Artificial intelligence still has some way to go

BJO? Man fuck that asshole!

lol

https://pbs.twimg.com/media/FyjCLKiXsAM-zKw?format=jpg&name=small
https://pbs.twimg.com/media/FyjCLKiWYAAkKjo?format=jpg&name=small

Business Basiness

What I feel like will likely continue to happen, and that I have already seen happen, is that every time people start to notice one of these glitches, these little cracks that break the illusion, companies like OpenAI will just look for ways to paint over the cracks through training, in order to make the illusion more perfect. But I am skeptical that it will stop being an illusion, it will just be harder and harder to tell.

Interesting.

I’m a lousy technology forecaster (especially compared to many people here) but I believe it’s equally likely that additional data will provide no benefit and additional data (e.g., Google Books) will provide tremendous benefit. If it’s the former, OpenAI is fucked without major research breakthroughs (e.g., I double RLHF is sufficient).

s/double/doubt

starp hop

it sounds like an extremely drunk seminar leader "Ieeem gn telllyawwooo howrna business yoar busness starp!"

I don't see the issue, this is what all these sites read like to me anyway

Haha I was just sitting here listening to Autechre when I scrolled past Eedttlpe, Plolk, Busiiness yoeir etc. Strangely fitting.

welp, here's one of my takes:

https://www.newyorker.com/culture/cultural-comment/what-to-do-about-fake-drake-songs

Wonderful essay by @henryfarrell and Cosma Shalizi. AI has us in a tizzy, but we've been living in a world dominated by strange superhuman intelligences for a couple centuries centuries now--they're known as markets and bureaucracy. https://t.co/oo9eBKGIOp

— Brink Lindsey (@lindsey_brink) June 22, 2023

It really has a long way to go.

https://www.theverge.com/features/23764584/ai-artificial-intelligence-data-notation-labor-scale-surge-remotasks-openai-chatbots

"But just as likely, the rise of AI will look like past labor-saving technologies, maybe like the telephone or typewriter, which vanquished the drudgery of message delivering and handwriting but generated so much new correspondence, commerce, and paperwork that new offices staffed by new types of workers — clerks, accountants, typists — were required to manage it. When AI comes for your job, you may not lose it, but it might become more alien, more isolating, more tedious"

“I read and I Googled and found I am working for a 25-year-old billionaire,” said one worker, who, when we spoke, was labeling the emotions of people calling to order Domino’s pizza. “I really am wasting my life here if I made somebody a billionaire and I’m earning a couple of bucks a week.

Just hilarious.

"This circuitous technique is called “reinforcement learning from human feedback,” or RLHF, and it’s so effective that it’s worth pausing to fully register what it doesn’t do. When annotators teach a model to be accurate, for example, the model isn’t learning to check answers against logic or external sources or about what accuracy as a concept even is. The model is still a text-prediction machine mimicking patterns in human writing, but now its training corpus has been supplemented with bespoke examples, and the model has been weighted to favor them. Maybe this results in the model extracting patterns from the part of its linguistic map labeled as accurate and producing text that happens to align with the truth, but it can also result in it mimicking the confident style and expert jargon of the accurate text while writing things that are totally wrong. There is no guarantee that the text the labelers marked as accurate is in fact accurate, and when it is, there is no guarantee that the model learns the right patterns from it."

Perhaps a clever AI firm will start hiring the subeditors and fact checkers who have been pushed out of the publishing and news industries.

more tedious

not possible.

Barred From Grocery Stores by Facial Recognition

Facewatch, a British company, is used by retailers across the country frustrated by petty crime. For as little as 250 pounds a month, or roughly $320, Facewatch offers access to a customized watchlist that stores near one another share. When Facewatch spots a flagged face, an alert is sent to a smartphone at the shop, where employees decide whether to keep a close eye on the person or ask the person to leave.

Mr. Mackenzie adds one or two new faces every week, he said, mainly people who steal diapers, groceries, pet supplies and other low-cost goods. He said their economic hardship made him sympathetic, but that the number of thefts had gotten so out of hand that facial recognition was needed. Usually at least once a day, Facewatch alerts him that somebody on the watchlist has entered the store.

truly a miraculous technology

and tying this back to AI is made of people:

Every time Facewatch’s system identifies a shoplifter, a notification goes to a person who passed a test to be a “super recognizer” — someone with a special talent for remembering faces. Within seconds, the super recognizer must confirm the match against the Facewatch database before an alert is sent.

The Button:

https://www.oneusefulthing.org/p/setting-time-on-fire-and-the-temptation

really hope one of these wide eyed imbeciles ends up being correct, can't wait for ai to cure cancer and alzheimers and climate change.

https://www.theguardian.com/technology/2023/jul/06/ai-artificial-intelligence-world-diseases-climate-scenarios-experts

This is not, now, cutting-edge stuff, but it is a reminder that a lot of the boundary is and remains shit jobs

https://www.nplusonemag.com/issue-44/essays/human_fallback/

tremendous memoir. this is what a LOT of work is going to look like very soon.

the mind boggles:

https://i.kinja-img.com/gawker-media/image/upload/c_fit,f_auto,g_center,q_60,w_1600/1dccf75c7243986c92b6e21dd2f331d2.jpg

(if that doesn't show up, it's a lovense vibrator with built in chatgpt)

Lock thread

and there it is

new eugenics just dropped, "The Carbon Emissions of Writing and Illustrating Are Lower for AI than for Humans" 💀💀💀 https://t.co/GN0v5mmzd9 pic.twitter.com/PO0dWdGWq5

— Kyle McDonald (@kcimc) July 19, 2023

and ChatGPT authored the study, unprompted

lol I took a look at that and you'll be surprised to hear it was garbage

https://www.forbes.com/sites/charliefink/2023/07/18/new-showrunner-ai-the-sum-of-all-hollywoods-fears/?sh=7f3375e65b72

As Hollywood actors and writers strike, a company called The Simulation (formerly Fable Studio) has introduced Showrunner, a new AI application that can create 22-minute generative AI TV fan and parody episodes of popular shows. To demonstrate the power of Showrunner, The Simulation released a 22- minute TV episode of South Park today, “Westland Chronicles,” which centers on the ongoing WGA (Writers Guild of America) strike and a Hollywood studio, Bizney, that uses AI with disastrous results. The episode, and Showrunner’s technology, confirm the threat of AI is a real, not existential, problem. Indeed, Showrunner’s creators say it's too dangerous to release to the public.

(the "it's too dangerous to release to the public" is total marketing speak)

It's what they said about GPT-3

remember when they told us chatgpt was reaching human-level intelligence and going to replace us? 🫠 pic.twitter.com/Xh2vbDeIHr

— Paris Marx (@parismarx) July 21, 2023

ChatGPT is now sniffing spray paint, as all adolescent AI's eventually do

Anecdotally, I’ve heard it’s been getting worse. I guess the programmers don’t understand how or why? What a weird technology.

I've been wondering if it's some kind of inevitable deterioration along the lines of the copy of a copy of a copy thing, or whether this is the specific result of endlessly "fine tuning" it for different purposes, or who knows what.

A relative who works in AI at a major tech company had previously told me--before GPT blew up--about the "catastrophic forgetting" problem in AI (which is a real thing that you can look up) and why it was holding back a lot of AI development. Maybe GPT is just subject to it like everything else.

https://dl.dropbox.com/s/6d2sho558ulcykm/IMG-20230720-WA0031.jpg

That's not actually true, ChatGPT is pronounced with a hard 't' at the end of 'chat' like in English, and unlike the French word for cat.

I was wondering about that but still found it amusing

I've been wondering if it's some kind of inevitable deterioration along the lines of the copy of a copy of a copy thing, or whether this is the specific result of endlessly "fine tuning" it for different purposes, or who knows what.

― longtime caller, first time listener (man alive), Saturday, 22 July 2023 bookmarkflaglink

It's been written about. I might have linked an article where it's describing a feedback loop where the gpt is churning 'ok' outputs because it has human inputs. As more gpt content floods the web it will use that as training data.

I haven't read the article but the headline is funny. It's similar to the lawyers who were fined by the judge because they used gpt for work and the docs it churned out were rubbish.

Also the teacher who complained on twitter how the quotes from novels in student essays were made up.

Just going by these small bits of anecdotes it tells you there is very little to fear.

i don't think ai being trained on data which now includes stuff written by ai will lead to a huge degradation in quality. now that the models work so well there's a bigger premium to feeding high quality data into the model, it's possible average quality of inputs go up

there's also a feedback loop where ai is enabling things like digitization of archival text at scales not previously possible, which then leads to higher quality data for ai to be trained on. an example is this research group at harvard (lead by an economic historian) who created a custom ai tool to identify layouts in newspapers, then applied it to create a data-set of headlines in local us newspapers, identifying pairs of local newspaper headlines describing the same underlying AP news story, which can then be used to train language models

A diversity of tasks use language models trained on semantic similarity data. While there are a variety of datasets that capture semantic similarity, they are either constructed from modern web data or are relatively small datasets created in the past decade by human annotators. This study utilizes a novel source, newly digitized articles from off-copyright, local U.S. newspapers, to assemble a massive-scale semantic similarity dataset spanning 70 years from 1920 to 1989 and containing nearly 400M positive semantic similarity pairs. Historically, around half of articles in U.S. local newspapers came from newswires like the Associated Press. While local papers reproduced articles from the newswire, they wrote their own headlines, which form abstractive summaries of the associated articles. We associate articles and their headlines by exploiting document layouts and language understanding. We then use deep neural methods to detect which articles are from the same underlying source, in the presence of substantial noise and abridgement. The headlines of reproduced articles form positive semantic similarity pairs. The resulting publicly available HEADLINES dataset is significantly larger than most existing semantic similarity datasets and covers a much longer span of time. It will facilitate the application of contrastively trained semantic similarity models to a variety of tasks, including the study of semantic change across space and time.

from my pov the advances in ai in the last year have been pretty incredibly useful. github copilot, a chat gpt application specialized at writing code, saves me an insane amount of time. since i now spend less time writing the code myself i have more time to de-bug and test it, which actually makes it less error prone (contrary to what one might expect given hallucinations). a prof i know (who's a bit of a "hacker" and uses the api versions of these tools) created a chatbot trained to help students with his courses. it answers questions, creates practice problems and gives students feedback on their solutions. he uses some tricks to reduce the rate of errors, like turning down the "temperature" parameter (which governs the amount of randomness in the answers) to zero, and somehow restricting it to focus only on the course material (using some kind of latent-space dimension reduction trick i don't understand). i haven't used it for writing yet but some of my friends are using it to write their dissertations, and say it's helpful in getting past writer's block cause you can just get it to start you off with a paragraph by giving it some stuff in point form, then edit from there

"there's also a feedback loop where ai is enabling things like digitization of archival text at scales not previously possible, which then leads to higher quality data for ai to be trained on. an example is this research group at harvard (lead by an economic historian)"

That highly specific use case doesn't disprove the point.

But yes it's not all terrible for sure. There is a lot to AI, mostly responding to the more outlandish stuff.

xp- i don't think it can be "disproven" one way or the other. but there are forces pushing it in both directions, it's not obvious that the proliferation of text written by ai online will be the dominant force. as allen said above, google training an ai on all of google books could lead to a compensating improvement in data quality

Good thread on AI being used by students, by an ex-academic/postgrad type. This is the key takeaway.

I would only really worry about the impact of ChatGPT on the situation if it were in some way seducing the students who did care, or might potentially be induced to care, into not caring. As for the rest, I suppose a convenient illusion is being dismantled -

— a furred tail upon nothingness (@dynamic_proxy) July 31, 2023

https://www.washingtonpost.com/technology/2023/08/10/san-francisco-robotaxi-approved-waymo-cruise/

SAN FRANCISCO — California regulators voted Thursday to allow self-driving car companies Waymo and Cruise to offer 24/7 paid taxi service in San Francisco, a major win for the industry that could pave the way for more widespread adoption of the technology.

Cars without drivers have become a common sight on San Francisco’s winding, hilly and often foggy streets. Thursday’s vote stripped most limitations on operating and charging for rides, essentially creating more ride-hailing services like Uber or Lyft — just without the drivers.

It’s a pivotal moment for the autonomous transportation industry, expanding one of the biggest test cases for a world in which many companies envision not needing drivers at all. For years, companies from Amazon to Google have experimented with self-driving vehicles, something that could prove incredibly disruptive to the labor economy if it ever materializes en masse.

In California alone, there are more than 40 companies — ranging from young start-ups to tech giants — that have permits to test their cars in San Francisco, according to the California Department of Motor Vehicles. According to a Washington Post analysis of the data, the companies collectively clock millions of miles on public roads every year — along with hundreds of mostly minor accidents.

put that on the t-shirt

"mostly minor accidents"

700k a day for a company sitting on $10b isnt really that much but the fact remains ai does use a lot of computers just to come up with some half to full bullshit

"OpenAI spends about $700,000 a day, just to keep ChatGPT going. The cost does not include other AI products like GPT-4 and DALL-E2. Right now, it is pulling through only because of Microsoft's $10 billion funding"

yoooo
https://t.co/k8qm6Lo0j3

— Olúfẹ́mi O. Táíwò (@OlufemiOTaiwo) August 12, 2023

In my SF neighborhood this weekend, I've seen a near constant parade of different (the cars have names) empty Cruise vehicles driving along various routes. I assume they are trying to collect as much training data as possible, but it feels a little like they are celebrating being unleashed.