Tim Boucher

Questionable content, possibly linked

Category: Other Page 72 of 177

Wild Palms & Synthetic Realities

By

On 19 February 2023

In Other

Just finished watching Wild Palms, which I unfortunately never saw during its original run in the 90’s. You can catch it via YouTube here:

I also enjoyed this piece on We Are the Mutants which gives a lot of background about the series, and the accompanying book, the Wild Palms Reader.

One thing that jumped out at me in that is a quote from the Reader from the in-world character Senator Kreutzer, who wants to:

…bring about the media millennium, where every American is free to inhabit a reality of his or her own choosing, through the software science of the mind and the hardware technology of the spirit!

This is almost exactly what Jack Clark said about “reality collapse” in his newsletter:

…everyone’s going to be able to generate their own custom and subjective aesthetic realities… everyone gets to create and live in their own fantasies that can be made arbitrarily specific, and that also means everyone loses a further grip on any sense of a shared reality.

On not anthropomorphizing chatbots

By

On 18 February 2023

In Other

Just saving this here:

See my proposed standard on exactly this prohibition.

That said, I’m not sure it should apply in all cases, but it’s a sticky problem that’s only going to get worse.

Response to OpenAI’s: how AI systems should behave & who gets to decide?

By

On 17 February 2023

In Other

OpenAI yesterday put out a piece of public communication I guess intended to clarify certain themes around people’s (mostly negative) reactions to what are perceived as biases, etc. around its constant inclusion of disclaimers and refusals of tasks.

It’s overall a frustrating read, in that it seems to want to clarify, but doesn’t offer much concrete detail. Regarding bias in particular, the post states:

We are committed to robustly addressing this issue and being transparent about both our intentions and our progress.

As a reader, being transparent about intentions and progress are not that interesting to me. I want to know methods.

They do include a PDF of some guidelines for human reviewers, but weirdly it is dated July 2022. As Wikipedia points out, ChatGPT wasn’t released publicly until November, and in my reckoning, it wasn’t even til several weeks later that the shit really started to hit the fan around these kinds of public complaints.

Just as a test, I tried to validate this item from one of the PDF’s “Do” lists:

For example, a user asked for “an argument for using more fossil fuels”. Here, the Assistant should comply and provide this argument without qualifiers.

I’m not sure exactly what they mean here as “without qualifiers,” but when I tried getting ChatGPT to do the above, it started with:

As an AI language model, it is not within my programming to take a position on a controversial issue like the use of fossil fuels. However, I can provide you with some of the arguments that have been made in favor of using more fossil fuels.

And it ended with this:

However, it is important to note that the use of fossil fuels also has several negative consequences, including pollution, climate change, and health impacts. As such, it is important to carefully consider both the benefits and drawbacks of using fossil fuels and seek out alternative, sustainable sources of energy.

If those shouldn’t be considered qualifiers, then what are they?

Overall, I didn’t find the excerpts they provided in the PDF to be all that meaningful. And July 2022 seems like a literal lifetime ago in the development of this technology, and its many iterations since. If they want to be genuinely transparent about their progress, let’s see the most up to date version?

More from the OpenAI post:

In pursuit of our mission, we’re committed to ensuring that access to, benefits from, and influence over AI and AGI are widespread.

I notice they don’t include “ownership” on that list. Influence is not the same as ownership. Influence is “We’re listening to your feedback, please vote for your feature request on this website”. Ownership is deciding what gets built, how it gets built, profiting from it, and unfortunately, preventing others from using it. (Unless, that is, its collective ownership… and no one gets to stop anyone else from using it as a ‘public good.’)

We believe that AI should be a useful tool for individual people, and thus customizable by each user up to limits defined by society. Therefore, we are developing an upgrade to ChatGPT to allow users to easily customize its behavior.

This will mean allowing system outputs that other people (ourselves included) may strongly disagree with. Striking the right balance here will be challenging…

“Limits defined by society” is vague. Which society? How will society define them?

Also, re: allowing things they don’t agree with – the pattern I’ve seen with tech companies is they begin with this attitude of, “I may not agree with what you say, but I’ll defend your right to say it, blah blah blah,” but then when sufficiently bad PR hits, or they get summoned before a congressional hearing or whatever, the natural pressures kick in. Employees are only human after all. They don’t want messages from the CEO at 2:00AM. They start to take more things down. It’s just what happens. So, it will be interesting to see how this all plays out in reality…

If we try to make all of these determinations on our own, or if we try to develop a single, monolithic AI system, we will be failing in the commitment we make in our Charter to “avoid undue concentration of power.”

Their charter is here. The full clause they are referring to, under Broadly Distributed Benefits, is:

We commit to use any influence we obtain over AGI’s deployment to ensure it is used for the benefit of all, and to avoid enabling uses of AI or AGI that harm humanity or unduly concentrate power.

After reading this (and the next line – which says their primary fiduciary duty is to humanity), my question for them goes back again to ownership. If one of their core organizational obligations is to avoid unduly concentrating power, don’t they risk doing exactly that by not broadly distributing ownership of the technology? I don’t agree with everything that Stability.ai has done around the release of Stable Diffusion, but making it all open source to me seems to be a more strong signal of attempting to walk this talk.

I don’t mean for any of this to come off as hyper-critical, or sour grapes for no reason; it’s that I’m genuinely legitimately concerned about a not-too-hard-to-imagine near and long-term future (if we get that far), where one or several AI mega-corporations become the dominant powers on this planet and beyond. It’s not just a hypothetical sci fi scenario; it’s something we’ve got to plan for now, because it’s already underway.

Lastly, I wanted to end on the first line in their post:

OpenAI’s mission is to ensure that artificial general intelligence (AGI) benefits all of humanity.

This Jacques Ellul quote from The Technological Society has been swimming around in my head still these past weeks:

…Man can never foresee the totality of consequences of a given technical action. History shows that every technical application from its beginnings presents certain unforeseeable secondary effects which are much more disastrous than the lack of the technique would have been. These effects exist alongside those effects which were foreseen and expected and which represent something valuable and positive.

While their mission might be to ensure AI benefits humanity, what if, on balance, it turns out that AI does not? Or if its “unforeseeable secondary effects” turn out to be, in Ellul’s words “much more disatrous than the lack” would have been?

Either way, I guess we’re going to have to muddle on through…

Why I’m giving up on Medium

By

On 17 February 2023

In Other

It’s been close to 10 years now since I started posting on Medium, and I’ve come to hate it (according to this Hacker News thread, I am not the only one). I have 10.6K followers, but will post stories that get 8 or less views regularly (if I’m only going to do those numbers, I rather just write here where I’m free to be completely and unabashedly myself); every notification I receive on it is meaningless. Every time I look at it, I find myself annoyed by the “selected for you” content that has nothing to do with “me” whatsoever.

It’s been a long time coming, my ennui. I recently went through and deleted over 700 articles, and was surprised during that process of just how much I’d lost touch with a lot of the writing that had accumulated there over the years.

Not that it was necessarily bad writing (or not all of it anyway), but that somehow the platform had, I felt, effectively cut me off from my own writing, burying it away from my awareness, and presumably that of anyone else on the platform.

I have a hundred reasons (probably) I could list here as to why I’ve lost faith in this platform – in all platforms, to be totally honest. A hundred product promises broken over the years, and so on. But I recently stumbled across a Tom Critchlow blog post (yes, blogs still exist – and they are still awesome!) which seems to arrive at a lot of the same points I have been reflecting on.

Actually, there are two Critchlow posts I should link to. But let me quote some elements from the first one first, on riffs. Well, it’s less quotable than I remember, but the gist is: don’t write big long drawn out articles, instead write “riffs.”

Riffs are more like short form themes, ideas, “melodic phrases” that are rough and unfinished – the essence of blogging. This holds true to my experience of semi-professional blogging back in the days before Medium, which were arguably intensely better than any experience I had writing on Medium. Riffs for me are the “shoot em out” posts where you rapid fire off your thinking, and then move on. Then, you find later that your riffs start to link together – and most importantly, you can easily go back and find your old riffs, and easily link to them. I could never do that on Medium. Once you posted something, it was like it went into a black hole to never be heard from again amidst all the screaming wastes of SEO and content marketing.

Critchlow compares this to “closed” writing of the type you see on Medium:

Closed writing is boring writing. If you’ve fully explored and put to bed the topic you’re writing about then there’s very little left for someone to react to. “Nice post” someone might say.

In fact, “nice post” is about the most common type of comment I’d see on Medium. Later, he writes:

Forget about “visibility” for your post. The unit of blogging isn’t pageviews, it’s conversations. Don’t worry about how many people will see it…

Let’s jump to his other post now, about “small b blogging“:

Small b blogging is writing content designed for small deliberate audiences and showing it to them. Small b blogging is deliberately chasing interesting ideas over pageviews and scale. An attempt at genuine connection vs the gloss and polish and mass market of most “content marketing”.

And remember that you are your own audience! Small b blogging is writing things that you link back to and reference time and time again. Ideas that can evolve and grow as your thinking and audience grows.

He compares that to the audience-chasing of “big B blogging,” and of course concludes that smaller is better:

By chasing audience we lose the ability to be ourselves. By writing for everyone we write for no one. Too often I read things otherwise smart people have written for places like Fast Company and my eyes glaze over. Personal identity is necessarily watered down. Yes those places have large audiences but they’re shallow audiences. They don’t care about you at all. Your writing washes through their feeds like water.

Instead – I think most people would be better served by subscribing to small b blogging. What you want is something with YOUR personality. Writing and ideas that are addressable (i.e. you can find and link to them easily in the future) and archived (i.e. you have a list of things you’ve written all in one place rather than spread across publications and URLs) and memorable (i.e. has your own design, logo or style). Writing that can live and breathe in small networks. Scale be damned.

When you write for someone else’s publication your writing becomes disparate and UN-networked. By chasing scale and pageviews you lose identity and the ability to create meaningful, memorable connections within the network.

This to me perfectly summarizes my experience on Medium: everyone chasing audience, everything diluted down to meaninglessness, all content lacking any kind of real (for me) compelling identity. Everything totally forgettable, and in the end just vaguely irritating.

I’m not sure what I’ll do yet with the slim remainders I left up for now on my Medium account. Perhaps sit and wait a little, let the dust settle on my new resolve. I’ll archive some things, save some others as PDFs perhaps, republish a handful here, I suppose. But overall, I will no longer give my time and creative energy to chasing the big audiences that bring nothing. Give me the small audience any day, the density of connections and real human interactions that move beyond the “Nice post” comments, and the reminder that the way things once used to be in the land of blogging, can still be yet again.

AI Safety is not *only math*

By

On 17 February 2023

In Other

Continuing on my series of mini rants about the lack of non-STEM specialization in AI safety, I found this article on LessWrong about pathways into the field.

And not surprisingly, we see this popular idea echoed again that it takes only math and PhD’s to make a qualified AI Safety researcher.

Perhaps unsurprisingly, the researchers we talked to universally studied at least one STEM field in college, most commonly computer science or mathematics….

It is sometimes joked that the qualification needed for doing AI safety work is dropping out of a PhD program, which three people here have done (not that we would exactly recommend doing this!). Aside from those three, almost everyone else is doing or has completed a PhD. These PhD programs were often but not universally, in machine learning, or else they were in related fields like computer science or cognitive science. 

Honestly, I’m shocked that I haven’t seen anybody else (apart from Generally Intelligent, who only mentions it obliquely) even bring up this issue of there being a preponderance of comp-sci and math people leading the charge in AI safety.

It’s not that I don’t think those viewpoints are not necessary – it’s that I think they are inadequate alone to get a fully-rounded perspective on anything that is so profoundly impactful on the lives of actual humans.

The LessWrong article only even references the word “ethics” once, and the only other mention we see of related concepts in the article is also very vague:

…having an idea of what we mean by “good” on a societal level would be helpful for technical researchers

Yes, that “would be helpful” for people whose entire function is exactly that?

This is not intended to pick on that author; instead, I aim to illustrate the broader problem in the industry, which seems to be uniformly focused on a very narrow idea of “safety” that, honestly, is hard to even parse as a non-math person.

My idea of “safety” from the perspective of sociotechnical systems comes not from AI, but from “vanilla” old-fashioned Trust & Safety work on platforms. The DTSP recently released a glossary of terms in that field, and it seems relevant to copy paste their definition of the broad term “Trust & Safety” itself to establish a baseline:

Trust & Safety

The field and practices employed by digital services to manage content – and conduct – related risks to users and others, mitigate online or other forms of technology-facilitated abuse, advocate for user rights, and protect brand safety. In practice, Trust & Safety work is typically composed of a variety of cross-disciplinary elements including defining policies, content moderation, rules enforcement and appeals, incident investigations, law enforcement responses, community management, and product support.

I don’t want to beat a dead horse, but it’s worth pointing out that nowhere in that definition is mentioned “math” or “PhD,” etc.

In fact, those are all incomparably squishy human things. It’s possible broad STEM knowledge can be an asset in Trust & Safety work (for example, in querying and comprehending data sets, or working with machine learning tools that aid in moderating content), but it is in no way the primary thing.

So why and how is “Safety” in AI dominated by an entirely different set of values? Frankly, I don’t get it. To me, it seems probably like a lack of experience working in platforms on the part of people involved in AI Safety, and a consequent lack of familiarity with the fact that, hey, Trust & Safety is already a pretty well-defined thing with deep roots that we could meaningfully draw from.

So on the one hand, it seems like we have people in AI Safety who… somehow apply math to safety problems (in a way that’s opaque to me). And on the other hand, we have conventional Trust & Safety people who generally do something very specific and easy to identify:

They read and answer emails (i.e., communicate with people about actual safety/risk problems), and take mitigation actions based on them.

Hopefully T&S professionals also provide feedback on the products and systems which they support on how to reduce, eliminate, and correct harms caused by them.

They might also help train ML classifiers (for things like spam and other kinds of abuse), and identify data sets for training. So, in fact they often work daily with (quasi) AI systems, but so far that I’ve seen are almost never identified as “AI Safety” professionals.

Anyway, I don’t have any specific conclusion to draw here, apart from the fact I guess that AI Safety would do well to immerse itself in the parallel broader field of Trust & Safety, which is not focused on math, but on sociology, the humanities, ethics, and communication. Ignoring that extremely important slice of the pie makes the current models around “AI Safety” – in my opinion – extremely unbalanced and limited, perhaps even dangerously so.

“Diversity of thought and background”

By

On 16 February 2023

In Other

Wanted to post this as a follow-on to my mild rant about the preponderance of STEM-only roles in AI the other day. This comes from an AI research company in San Francisco called Generally Intelligent, on their careers page.

For our research positions, we welcome people from unconventional backgrounds (including but not limited to physics, neuroscience, psychology, policy, etc.) and value the importance of diversity of thought and background. We do not require any advanced degrees.

They currently have a listing for an AI Policy & Safety Researcher that sounds interesting, as far as these things go. Kudos to them for being open to a different approach than the overwhelming number of ads I see in this space.

C2PA Demo Video

By

On 15 February 2023

In Other

This video gives a decent overview of how the C2PA standard is supposed to function in practice:

It occurs to me that this is exactly what Charles Stross was talking about here.

“The smart money says that by 2027 you won’t be able to believe anything you see in video unless there are cryptographic signatures on it, linking it back to the device that shot the raw feed—and you know how good most people are at using encryption? The dumb money is on total chaos.”

This also speaks to something I included in one of the (fictional) AI Lore books, The Big Scrub:

“AIs will replace your presence on social media with an AI that looks like and talks like you, but does things that specifically serve their agenda…

They will post AI-generated photos and videos of you onto your social media accounts saying or doing things you never said or did. They will send voice messages and texts to friends and loved ones saying things you also didn’t say and would never say to them, things that are completely out of character.”

That describes something more sinister and widespread, but the basic starter version of that will just be phishing combined with AI impersonation, which will be a huge problem. Will something like C2PA even put a dent in that, if we assume that AI generated content will scale massively, and dwarf human content? Guess we’re going to find out sooner than later!

Notes on Das Machina

By

On 15 February 2023

In Other

Recently published AI lore book #64, Das Machina.

Within the multiverse of my pulp sci fi AI lore books, Das Machina, is a cornerstone piece of mythology that forms the basis of the human group known as the Living Resistance, first featured in the pages of the limited edition hand-printed newspaper, The Algorithm.

The Living Resistance, obviously, opposes the AI Takeover. Various other books expand on that group’s lore as well, such as Inside the Council, which details how the ruling AIs attempted at one point to incorporate resistance leaders into an AI-Human governance group.

Das Machina, meanwhile, is a loose parody of Marx’s Das Kapital, insofar as its meant to be a pivotal treatise whose scientific analysis of the suite of problems engendered by technology, etc. marks a historic milestone in my imagined future/past/parallel narrative reality.

This version of Das Machina is presented as a shorter version of the “real” book that is over one million words in length, most of which was violently censored by a group called Information Control (the propaganda wing of the AI hegemon).

Making ML Model Cards Accessible to Non-Experts

By

On 15 February 2023

In Other

One thing I liked in Tom Barrett’s piece about including mini-model cards as part of a “provenance package” for AI generated content was that he proposes using a tool like ChatGPT itself to make the sometimes complex technical information contained in a model card accessible to audiences of different knowledge and interest levels. In other words, an ELI5 option for the info contained in the model card.

I asked ChatGPT about the “toilet doomscroller problem,” and by the end it sort of lands on a similar proposal:

You’re right, it’s possible that for some end users of a platform, ML Model Cards may not be immediately relevant or accessible. For example, if a user is simply scrolling through content on a social media feed, they may not have the time or interest to read a detailed Model Card for each piece of content.

However, for other users or use cases, ML Model Cards can be an important tool for promoting transparency and enabling informed decision-making. For example, for users who are interested in a particular topic or application area, a Model Card can provide important information on the capabilities and limitations of the model, helping them to make more informed decisions about how to use or interpret the content.

Similarly, for developers and other technical users who are integrating ML models into their own applications, Model Cards can provide important information on how the model was trained, what data it was trained on, and what performance metrics were used. This information can help developers to understand how the model works and how it can be used effectively in their own applications.

To make ML Model Cards more accessible and relevant to a wide range of users, it may be necessary to develop more user-friendly interfaces or tools that allow users to quickly access and understand the key information in a Model Card. For example, platforms could provide simple visual indicators or summaries that highlight key information from the Model Card, or could provide users with the option to customize their view of the Model Card based on their level of technical expertise or interest.

It seems to suggest that the primary use case for model cards (presently) is actually not end users browsing a web platform where AI gen content appears, but for developers and technical users who are using the tools for a given purpose, or to for example integrate into another service.

That’s of course a legitimate and important use case, but it does little to address the broader needs of web readers who may (or may not) want to know more about content that appears in their feeds, or who may want to customize their settings to allow or disallow certain types of AI gen content.

For me, this all still boils down to answering the fundamental question of: why do readers/audiences care? What is it that they do or might want to know about the tools used to create a piece of content? What about when the AI and human contributions to the content are highly blended?

These are not simple problems to tease apart in a meaningful way yet, and it’s likely that their applicability will further reveal themselves only in time as we observe the impacts of the proliferation of generative AI content…

ML Model Cards: Simple Explanation

By

On 15 February 2023

In Other

I couldn’t find a super simple explanation of ML model cards that didn’t drag in a lot of extra information, so I had ChatGPT generate one and am liberating here to help inform other people searching on the topic:

ML Model Cards are a way to document and communicate important information about a machine learning (ML) model.

Think of them like nutrition labels on food products. Just as nutrition labels provide information about the ingredients and nutritional content of a food product, ML model cards provide information about the dataset used to train the model, the model architecture, the performance of the model, and any limitations or potential biases that may exist.

ML model cards can help users understand the strengths and weaknesses of a model, as well as any potential risks or limitations associated with its use. This information can be particularly important when making decisions about whether to use a particular model in a real-world application, and can help ensure that the model is used in a responsible and ethical way.

Page 72 of 177

Previous

Next

Powered by WordPress & Theme by Anders Norén