I like the idea of cloning my own voice and having it speak in a foreign language
2,943 karma · joined October 15, 2007
Occlusions, short-lived tracks, misassociations, low frame rate + high-rate-of-change features (e.g. flashing lights) are all still very challenging when you get down to brass tacks.
Also, pretty sure there are no DEI quotas for Asian men.
Even an understaffed team can crawl along with a skeleton crew for a while, but you will burn people out. A burned out team ceases to make forward progress and is going to act like things are in maintenance mode. You're liable to spend more money and SWE-hours looking for backfills vs if you'd just kept your team intact.
Without being an expert in this area, I imagine one such way would be to train a second neural net that learns to regress traditional workflow parameters to best approximate the NeRF output, then let a user post-edit those parameters to fine-tune.
If the NN doesn't get it right, it's absolutely not the end of the world. First of all, triaging the failures usually turns up something interesting / some connecting thread. And as for fixing issues, your options are typically:
1. Retrain with more examples to help it get it right (easier these days since you can take advantage of synthetic data / self- or semi-supervised approaches). Specifically it will help to generate or find examples that approximate the failures.
2. Shape your loss to penalize undesirable behavior / tendencies you've identified in triage
3. Add inputs or outputs that help your net solve the problem better (e.g. side inputs that may help the net not overfit to its available data)
These "inscrutable black boxes," while hard to edit directly, are not impossible to evaluate which means you can just measure to see if they're better than more explicit systems. And most real world systems are patchworks of ML + heuristics to give some desired level of performance.
Furthermore there is a big field of folks dedicated to elucidating what your net is learning (e.g. saliency). The tools are pretty cool -- for example in computer vision you can visualize whether your net is overfitting to background data in your training examples.
A valid critique of all of this is simply that it's expensive. Very expensive to build teams to warehouse and clean huge amounts of data, then build unbiased / fair model evaluations, then build models, then integrate them into real world stuff, usually to be used in conjunction with non-learned algos.
Another valid critique is that many companies have started using models they can't retrain or fix (e.g. off the shelf LLMs). This is dangerous territory since you lose so many of the options I outlined above.
Pixel is assuredly NOT leading in video recording quality though. And the overall package still matters: for example X% superiority in still photos didn't make up for Pixel 6 having a very small photo buffer / a slow camera startup speed. Didn't make up for lens cracking issues in Pixel 7. iPhone just works.
This is possible and even likely depending on how you hire.
But I think the OP might be assuming you can onboard whomever you've hired and instead referring to the relative incentives in each environment. For example, big companies tend to plan quarter to quarter, while small companies typically have a pile of things they need done urgently and are incentivized to help folks move as fast as possible (company runway may only be 1 year). That sets up a dynamic where large companies make you spend time derisking new project proposals and getting buy-in, in a way that small companies don't (this is where the agility comes from). Bigco also may have higher standards for getting something off the ground and a lower tolerance for code that is likely to be thrown away, whereas at startups you often write a pile of not-so-great code just to prove the thing works + understand your customer, with the knowledge that you'll have to throw it away when it's time to grow and scale.
One more annoying wrinkle: big companies sometimes have investments that mean you can't use the best tool for the best job. At Google you have to use borg and a very annoying deep stack of stuff to expose a service to the public -- you can't just buy a domain and spin up your website on a $10/mo server with nginx. You can't use React. You can't use PyTorch and just fork the hottest new model that came out -- you'll have to translate it into Tensorflow or JAX first. You also have to build using tools that assume and can handle scale, but sacrifice agility and developer ergonomics for that.
Oh one more thing: people at big companies are so obsessed with level. If you have good ideas and want to try them out but you're underleveled, you're going to have an impossible time geting buy-in from leadership to let you lead or even try the thing. At small companies if you have an idea that's good for the company, it's obvious to your leadership and they're just like go do it, you're in charge of this project, show me what can happen in a week. And you can make stuff happen. Low friction. Big companies might also have an entire team of PMs trying to decide what to prioritize (all the while playing political games since it affects their standing in the company), who also have to sync with eng and design, and then some pushback between the two happen. At small companies the number of people to get on board is just smaller and it makes things go faster.
YMMV. Finding and papering over the things that prevent a model from being deployable can also require a deep bag of engineering tricks that an average ML research scientist does not have. In my personal experience I've seen teams burned by this more often than the other way around.
They might just learn they need to add back modalities they neglected previously, or explore some new ones.
If you're familiar with the political infighting between Brain, Deepmind, and subteams within Brain, you'd know there was a lot of room for improvement.
Also Google would benefit from being able to use and build off more open source stuff. Tensorflow losing to Torch for researchers was a pretty big blow imo.
Macro cultural trends at Google didn't help either. Stuff like not incentivizing work with longer term payoffs, and too many layers of middle management.
> Only that the tech became production-ready and the rest of the world took notice, and these researchers and engineers decided that they no longer needed Google's money since there was so much available from everyone else.
All that said this is very true and probably unavoidable.
A more correct statement with the same thrust would be "copyright is not _natural_ -- it's an artificial, bolt-on construct."
And yet lots of artificial constructs are useful for incentivizing / de-incentivizing massed behavior in a way that is maximally beneficial to all. For sure, getting it right is difficult but let's not write off all artificial constructs as worthless or impossible to get right (especially when you just have to get it right _enough_).
For whatever they're being written for. Alignment's goal is to have models to do what they're being trained to do and not other random things. It won't be uniform; for example, determining "what does inappropriate mean" will vary between countries.
Would be nice if rents were a smaller total of our living expenses though. Everyone would have more breathing room.
Labor aside, a 2 bedroom in Mountain View is quite expensive and doesn't come with learning materials, toys, classmates, etc.
> You can't actually get in. There was a waitlist and your position on it actually got worse over time
This changed after covid. My understanding is that they struggled to fill classrooms due to people moving away from MTV / Palo Alto. We tried to convince friends to enroll their kids for fear of something like this happening. The closure was way worse than we expected (we thought they'd at least do a phased closure) and we kinda regret the advice we gave to friends since they're scrambling to find replacement care now.
> From what I heard recently from a daycare provider (I can't directly verify), it wasn't very good if you did get in
Daycare turnover has been bad industry wide, and our classroom had bad turnover in 2021 but it stablized since then. We've super loved our experience. Really great sanitary procedures, high quality learning objectives and measurement, healthy food, etc.
> It was supposedly subsidized but still strangely expensive.
Yes it's still shockingly expensive. Maybe this was self inflicted at first: I heard that when one of the Wojicki sisters had her kids there they changed out all classroom furniture every year, looked for way overqualified people to do the childcare (like Ph.Ds) and all that. It's not that over the top anymore today, but today they still pay the childcare workers very well by that industry's standards (they also get all the same days off as other alphabet workers) and the facilities are way bigger than others that you'd find on the peninsula, with a playground / yard for every N classrooms.
My personal feeling is: I understand why they'd cut this benefit. I _don't_ understand why they had to cut it in this fashion. Having another daycare company buy it out and handing over operations, or raising prices so it's not subsidized, or phasing out classrooms over N years to give parents plenty of time to find alternative care -- there are just better ways they could have ended this longstanding benefit.
Maybe a specific young person with a lot of relevant experience and strong desire for this sort of life could, but it would be pretty hard and the initial costs are nothing to sniff at.
Water is especially prohibitive. The American West is going through massive aridification and the water table is overdepleted in many places. Gotta watch out for water rights too.
Electricity and internet are doable if you can afford the right systems.
Food is really hard unless you're willing to buy everything at dollar general. Modern farming is highly reliant on artificial sources of nitrogen. And it's just not a good use of your time if you have a tech job. Social needs of humans cannot be overlooked.
People even a couple hundred years ago would move whole communities. They had access to abundant natural resources and divided responsibilities differently.
Maybe you're just interested in van life but + farming? It's expensive and probably a bad investment overall, but makes more sense than owning random cheap but poorly situated land outright.