In an age of LLMs, is it time to reconsider human-edited web directories?
Back in the early-to-mid '90s, one of the main ways of finding anything on the web was to browse through a web directory.
These directories generally had a list of categories on their front page. News/Sport/Entertainment/Arts/Technology/Fashion/etc.
Each of those categories had subcategories, and sub-subcategories that you clicked through until you got to a list of websites. These lists were maintained by actual humans.
Typically, these directories also had a limited web search that would crawl through the pages of websites listed in the directory.
Lycos, Excite, and of course Yahoo all offered web directories of this sort.
(EDIT: I initially also mentioned AltaVista. It did offer a web directory by the late '90s, but this was something it tacked on much later.)
By the late '90s, the standard narrative goes, the web got too big to index websites manually.
Google promised the world its algorithms would weed out the spam automatically.
And for a time, it worked.
But then SEO and SEM became a multi-billion-dollar industry. The spambots proliferated. Google itself began promoting its own content and advertisers above search results.
And now with LLMs, the industrial-scale spamming of the web is likely to grow exponentially.
My question is, if a lot of the web is turning to crap, do we even want to search the entire web anymore?
Do we really want to search every single website on the web?
Or just those that aren't filled with LLM-generated SEO spam?
Or just those that don't feature 200 tracking scripts, and passive-aggressive privacy warnings, and paywalls, and popovers, and newsletters, and increasingly obnoxious banner ads, and dark patterns to prevent you cancelling your "free trial" subscription?
At some point, does it become more desirable to go back to search engines that only crawl pages on human-curated lists of trustworthy, quality websites?
And is it time to begin considering what a modern version of those early web directories might look like?
You know why some blind people are really leaning into AI to fix accessibility issues? No, not like overlays that probably barely have any if/else statements in them, let alone AI, but stuff like Be My Eyes, and gasp screen recognition in VoiceOver for iOS? Because shit sucks, and it's sucked for the last 40 years of computing history for blind people. That's why whenever we get even a bit more light, even if 20% of what an AI says is fake, that 80%, that gives us 80% more info than we didn't have before. And yeah, we should all, every single one of us, know that AI can give false info by now. Hell, Mastodon folks have been shoving that into our ears with an oversized cue tip since the day ChatGPT came out. We get it. But hot damn, being able to point my phone out the bus window and take pictures as I'm going to work, hearing about a fire station, or a house with a dog in the yard, or that it's a sunny, clear, nice day outside even, is really freaking nice. And sure, maybe it's not a firestation. Maybe it's a courthouse, or a post office, or something else. but it's something that I would never have known before. Because I don't have some sighted person telling me about what's around, and I wouldn't want any other human to have to do that for me. Like, this is the thing. In order to get 100%, perfect info, I'd have to hire another human who, all they do is look around and tell me in extreme detail, what's around me? Now, sighted people of Fedi, would you want that job? Maybe for a day. Maybe for a week. But months of that? I doubt it. And that is where AI comes in. No, it ain't perfect. And the more you deviate from its training data, the less accurate it gets. And maybe eventually we'll get to a point in the middle of what VoiceOver Recognition is, and what LLM's are. But I'm just getting tired of this OMG AI is the end of the world rhetoric. It's really getting old.
We call it AI because no one would take us seriously if we called it matrix multiplication seeded with a bunch of initial values we pulled out of our asses and run on as much shitty data as we can get our grubby little paws on.
I am really, really, REALLY irritated by what I just saw. The #ImageDescription function of Microsoft's #Bing is outright lying to people with vision impairments about what appears in images it receives. It's bad enough when an #LLM is allowed to tell lies that a person can easily check for veracity themselves. But how the hell are you going to offer this so-called service to someone who can't check the claims being made and NEEDS those claims to be correct?
How long till someone gets poisoned because Bing lied and told someone it was food that hasn't expired when it has, or that it's safe to drink when it's cleaning solution, or God knows what? This is downright irresponsible and dangerous. #Microsoft either needs to put VERY CLEAR disclaimers on their service, or just take it down until it can actually be trusted.
I've been doing a bit more experimenting with #LargeLanguageModels#LLM and truth, and I've got an interesting one.
my experimental design was that I'd start asking about relationships between European monarchs, and then start introducing fictitious monarchs, but I didn't get that far...
#Amazon releases details on its Alexa #LLM, which will use its constant surveillance data to "personalize" the model. Like #Google, they're moving away from wakewords towards being able to trigger Alexa contextually - when the assistant "thinks" it should be responding, which of course requires continual processing of speech for content, not just a word.
The consumer page suggests user data is "training" the model, but the developer page describes exactly the augmented LLM, iterative generation process grounded in a personal knowledge graph that Microsoft, Facebook, and Google all describe as the next step in LLM tech.
We can no longer think of LLMs on their own when we consider these technologies, that era was brief and has passed. Ive been waving my arms up and down about this since chatGPT was released - criticisms of LLMs that stop short at their current form, arguing about whether the language models themselves can "understand" language miss the bigger picture of what they are intended for. These are surveillance technologies that act as interfaces to knowledge graphs and external services, putting a human voice on whole-life surveillance
People keep telling me that #ChatGPT is amazing for proofreading text and improving scientific writing.
I just gave #GPT4 a section of a grant proposal and it made 11 suggestions, none of which were worth keeping (often adding or removing a comma, or repeating a preposition in a list).
More interestedly, a number of its suggestions were identical to my originals.
Welp, here we go. Be My Eyes is so great, people are just... Calling their projects Be My Eyes web apps now. Cause it's 2023 and here we are. No, I haven't tried this thing. I hope it's worth the absolute shitstorm of a hype train this will cause in the blind community once people get to looking at it.
Mise à jour par rapport à mes vidéos précédentes du printemps.
Petit essai en vidéo de #Openchat3.5, un nouveau modèle de #LLM libre utilisable avec #llama.cpp, et qui prétend être comparable à GPT 3.5 Turbo, l'avant-dernier modèle d'OpenAI.
La vidéo est aussi faite pour montrer ce que ça donne en rapidité sur un processeur "normal", sans usage du GPU.
Note that the training data heavily relies on the Bible and its translations. Lots of bias there.
Meta unveils open-source #AI models it says can identify 4,000+ spoken languages and produce speech for 1,000+ languages, an increase of 40x and 10x respectively.
A thing that bugs me wrt to #AI, #LLM, whatever pattern synthesis tech you want:
BINARY THINKING
I understand the hype is exhausting.
But this thing where someone plugs their ears and repeats reductive catechisms about AI is just as silly. The common trope lately is comparing it to Eliza.
The truth is somewhere in between, unevenly distributed across use cases.
Still, when this stuff works well it works well in life-changing ways.
People who believe in all of the #ai hype have the same problem as people who live in denial of the technology being as powerful as it is: Thinking the purpose of #generativeAI is to create unique works
Consider a programmer who has lost their hands. An AI tool could be made using a #llm to generate keystrokes based on what the programmer says
User: “Clear the terminal”
GPT: generates clear command
That is the power of #gpt that everyone who knows what they’re talking about is excited for
The teacher profession has lately been hard in the US, and is going to be made even harder by LLMs. I reject the article's comparison with calculators, these are exact and you need to know what to ask before getting a useful answer from them. On the contrary LLMs satisfy neither of these propositions by accepting arbitrary prompts and outputting only plausible answers which might be useful or not.
I believe the introduction of accessible LLMs will further the divide between privileged students who will reap the benefits of homework vs the others who will use free LLM tools to skip homework, a cheap short-term win that will end up costing them in the long run. #education#LLM
♲ piaille.fr/@eglantine/11070143…
i’ll say it — #LLMs can and will spit out any topic they’ve been trained on
an absurd amount of research is going into preventing the #LLM from explaining how to make a bomb, when they could just do some dumb tricks and remove the “how to make a bomb” manuals from the training corpus.
The racism in chatGPT we are not talking about....
This year, I learned that students use chatGPT because they believe it helps them sound more respectable. And I learned that it absolutely does not work. A thread.
A few weeks ago, I was working on a paper with one of my RAs. I have permission from them to share this story. They had done the research and the draft. I was to come in and make minor edits, clarify the method, add some background literature, and we were to refine the discussion together.
The draft was incomprehensible. Whole paragraphs were vague, repetitive, and bewildering. It was like listening to a politician. I could not edit it. I had to rewrite nearly every section. We were on a tight deadline, and I was struggling to articulate what was wrong and how the student could fix it, so I sent them on to further sections while I cleaned up ... this.
As I edited, I had to keep my mind from wandering. I had written with this student before, and this was not normal. I usually did some light edits for phrasing, though sometimes with major restructuring.
I was worried about my student. They had been going through some complicated domestic issues. They were disabled. They'd had a prior head injury. They had done excellent on their prelims, which of course I couldn't edit for them. What was going on!?
We were co-writing the day before the deadline. I could tell they were struggling with how much I had to rewrite. I tried to be encouraging and remind them that this was their research project and they had done all of the interviews and analysis. And they were doing great.
In fact, the qualitative write-up they had done the night before was better, and I was back to just adjusting minor grammar and structure. I complimented their new work and noted it was different from the other parts of the draft that I had struggled to edit.
Quietly, they asked, "is it okay to use chatGPT to fix sentences to make you sound more white?"
"... is... is that what you did with the earlier draft?"
They had, a few sentences at a time, completely ruined their own work, and they couldnt tell, because they believed that the chatGPT output had to be better writing. Because it sounded smarter. It sounded fluent. It seemed fluent. But it was nonsense!
I nearly cried with relief. I told them I had been so worried. I was going to check in with them when we were done, because I could not figure out what was wrong. I showed them the clear differences between their raw drafting and their "corrected" draft.
I told them that I believed in them. They do great work. When I asked them why they felt they had to do that, they told me that another faculty member had told the class that they should use it to make their papers better, and that he and his RAs were doing it.
The student also told me that in therapy, their therapist had been misunderstanding them, blaming them, and denying that these misunderstandings were because of a language barrier.
They felt that they were so bad at communicating, because of their language, and their culture, and their head injury, that they would never be a good scholar. They thought they had to use chatGPT to make them sound like an American, or they would never get a job.
They also told me that when they used chatGPT to help them write emails, they got more responses, which helped them with research recruitment.
I've heard this from other students too. That faculty only respond to their emails when they use chatGPT. The great irony of my viral autistic email thread was always that had I actually used AI to write it, I would have sounded decidedly less robotic.
ChatGPT is probably pretty good at spitting out the meaningless pleasantries that people associate with respectability. But it's terrible at making coherent, complex, academic arguments!
Last semester, I gave my graduate students an assignment. They were to read some reports on labor exploitation and environmental impact of chatGPT and other language models. Then they were to write a reflection on why they have used chatGPT in the past, and how they might chose to use it in the future.
I told them I would not be policing their LLM use. But I wanted them to know things about it they were unlikely to know, and I warned them about the ways that using an LLM could cause them to submit inadequate work (incoherent methods and fake references, for example).
In their reflections, many international students reported that they used chatGPT to help them correct grammar, and to make their writing "more polished".
I was sad that so many students seemed to be relying on chatGPT to make them feel more confident in their writing, because I felt that the real problem was faculty attitudes toward multilingual scholars.
I have worked with a number of graduate international students who are told by other faculty that their writing is "bad", or are given bad grades for writing that is reflective of English as a second language, but still clearly demonstrates comprehension of the subject matter.
I believe that written communication is important. However, I also believe in focused feedback. As a professor of design, I am grading people's ability to demonstrate that they understand concepts and can apply them in design research and then communicate that process to me.
I do not require that communication to read like a first language student, when I am perfectly capable of understanding the intent. When I am confused about meaning, I suggest clarifying edits.
I can speak and write in one language with competence. How dare I punish international students for their bravery? Fixation on normative communication chronically suppresses their grades and their confidence. And, most importantly, it doesn't improve their language skills!
If I were teaching rhetoric and comp it might be different. But not THAT different. I'm a scholar of neurodivergent and Mad rhetorics. I can't in good conscious support Divergent rhetorics while supressing transnational rhetoric!
Anyway, if you want your students to stop using chatGPT then stop being racist and ableist when you grade.
Here's a test I'll be adding to my repertoire of LLM tests.
"I left Singapore at 9PM on Oct 20, flying east to San Francisco where I will arrive at 9PM on Oct 20. How much daylight will I see looking out the window?"
All of them - ChatGPT4, Claude, Bard - got it spectacularly wrong, confidently asserting I'd see lots of daylight.
When I told them I'm 10 hours into the flight and have seen none so far they were all like "oh, yeah, right, sorry about that."
I just issued a data deletion request to #StackOverflow to erase all of the associations between my name and the questions, answers and comments I have on the platform.
One of the key ways in which #RAG works to supplement #LLMs is based on proven associations. Higher ranked Stack Overflow members' answers will carry more weight in any #LLM that is produced.
By asking for my name to be disassociated from the textual data, it removes a semantic relationship that is helpful for determining which tokens of text to use in an #LLM.
If you sell out your user base without consultation, expect a backlash.
AI Lie: Machines Don’t Learn Like Humans (And Don’t Have the Right To) (www.tomshardware.com)
Some argue that bots should be entitled to ingest any content they see, because people can.
ChatGPT Can Be Broken by Entering These Strange Words, And Nobody Is Sure Why (www.vice.com)
Reddit usernames like ‘SolidGoldMagikarp’ are somehow causing the chatbot to give bizarre responses.