[HN Gopher] Multiple Stability AI researchers are departing, CEO...
       ___________________________________________________________________
        
       Multiple Stability AI researchers are departing, CEO says
        
       Author : ilamont
       Score  : 121 points
       Date   : 2024-03-20 15:58 UTC (7 hours ago)
        
 (HTM) web link (twitter.com)
 (TXT) w3m dump (twitter.com)
        
       | hbrav wrote:
       | This seems like a general problem for a company like this: your
       | great success demonstrates that your employees are incredibly
       | valuable. And sometimes businesses may say things like "our
       | people are our greatest asset", but that's not actually true.
       | They're not an asset. Your IP is an asset, and you own that. But
       | your people are often capable of going elsewhere and producing
       | equally good work, and they're likely to do that if they get a
       | better offer.
        
         | jackblemming wrote:
         | So give them real ownership of the company and real fruits of
         | their labor. Otherwise yea why wouldn't they go to the highest
         | bidder?
        
           | lil-grandson wrote:
           | That would be ridiculous. All the upside is supposed to go to
           | a small handful of people who already have more money than
           | they need.
        
           | CaptainFever wrote:
           | Are you talking about options?
        
         | open592 wrote:
         | It's no different than professional sports. When you win the
         | super bowl for instance, the value of many of your players
         | (employees) suddenly jump. At that point it's infeasible to pay
         | everyone what they can realistically get on the market. Thus
         | you identify and pay your most essential assets, and the others
         | leave and receive huge pay bumps.
         | 
         | I think it's very valuable for the industry, and as long as the
         | teams who are picking up these employees are able to utilize
         | them effectively, will result in broader innovation.
        
           | dragonwriter wrote:
           | > At that point it's infeasible to pay everyone what they can
           | realistically get on the market.
           | 
           | Largely, because of salary caps created fairly explicitly to
           | break the winning -> popularity -> money -> monopolizing
           | talent -> winning positive feedback loop.
           | 
           | Success feedback loops aren't quite so constrained in most
           | other industries.
        
             | entangledqubit wrote:
             | With the compensation deferral setup (like Ohtani in MLB),
             | I'm curious if the influence of the caps will be eroded.
        
           | golergka wrote:
           | Interesting dynamic, never thought of that.
           | 
           | Were there instances where a sports club actually didn't want
           | to win some title to prevent exactly this from happening? How
           | would the sports association react if the club decided to
           | throw the match for this reason?
        
             | passion__desire wrote:
             | Sumo wrestlers throwing away matches so that the other
             | player can record a winning streak.
        
             | scarface_74 wrote:
             | Well, I know it's not the same thing. But some sports clubs
             | do try to lose so they can get a higher place in the draft
             | picks.
             | 
             | https://en.m.wikipedia.org/wiki/Tanking_(sports)
             | 
             | And of course when kids don't want to play during the
             | summer /s
             | 
             | https://en.m.wikipedia.org/wiki/The_Losing_Edge
        
         | galoisscobi wrote:
         | You assume that IP once created stands on its own. The
         | programming as theory building view suggests that IP will
         | degrade once people who have robust mental models about how the
         | IP was created and works. So software/IP in essence has two
         | critical parts that need to be there, the actual work that was
         | created and robust mental models living in people's heads who
         | continue to work on those projects.
         | 
         | It also goes against the idea that programmers are replaceable
         | cogs that management can change out as and when they want to.
        
           | Filligree wrote:
           | I think you're misunderstanding the GP. He's not saying the
           | people aren't valuable; he's saying they aren't assets. That
           | is to say, they aren't property.
           | 
           | Which isn't what 'asset' means, but it _is_ the central
           | meaning; using it on people always felt squicky to me.
        
         | kcorbitt wrote:
         | In my experience this kind of mass exodus is more associated
         | with failures in leadership/management than an inevitable
         | result of success. For example, OpenAI has been far more
         | successful than Stability and its senior employees obviously
         | have lots of options, but its staff turnover ratio feels much
         | smaller than Stability's.
        
           | brandall10 wrote:
           | Isn't that largely because they're all locked into comp
           | structures that can net top researchers something ~$10M?
           | 
           | I guess one could argue it's a management issue that Sam
           | Altman figured out a way to structure his company to make
           | that possible.
        
         | darth_avocado wrote:
         | > They're not an asset. Your IP is an asset.
         | 
         | While true for existing IP for industries like pharma, for
         | industries like AI, your IP isn't worth anything unless it
         | keeps evolving. If you lose your key researchers, you're done.
        
         | stavros wrote:
         | "Our people are our greatest third-party suppliers" just
         | doesn't have the same ring to it.
        
       | close04 wrote:
       | The source article [0] might be a better link. The tweet (from
       | the article's author himself) isn't adding anything.
       | 
       | [0] https://www.forbes.com/sites/iainmartin/2024/03/20/key-
       | stabl...
        
         | jprete wrote:
         | Source was posted as
         | https://news.ycombinator.com/item?id=39768402
        
         | ilamont wrote:
         | It was paywalled for me.
        
       | mupuff1234 wrote:
       | Im surprised to hear they are struggling to raise money, I
       | figured investors were throwing money at them.
        
         | Filligree wrote:
         | Their value proposition thus far has been "Give stuff away".
         | 
         | I'm grateful for it, and I think they should have government
         | funding if necessary -- the upsides are huge, just little of it
         | goes to Stability -- but it's easy to see why investors would
         | be wary.
        
           | AuryGlenz wrote:
           | Yeah. Every time Emad has talked about their business model
           | changes I've always thought "...that's it?"
           | 
           | I personally think they're missing some low hanging fruit,
           | though I suppose it might be in the name of "safety." I
           | believe Stability (maybe Clipdrop?) at least did have some
           | sort of paid Lora or other training, but I tried it and it
           | was awful. Considering they made the models surely they would
           | have the absolute best insight into full fine tuning it and
           | could roll out a service to do so.
           | 
           | I think they talked about doing it B2B at a presumably much
           | higher cost, but a consumer facing, easy to use way of doing
           | it would at least pull in some money. Of course, the second
           | someone uses that to train child porn or whatever they'll be
           | in hot water.
        
           | 1letterunixname wrote:
           | Failure to monetize. (How is this going to make money? [0])
           | 
           | 0. https://websd.mlc.ai/#text-to-image-generation-demo
           | 
           | It's readily apparent that any sufficiently useful AI/CV/ML
           | should be deployed to solve specific, immediate, niche
           | problems of individual people that they're willing to pay
           | small amounts of money or assign referral commissions for:
           | 
           | - Product personal shopper-recommender goes out and crawls
           | reviews and pricing to find what and where someone should buy
           | something based on a prompt.
           | 
           | - Where to live.
           | 
           | - Where to go out.
           | 
           | - What to have for dinner.
           | 
           | - Find a plumber, carpenter, etc. and do open source due-
           | diligence on them.
           | 
           | - How to optimally invest money based on circumstances,
           | assumptions, and speculative outcome distribution.
           | 
           | - How to redecorate a room.
           | 
           | In a business context, there are many classes of problems AI
           | can semi-automate including:
           | 
           | - Decision support
           | 
           | - Feature prioritization based on support data and social
           | media sentiment
           | 
           | - Supply (inventory) and demand forecasting
           | 
           | - Pricing optimization
        
       | TrackerFF wrote:
       | 1. Work for some prestigious ML/AI research lab / startup.
       | 
       | 2. Identify a demand in the market.
       | 
       | 3. Quit the lab/startup, launch your own startup.
       | 
       | 4. Get acquired by some established AI company or FAANG tech
       | megacorp.
       | 
       | 5. Retire.
        
         | Solvency wrote:
         | You don't need to identify a demand. Just make an AI startup
         | and nail the brand/marketing and you're equally likely to
         | attract offers.
         | 
         | It's a gold rush, get in now.
        
           | CapeTheory wrote:
           | So far it feels like a gold rush without much gold, just a
           | lot of very expensive picks and shovels.
        
             | candiddevmike wrote:
             | And all of the shovels have Nvidia written on them
        
             | fragmede wrote:
             | Because everybody's so busy selling shovels that no one's
             | digging for gold, so there's no one to sell shovels to.
             | 
             | That's not totally true, but we're all here trying to
             | outsmart one another in order to have a Sure Thing. Well
             | I'm here to tell you that there are no Sure Things in life
             | except for death and taxes, so take a risk, jump on a
             | random plot of land, and start digging.
        
         | hipadev23 wrote:
         | Up through #3, literally the playbook for the Google Research
         | team (minus one who took at job at OpenAI) that authored the
         | Attention is All you Need paper.
        
         | mattgreenrocks wrote:
         | In my old age, I have to say: this is not _just_ cynical, it is
         | smart.
         | 
         | It acts on the same dynamics companies act on. People build
         | reputations by doing this with far less effort than it would
         | take otherwise.
        
           | nine_zeros wrote:
           | It takes decades to figure out that playing the game under
           | layers and layers of management in mega corps - is equivalent
           | to undervaluing yourself.
        
             | appplication wrote:
             | Well... yes definitely. You are worth more than your
             | salary, otherwise your employer wouldn't keep you around.
             | 
             | The problem is that it's definitely a lot _easier_ to just
             | cruise along, punch 40 hours, and leave work at work.
             | Startup founder life is _rough_. I don't blame anyone for
             | not wanting to extract that worth from themselves.
        
               | nine_zeros wrote:
               | It is not easy to cruise along in some large corps. You
               | always have to be on guard about layoffs or managers
               | throwing you under the bus.
               | 
               | Not always true but it only needs to happen once to you
               | for your 40 hour cruising to end. At that point, you will
               | question why you spent so much time slaving away.
        
         | verticalscaler wrote:
         | There's clearly an Oceans 11 type plot here. Or perhaps The
         | Producers is more apt, not sure.
        
         | tkellogg wrote:
         | wow, so easy!
        
           | mhuffman wrote:
           | billionaires hate him for this one weird trick!
        
         | bongodongobob wrote:
         | 1. Win the lottery
         | 
         | 2. Retire
         | 
         | Probably similar odds.
        
           | pxeger1 wrote:
           | Yeah but at least if you fail at #1 you still have a job
        
         | ldjkfkdsjnv wrote:
         | This works in finance between investment firms
        
         | deadbabe wrote:
         | Why quit? Most people will mess up #2 or execute badly on #3 or
         | even then #4 is unlikely to happen when FAANGs can just copy
         | you and put you out of business instead of buying you out and
         | rewarding this behavior.
         | 
         | #5 will almost certainly happen even if you keep doing #1,
         | earning a huge salary.
         | 
         | The numbers and risks just don't make sense anymore.
        
           | sangnoir wrote:
           | > Why quit?
           | 
           | ...because there's a glut of investor money chasing "AI",
           | raising money for your new AI startup is not particularly
           | hard if you are linked with any of the big names in AI.
           | 
           | Even if the startup "fail", the founder could have been
           | paying themselves a market-rate salary with the added
           | possibility of an FU-money exit. The downsides are limited.
        
             | deadbabe wrote:
             | It's not that simple.
        
         | golergka wrote:
         | Didn't all authors of Attention is all you need do this?
        
         | stevofolife wrote:
         | What step do most find the most challenging?
         | 
         | For me, it's the first step. It seems once you get started,
         | everything else tends to fall into place more smoothly.
         | 
         | I'm eager to know others' perspectives.
        
         | scarface_74 wrote:
         | Because of the FTC, BigTech is more reticent to do an
         | acquisition.
         | 
         | So like in the case of Inflection.ai, Microsoft just did an
         | acquihire of the entire team and "licensed" the technology/IP
         | from what now is a shell of a company.
        
       | robotnikman wrote:
       | It makes me wonder if I should switch from my current position as
       | a Java dev to something related to AI and machine learning. It
       | seems that is going to be the future now, and there seems to be a
       | big need for people with that knowledge right now.
        
         | michpoch wrote:
         | Yes, if you're willing to invest X years into becoming relevant
         | on the AI/ML market, and you believe than in X years there will
         | be still too few AI/ML specialists - then it's a great idea.
        
           | khazhoux wrote:
           | There's a huge amount of help needed in AI/ML in non-
           | researcher roles (infrastructure, pipelines, testing, etc).
           | I'd say that the bulk of AI/ML orgs are actually just that,
           | in my experience.
           | 
           | (I'd further argue that the number of actual "AI researchers"
           | that actually contribute enough to justify their mega-million
           | comp packages is small indeed... but that's a separate topic)
        
             | robotnikman wrote:
             | Infrastructure was another area I was thinking of possibly
             | moving into as well, since AI of course still has to rely
             | on that as well.
        
         | AznHisoka wrote:
         | yep I did an analysis on this and that's definitely where the
         | software jobs are right now: https://bloomberry.com/how-ai-is-
         | disrupting-the-tech-job-mar...
        
           | digging wrote:
           | Good article, thanks for posting.
        
           | robotnikman wrote:
           | Interesting Article. Looks like it would be worthwhile to
           | dedicate some of my free time into specializing in AI more
        
         | zer00eyz wrote:
         | Dotcoms, then add tech (the money of Web 2.0),
         | crypto/blockchain, AI/ML...
         | 
         | If you are hearing about it, and you aren't already in, it
         | might be too late.
         | 
         | That having been said, it's a prime time to have project that
         | you can bootstrap into a startup. You dont need to be google,
         | you need to be google, small teams can clear a few mill a year
         | with the right product and not get much larger.
        
           | cm2187 wrote:
           | AI has this difference to the other booms that the resources
           | required to train a large model are out of reach of a student
           | in a college dorm.
        
         | khazhoux wrote:
         | I'd say this is a "yes", and remember that there's a _ton_ of
         | work that doesn 't involve actually training models. Every
         | company right now has a big need for help in infrastructure,
         | data pipelines, ops, monitoring + dashboards, internal tools,
         | CI/CD, release management, etc.
        
         | UncleOxidant wrote:
         | Probably too late now. At least for this cycle.
        
       | pksebben wrote:
       | Hey now, there's a real lack of information in this tweet. Just a
       | link to a paywalled article from a dodgy financial journal.
       | 
       | This is the second time this has been referenced on the front
       | page and I still don't know who they're talking about or what the
       | details are (and I'm not about to give money to _forbes_ , of all
       | people.)
        
       | ChrisArchitect wrote:
       | [dupe]
       | 
       | Some more discussion on Forbes article here:
       | 
       | https://news.ycombinator.com/item?id=39768402
        
       | Jackson__ wrote:
       | Interesting. It's not mentioned in the article, but the team
       | behind the recently released "Stable Cascade" models has left
       | stability as well.
       | 
       | Personally, I assume stability has hit a scaling/money issue on
       | image generation models, which is why they might be pivoting to
       | other domains where you can still make headlines spending <$50M
       | on training runs, like their 3D models.
        
       | Bjorkbat wrote:
       | I mean, since we're sharing Forbes links
       | https://www.forbes.com/sites/kenrickcai/2023/06/04/stable-di...
       | 
       | The CEO of Stability seems to be rather controversial person at
       | the helm of a rather controversial company. I'm honestly
       | surprised more people haven't left the company by now, unless
       | they had some absurd incentives to stick around.
        
         | achow wrote:
         | https://archive.is/NHdKM
        
       | artninja1988 wrote:
       | Sad to hear, they really created some awesome models for the
       | community. Looks like their entire diffusion team left. Do we
       | know where they went to? Hopefully another organisation pops up
       | without the incompetent discord based leadership and that
       | actually finds a way to finance themselves while releasing
       | weights in some capacity.
        
       ___________________________________________________________________
       (page generated 2024-03-20 23:01 UTC)