Lisp hacker at the MIT Artificial Intelligence lab when AI research was peaking in the 1970s.
The commercialization of Lisp and early AI effectively destroyed the hacker paradise he worked in and the following "AI Winter" devastated the research and its funding.
These experiences lead him to start the GNU project, so the net result was good for software, but definitely colored his perspective.
See https://en.wikipedia.org/wiki/Richard_Stallman#Harvard_Unive... for more details.
I think governments were first in this space of “engineering people’s brains to be ideal victims” and will remain active in that long after any regulation is crafted to limit individuals and corporations from doing the same. Engagement optimization seems like a thing that can be regulated, but it’s just the same thing tv stations have been doing forever, basically marketing. I think really the problem is a different one. People need better education especially when it comes to ethics. Psychological experiments (a/b testing for example) should be governed by boards that oversee the ethics of experimenting on human subjects. Stallman identifies the problem as “popularity” but what if popularity required behaving ethically and not doing what is incontrovertibly unethical human subject research for personal greed.
On the other hand you will be able to get adults interested, there are likely plenty of us out there that are completely jaded from using proprietary software most of their lives. The ideals of freedom, privacy and openness are still valued by some and they should be the core of the recruitment message.
I chuckled ;-)