treesprite82

Probably expected just due taking a model trained on all sorts of content and fine-tuning specifically on non-fiction data.

Wouldn't conclude anything about /pol/ in particular without at least comparing the same done for HackerNews/Reddit/etc.

> you forget that the human had to predict that the combination of two would be interesting for other humans, and then construct the prompt, possibly selecting the best pictures. That's who did most of the work here

If the Twitter user claimed that the text prompts themselves were generated by asking GPT-3 for "An interesting sequence of text prompts to feed an image generation AI" or something like that, I would have believed them.

Presumptively, I imagine it's harder to create a model that generates images matching a certain human-language input prompt than to create an image generation model with no language component and have it pick its own scenarios internally. I don't think the former is done to palm off "most of the work" to humans, but rather because people want an easy way to see create own ideas so there's more demand for it.

> You have to re-train it from scratch every time you want it to remember something truly new, there's no feedback loop to do that.

As far as I'm aware, this isn't true. Deep learning is perfectly compatible with fine-tuning an existing model using new data. OpenAI/MS have been doing this with Codex to improve it based on Copilot telemetry and code from new languages/libraries.

Seems like it adapted Kermit's features to fit with the world of the movie, as if he was actually a character from that movie.

I don't have access to DALL-E 2, but I wonder if a prompt like "A cameo from Kermit the Frog in ..." would give more literal Kermits.

> It would have been better to ignore the emergency vehicle and continue driving, which would have at least cleared the jam.

With the fire truck using the oncoming lane to overtake the garbage truck, and with the articles saying a human could have reversed back into the intersection to clear the lane, it sounds like the fire truck would be in the way of the autonomous car just driving forwards.

> The same thing would have happened with 2 autonomous vehicles, because neither even attempts to understand what the other is trying to do

I don't think this is true in general - they seem to rely heavily on judging the intent/target of other vehicles to predict future path and react to those possibilities.

Likely that the autonomous car A (in the position of the fire truck) would not overtake the garbage truck when it can see car B oncoming in that lane in the first place, but in the event that it does occur, car B would likely slow to prevent a potential accident (understanding car A is attempting an overtake and may continue forward) and car A would probably pull back in behind the garbage truck (understanding that it'd be in the way of car B continuing forwards).

> If there’s no vehicles or people directly behind the car, no traffic close enough to pose a danger, and a quick spot check confirms this

No traffic close enough to pose a danger isn't a given in this scenario - there may well have been traffic.

> why would it take you longer than 5 seconds to reverse out of the way?

I think many would be at least slightly hesitant to back into an intersection. There would be initial reaction/braking time, time to check for and evaluate options, then time to reverse backwards with caution.

> In the end the garbage truck moved, which is why the fire truck was “able to pass within 25 seconds of encountering the Cruise car”.

The autonomous car did yield to the right to the extent it could, so wasn't doing nothing for 25 seconds. Just that (I'll go with the first responders' judgement) doing so didn't give sufficient space.

From what I can tell, it detected the emergency vehicle and yielded to the right. Just that it didn't reverse back into the intersection to clear the lane completely.

> A human driver in that situation should not spend longer than 5 seconds figuring out and implementing a way to safely let the fire truck pass. The Cruise car took longer.

Impossible to say for certain, but I'd bet that a significant number of human drivers would yield to the right (as the autonomous car did), or take longer than 5 seconds to reverse back into the intersection (as was desired of the autonomous car).

Autonomous car didn't seem to fail to notice the fire truck, just that it arguably took the wrong course of action by yielding to the right instead of reversing into the intersection.

Rough diagram of my understanding from information I can find: https://i.imgur.com/VOp8cWi.png

Fire truck wants to use oncoming lane to overtake the double-parked garbage truck, autonomous car in that oncoming lane yields to the right to the extent allowed by more parked cars - but doesn't back up into the intersection to completely clear the lane.

I don't perceive the benefit from risk reduction of helmets in cars to overcome the hassle hurdle. But I wouldn't advocate banning others from wearing helmets in cars if they so wished.

> Do you wear a helmet when you drive inside your car? If not, why not?

Cars already have airbags and seatbelts which help a lot for the kind of collisions that would otherwise result in head injuries.

> It might make you much safer in case of a crash, according to your reasoning.

I don't see what in their comment could be construed to say that helmets make you much safer regardless of vehicle.

> Wearing a helmet can itself become a leading factor to cause an incident

Is there a source for this? Should be a randomized A/B test as you mentioned, not just a correlation - wearing a hi-vis jacket or other precautions taken more often in dangerous situations probably also correlate with accidents.

Even if helmets do cause accidents through increased carelessness, some may still take issue to intentionally making a scenario more dangerous such that people are more careful. It's kind of settling for a local minimum, rather than aiming to reduce inherent risk alongside aligning people's risk estimates to not overestimate the precautions.

> a stupid bald c** and threatened to deck me

The "harassment" part makes sense at least.

> and it related to the claimant’s sex

I'd have assumed "sexual" meant relating to sex (intercourse), rather than relating to sexes (male/female).

I think the relevant act would be: https://www.legislation.gov.uk/ukpga/2010/15/section/26

> The issue with CSV, as far as I am concerned, is the escaping. And that isn't minor from either a readability or parsing point of view.

For parsing I'd agree, as you've eliminated the need to handle escaping by declaring some characters off-limits. But for manual editing, to start a .usv file with most editors people would be copying the characters from google rather than just being able to type commas and newlines.

> Are \u001F or \u001E even legal filename characters on any OS? These codes could turn up by chance in a blob of binary data. But it isn't intended for binary data.

Generally I believe NUL and / are the only ASCII characters considered safe to never occur in file/directory names.

> easy to manually edit (Notepad++ shows the unit separator as a ‘US’ symbol)

Won't it all be in one long line?

> Typing \u001F or \u001E in some editors might be a faff, but it is hardly a showstopper.

Doesn't seem negligible when the editability issues with CSV were also minor.

> No escaping. If you want to put \u001F or \u001E in your data – tough you can’t. Use a different format.

Wanting those characters specifically is rare, but wanting to safely store an arbitrary unicode string - maybe a filename - is very common. Parsers might invent their own escaping to handle this.

There are areas of art where the main draw is that it was made by a human. Despite cameras and image filters, hyperrealistic pencil drawings are still appreciated because they're an impressive demonstration of human talent.

However, I'd assume those employ very few people compared to more "functional" creative jobs where the end product is what matters most, like creating clip-art for a company website or character portraits for a video-game. I think there's a large risk here that teams of artists will be collapsed into a single "art director" job, for cost-saving and faster results.

In terms of creative jobs that have safety due to difficulty in automating, maybe something involving bespoke physical products like prop/set/costume/lighting design for a movie?

----

It hasn't yet been 10 years since AlexNet (often seen as kicking off the current deep learning trend, though not strictly the first at anything). I think people are being overconfident in the idea that current specific DL weaknesses will remain weaknesses in the upcoming decades of your career.

I've set a long-term reminder to look back at this thread since I'm interested in seeing how well predictions/sentiments have held up, some I'll collate below along with questions for my future self:

> This really isn't going to change much. DALLE-2 has and will likely for a long time have a lot of imperfections that will never be fixed.

> Even if AI can mimic style with style transfer, it will never achieve greatness this way.

Using baseline of a human artist, do generated images still exhibit a large number of imperfections? Do blind tests show generated artworks are considered inferior quality to human artworks?

> Due to the inherit way that these GANs work it's not going to force anyone really out of business.

> Art is about human expression. It is less about the end product and more about the journey and expression that got to that end product. This machine learning art generation skips all of that. I personally wouldn't worry about it.

> They will no more replace artists than photoshop or illustrator did.

> All of 'em. All of them will be safe.

Have any artists been made redundant due generative models? Should artists starting their career in the 2020s have been worried?

> There is also something cool and interesting about AI generated art currently because it is new and fresh. At some point though that will wear off.

Has the interest in generated art worn off? If so, was that because it was passing fad, or because it's now so commonplace that it's considered normal?

> After all, knowing that the tosses are independent is just knowing that a heads is not more (or less) likely after a string of tails; therefore anyone who thinks that a heads is more likely after a string of tails does not know that the tosses are independent.

I think you could similarly dismiss any formal fallacy. The fact that X implies Y and some accept X (coin tosses are independent) but not Y (results are not more likely when "overdue") is what makes it a fallacy.

From then on the post redefines gambler's fallacy to a scenario in which you know the outcome percentage but not if results are independent. Still an interesting post, but not really what I've seen meant by gambler's fallacy.

A few jobs like football player or chess GM should be inherently immune to automation; they exist because we're interested in seeing what humans can do, even if machines can outperform them.

For some artistic/creative jobs we might also value it coming from a human, but I'd be cautious of overstating the effect here. Art in a gallery may be safe for now, but more "functional" art like character portraits for a videogame or clip art for a company website seems in imminent danger.

Plumbers, electricians, mechanics, etc. can be replaced for repetitive known work, like mass-production of vehicles or household appliances, but I don't think robotics is advanced/cheap enough yet to be approaching the point where a robot would turn up to your home, navigate around, access the right areas, and fix some variable maintenance issue.

Waymo arguably reached level 4 with commercial taxis with no human at the wheel a while back: https://www.youtube.com/watch?v=__EoOvVkEMo

Lower stakes because of the speed, but a lot of special cases. Higher chance of pedestrians and parked cars in the road for both. Parking lots will have a lot of cars moving in atypical ways, and markings aren't as standardized as on roads.

> Destiny for example is now permanently banned from twitch because he took the position that trans shouldn't compete in physical sports.

Destiny's ban reason and length are only speculated (Twitch uses "indefinite" in its literal sense). May also be from having Nick Fuentes on stream recently, since Twitch can be strict about featuring banned streamers [0]. Others guessed that it was this tweet [1] but I doubt it - seems to be something on-stream.

If the ban is from the trans athletes take (I do agree that it's the most likely cause), Destiny's usual edginess [2] probably contributed. Others have been fine with the same position.

[0]: https://news.ycombinator.com/item?id=30801959

[1]: https://archive.ph/wRM6u

[2]: https://streamable.com/trlxfu (not my title, I'd say there's deniability in the exact (sub)group being referred to)

At the extreme: Keemstar has long been banned from Youtube, but remained prominent through DramaAlert which is not technically his channel.

> Article includes no JKR quotes

> She uses an Orwell quote which doesn’t refer to rape at all

Third paragraph of the article has a JKR quote directly referencing rape.

Not sure if it backs up the original "freaking out about..." claim, but maybe "by placing undue emphasis, perpetuates the narrative that...".

> But that's not a question of fairness, rather it's a question of feasibility.

I initially took the challenge's intention to be about highlighting modern machine learning's weaknesses compared to biological intelligence, and so barring certain already-existing generalization techniques seemed an arbitrary and asymmetrical restriction.

If it's more meant as "Classifiers can already achieve this particular goal, but I have a theory that human-determined 'predicates' will scale up better in the long run, I challenge you to progress my idea", then I currently disagree but understand.

> If it took us many thousands of years to learn our background knowledge from the real world over many human generations, it's difficult to see how we can reproduce this result with the comparatively poor computational resources and data in our disposal.

My belief would be that we can surpass this result with a combination of using our existing intuition alongside techniques that outperform evolution's hypo-glacial pace and inefficient data utilization.

Given the same narrow problem, some human insight for the search space and a couple of hours of gradient descent on a GPU can match what would take evolution many generations. That doesn't prove we'll get such a speedup on achieving broader intelligence, but at least natural selection hasn't appeared to be a speed limit so far.

> or, we can find a way to transfer the background knowledge bestowed upon us by thousand years of evolution to guide the training of our learning systems towards the goals we want them to achieve, whatever those are.

> To clarify, I'm not saying we should go back to feature engineering

> Clearly not explicitly coding expert knowledge in production rules as in expert systems. Much of our knowledge is maybe impossible to articulate explicitly. So we must find a way to encode implicit knowledge, also.

I'm all for finding ways to use human domain knowledge to guide the network in the right direction, essentially making use of a gigantic dataset from life's history. The trend seems to be to do this at an increasingly high level: weights are found by gradient descent, and hyperparameters by NAS or similar, but humans still designing various layers and blocks.

"Predicates" being a probably-small set of feature detectors which can describe all 2D images makes me think of something like eigenfaces, which it felt backwards to have humans determine. Maybe intended to be broader than that?

> It's basically a trade-off. If you have good background knowledge, you don't need a lot of data. Good background knowledge helps you build robustly generalisable concepts. And if you can reuse the learned concepts as background knowledge, then the sky is the limit.

I'd claim that this is in effect also what pretraining is. Pretraining and instilling a model with human background knowledge both allow faster low-data generalization to new tasks by utilizing large amounts of prior data. Difference is in whether the base data is organic or digital. Using both to find useful predicates seems most promising so far.

> they push the problem of training with big data to the pre-training stage and then claim to do "few-" or "one-shot" learning at the end

Humans have had 4 billion years of natural selection and then 4 years of input from all senses before they start identifying digits. I've seen studies suggesting that we're already born with an area of our brain for recognizing letters and words.

Seems at least fair in comparison to allow MAML/pretraining to find a good starting model (e.g: can recognize lines and shapes) by utilizing data other than the classes of interest.

> It's like the Aesop's fable where the sparrow hid in the eagle's feathers and jumped up at the last moment to claim "I'm the bird that flies the highest!".

> you should not need a lot of data for anything

Is choosing suitable starting weights/architecture/"predicates" by hand-designing based on our own built up information qualitatively any different? It still seems like "hiding" utilization of a huge amount of background knowledge about digits/symbols/images/reality.

Arguably harder to expand that way too. I think techniques such as unsupervised learning are probably going to be a more feasible way to utilize the increasing amount of data we're collecting about the universe.

At our current stage, both seem useful. Broad strokes like moving from dense networks to convolutional networks to add locality and translational invariance based on our knowledge that this is an appropriate search space for vision tasks, and then automated methods like NAS and pretraining to determine relevance on a finer level.

We definitely haven't exhausted ways for us to use our intuition to guide networks in the right direction, such as transformers with their attention mechanisms or say a network inherently agnostic to horizontal flips rather than teaching that with data augmentation, but I'm skeptical about what sounds like stepping back into hand-crafted feature extraction which automated techniques have been far more effective at.

> but the reliance on big data for pre-training suggests that the models are still not learning good representations that generalise well - so they still need big data to make up for it.

Wouldn't it be lack of generalization to new tasks after the fact which indicates poor predicates? I don't see why good feature detectors should necessarily themselves be discoverable by hand or with low data, as that doesn't appear to have been the case for organic intelligence.

> To summarise: learn to identify MNIST digits from 60 examples of each class, rather than 6000,

SOTA accuracy on a similar but more challenging problem (5-shot 20-way rather than 60-shot 10-way) appears to be around 99.6%: https://paperswithcode.com/sota/few-shot-image-classificatio...

> while retaining current accuracy.

Depends how strict you're being with this. There's room for the gap to shrink, but I think on average classifiers (whether organic or machine) with a large number of examples to go off of will always perform at least marginally better than classifiers with fewer examples.

My guess as to what will concretely emerge is protocols/standards for things like 3D user avatars and world hopping, so that worlds hosted by anyone can be linked up, rather than only uploaded to a proprietary game with rules/technical limitations decided by the game's devs. The "Metaverse" would be to 3D environments what the Web is to HTML pages.

But the article's "Ask 50 people what the metaverse means, right now, and you’ll get 50 different answers" is very true. I have no idea if this is what Facebook or Epic Games have in mind.

> Like in no universe is it ever OK to do this even if nobody responds to your discord poll

As far as I can tell, they didn't even really make a Discord poll.

The first mention of "spider" (or "IP" or "proxy" or any related words) is from the release announcement of a version which already contains Spider, which got negative reactions: https://i.imgur.com/gLKce8W.png

It's very deceptive wording since "I made an announcement to poll people's reaction" is technically true, but it wasn't a poll or before the changes were made as is implied. Same with "There were very little comments", which leaves out the fact the comments they did get were negative and the low quantity is just from it being a small unknown Discord server.

It's worse than it seems.

The very first mention of "spider" on that Discord was already the release announcement of a version containing it and got negative reactions: https://i.imgur.com/gLKce8W.png

I skimmed through the dev's messages prior to that (not many) and couldn't find anything else related.

I believe "I made an announcement to poll people's reaction" is intentionally deceptive wording to imply they made a poll beforehand, but they're actually just referring to the release announcement. Same with "There were very little comments" which misses out the fact that all those comments were negative, and there were few comments primarily because it's a small unknown Discord server.

I'd say anyone who buys into the wave of fearmongering against vaccines is antivax, even if they've had previous vaccinations. Two twins in their 70s choosing not to get vaccinated would fall into this camp.

Conversely, if there's someone for whom medical consensus agrees is legitimately better off without vaccines (e.g: due to some immune system condition), then I wouldn't call them antivax.

To clarify, it was the banned user spamming dogebonk cryptocurrency posts - not the mod. I don't think there's any wrongful behavior on part of the mod, assuming the offer was a joke.

If someone banned you and your 3 alts then I believe that would've been an admin (paid reddit employee). Subreddit mods (volunteers) can only prevent you posting in the subreddit(s) they moderate, and can't see your IP or anything.

Reddit as a whole doesn't forbid Youtube videos. Some specific subreddits may disallow videos.