[HN Gopher] AI's Instagram Problem
___________________________________________________________________
AI's Instagram Problem
Author : deepwaters
Score : 47 points
Date : 2023-02-23 20:38 UTC (2 hours ago)
(HTM) web link (www.deeplearning.ai)
(TXT) w3m dump (www.deeplearning.ai)
| minimaxir wrote:
| All AI projects are valuable, but it's annoying and demotivating
| to work hard on a unique and useful project and have no one
| read/use it because it doesn't stand out in the extremely
| rapidly-evolving ecosystem.
|
| Most of my AI text-generation and image-generation projects and
| tools have already become obsolete by technology released in the
| past few months, and I've almost given up competing.
| gnramires wrote:
| One advise I'll never forget (from an interview with late
| mathematician Maryam Mirzakhani) is that you don't have to race
| faster than everyone else (although if you can, that's great!).
| You can just run in a direction no one is running toward (and
| hopefully be persistent). Then you're only racing yourself and
| speed is largely irrelevant.
|
| I myself know I'm very slow, specially developing projects. So
| I don't try to race anyone, except contribute to things I know
| are neglected. There are so many neglected problems, you don't
| need to work on the most coveted fields.
| satvikpendem wrote:
| Problem is that's not how academic funding works. Grants are
| generally not given to possible dead ends.
| nl wrote:
| Don't be discouraged. Your work is good.
|
| And remember we are 6-12 months off having good, open source
| chatGPT class open source models and the software support to
| make it possible to run them at home.
| eternalban wrote:
| Ironic mantra for you: Geoffrey Hinton in 70s, Minsky, Symbolic
| Logic, Neural Nets, stay the course.
| knicholes wrote:
| When asked how to keep up with all of this AI/ML stuff when
| starting your own AI business, Jeremy Howard (general genius
| and co-creator of fast.ai) replied that you don't need to know
| everything and keep up with everything. Pick a niche and focus
| just on that. And if that keeps changing too fast or is too
| much, pick a niche inside that niche.
| Xeoncross wrote:
| Stuff does happen though.
|
| - Facebook enters the geolocation game
| (dowalla/foursquare/instagram)
|
| - Google releases reviews to maps (yelp & co)
|
| - Amazon starts it's own airline (Fedex)
|
| - Microsoft adds ChatGPT to search (Kagi, Duckduckgo, etc..)
|
| Sometimes it's hard to recover your startup when a big player
| does something to fill the void you were targeting.
| aziaziazi wrote:
| How does DDG and Kagi competes with ChatGpt ? In my
| understanding their "responses" above search results works like
| Google graph, but I'll be happy to be corrected.
| Xeoncross wrote:
| Most people don't use search engines to search, they use them
| to answer.
|
| That's why Google, DDG, Kagi and others are all putting
| summary blurbs.
|
| Now that we have ChatGPT a lot of people are going to cut out
| searches in exchange for prompts.
| baxtr wrote:
| _> AI develops so quickly that waves of new ideas keep coming:
| quantum AI, self-supervised learning, transformers, diffusion
| models, large language models, and on and on._
|
| I wonder what's wrong with us as a species. We love to run into
| one direction all together chasing the next big thing just until
| the next big trend comes along.
|
| Quantum AI crypto LLMs for the win!
| xg15 wrote:
| That's definitely not "us as a species". I would say this is a
| specific effect of post-industrial, market-driven economy - and
| even there, probably restricted to specific communities such as
| entrepreneurs.
|
| By and large, I believe, people are conservative and would
| prefer for things to stay as they were, unless the change
| solves a specific problem.
|
| Sometimes people actively tried to resist "progress" and had to
| be forced into their luck by those in power. E.g.:
| https://en.wikipedia.org/wiki/Luddite#Government_response
| pixl97 wrote:
| And I would say its even more detailed than that.
|
| In general your most staunch conservative doesn't mind things
| changing if it personally benefits them greatly (especially
| if they don't bear most the burden of its effects).
|
| Your most avid progressive commonly want the things that
| bring them comfort to stay the same, even if its potentially
| harmful to someone else.
|
| I believe the issue here is we're playing "everything
| everywhere all at once". Everywhere on the globe is pretty
| much instantly connected to everything else via digital
| communications. There is no more sit by the sidelines zone
| that gets to avoid this connectivity. You may not want it,
| but you have to stop everyone else from bringing that
| connectivity too, and somehow maintain casual connectivity
| economically unless you want to play caveman.
| phailhaus wrote:
| New technology = race to figure out how to make the most money
| off of it. Not that complicated, nor does it have anything to
| do with "us as a species".
| uoaei wrote:
| > what's wrong with us as a species
|
| We ask for faster horses, when automobiles would get us where
| we want to go. It is a lack of imagination for what is possible
| in the future that keeps us risk-averse and focused instead on
| a quest for incremental improvements to the status quo.
|
| This was the focus of the idealistic vision of the whole
| "disruption" meta-meme as it spread through Silicon Valley.
|
| Not to say that maintenance and refinement aren't important,
| but innovation doesn't look like just adding more parameters
| and ways to scale massive computations.
|
| The main concept that AI laypeople (even those who burn OpenAI
| credits on a hobbyist basis) are missing to really grasp the
| possibilities of these kinds of statistical analysis techniques
| is a deep understanding of the notion of _inductive bias_. For
| instance, Transformers are so powerful precisely because they
| have generalized that idea of inductive bias one level up by
| using multi-head attention to project onto many different
| linear spaces and attend to them differently in different
| contexts, rather than just filtering through effectively one
| linear perspective at a time as with your traditional dense
| NNs.
|
| The kinds of innovations we have witnessed most recently are
| actually pretty small steps in the space of possible
| architectures, not to mention training methods, etc. For
| instance, predictive coding approximates gradient descent[1]
| which opens up all sorts of new architectures (feedback arch,
| modular/federated/local compositions, etc.) that are
| intractable with traditional backprop-based techniques unless
| you can manage the infrastructure around periodic global
| parameter/state syncs.
|
| [1] https://openreview.net/pdf?id=PdauS7wZBfC
| potatolicious wrote:
| Are these really separate phenomena though? I'd strongly argue
| that between self-supervised learning, transformers, diffusion
| models, and LLMs we're looking at phases of the same
| phenomenon, each building on top of the knowledge that came
| before.
|
| Like vaccines - inactivated viruses isn't a different "wave"
| than mRNA, they are logical successors within the same
| intellectual pursuit.
|
| Or Vulkan vs. OpenGL. They aren't separate "fads" that gained
| traction.
|
| Or ARM vs. x86. They aren't "fads" so much as entirely
| reasonable evolutions of technologies based on the evolving
| knowledge of the field and its needs.
|
| Not everything is a popularity contest.
| softwaredoug wrote:
| There is a pretty big gap between what's talked about at
| conference and blogs vs the actual, not very sexy, hard data
| science work that goes into doing "machine learning" at scale.
| Animats wrote:
| _" Maintain your faith and keep going"_.
|
| Why? You're not going to make money if you're losing in a fast-
| moving field.
| c9capital wrote:
| Instagram AlwayS have a problem....
| keerthiko wrote:
| This is so true, but to follow this advice takes a respectable
| amount of personal discipline, combined with picking a very
| specific niche to focus on. A business with an R&D core such as
| AI should only be considering generalizing (if at all) after
| developing concrete confidence that you are on, and will stay on
| the bleeding edge of your initial niche. It's easy to feel
| ashamed when you tell someone you're doing X, and enough lay
| (tech) persons ask if it uses the trending research technology,
| and you're not. It may also make you think like you _should_ use
| that tech, but it 's likely a costly distraction more often than
| not.
|
| Our company [0] developed a cutting edge computer vision system
| focused on detecting cardio machine exercise cadence (hyper
| specific!) and became the only reliable camera-based solution to
| do so. We then tried to generalize to all exercise motion (rep
| tracking, a still unsolved problem), achieved mediocre success,
| and put the exploration to sleep later because we think waiting
| for other technologies to mature would be easier and faster
| (better 3D cameras, AI pose models, etc). On the other hand we've
| picked other niches that meet our business needs to expand our CV
| R&D into, with pretty good success but mostly just for internal
| use (video content creation tools). More importantly, we're still
| the best camera-based indoor cardio detection tech out there, and
| that's a big part of why we're still alive as a bootstrapped
| business founded in 2010.
|
| Quadruple down on your niche first!
|
| [0] https://www.activetheoryinc.com/
| lakomen wrote:
| Idk about AI having an Instagram problem, but I know Instagram
| has an AI problem. So many fake accounts with AI as actors and
| they send you chat messages trying to pretend to be real people.
| They even react to comments and can discern good from bad
| comments.
|
| At first it was interesting. Now it's just annoying.
|
| Then the Instagram algorithm, if you comment on coffee ads that
| you drink tea, you'll get tea ads. It will show you progressively
| naked women, cameltoes and similar softcore erotica. Sometimes
| even real porn. When you report that porn accounts your report
| will not even be reviewed and denied. But God beware you post the
| word tits. Instant harassment automoderation. All the women that
| follow you instantly message you saying hi, how are you, how old
| are you, what is your name, where are you from etc. I think
| they're infobots trying to build profiles and sell your data.
|
| Some play the old " I'm a hot girl but can't post my picture,
| please buy me an iTunes gift card ", like I was born yesterday.
| And how dumb their game is. They see you commenting on a picture
| you liked, take that picture, claim to be that person, only on a
| private account, and they picked you, their loyal fan. Many of
| them are cheap porn actresses or accounts of people having built
| profiles from the pictures those porn actresses post, only
| modified by AI.
|
| I am so tired of that platform. And then they serve you small 1
| minute clips of stand up comedy, cats or dogs, or some stupid 1
| person having a chat with themself. Let's not forget those "how
| do I say it at my workplace" "expert" vids. And lots of ads in
| between.
|
| Instagram is so disgusting.
| notpachet wrote:
| > Instagram is so disgusting.
|
| Why are you still using it? (Not trying to troll you, I'm
| legitimately curious -- it really doesn't sound like it's
| offering you anything positive at this point)
| mey wrote:
| As someone using Instagram but has a Todo list item to close
| my FB/Instagram/Meta account, the reason I still use it (and
| only Instagram) is to follow artists. A hyper curated list of
| individuals that post interesting things that I can check up
| on. It's what I lost when I shuttered my Twitter account. I
| use it as a glorified RSS feed rather than as a way to
| interact socially.
|
| But yeah, the bad outweighs the good, and other than spending
| the time to archive data, I do plan to shut it down.
| version_five wrote:
| This is correct, but there's also some meta comment to be made
| that research has an "instragram problem". What he's talking
| about isn't really even AI so much as deep learning: there are
| lots of other branches that now get virtually no play relative to
| deep learning. And then there are all sorts of adjacent CS, image
| processing (when was the last time you saw someone talk about
| wavelets), language analysis, and other research avenues,
| probably often better suited to many problems, that get buried
| under the hype.
|
| This is why academic research is so important. All the corporate
| money is flowing to deep learning, grant agencies should be
| sustaining the other more fundamental areas.
| psyklic wrote:
| Academic research has the same problem. When I studied
| neuroscience, flashy fMRI studies received substantially more
| funding than fundamental research. Without understanding how
| simple neural nets work, it's difficult to construct bottom-up
| theories of the brain.
___________________________________________________________________
(page generated 2023-02-23 23:00 UTC)