Yes, the concept is still powerful and in use today.
As I understand the RLHF method of training LLMs, this involves the creation of an internal "reward model" which is a secondary model that is trained to try to predict the score of an arbitrary generation. This feels very analogous to the "discriminator" half of a GAN, because they both critique the generation created by the other half of the network, and this score is fed back in to train the primary network through positive and negative rewards.
I'm sure it's an oversimplification, but RLHF feels like GANs applied to the newest generation of LLMs -- but I rarely hear people talk about it in these terms.
I think diffusion models are useful too, I’m currently working on a project to use them to generate medical type data. It seems they'd both be useful as they are both targeted towards generation of data, especially in areas where data is hard to come by. Doing this blog made me wonder of the application in finance too.
I agree -- I would love to see diffusion models applied to more types of data. I would love to see more experiments done with text generation using a diffusion model, because it would have an easier time looking at the "whole text" rather than the myopia that can occur from simple next-token prediction.
To my understanding, it does not cache (but I haven't reviewed the code). "Expensive" here really just means expensive in time. If you're running DuckDB on an EC2 instance to query S3 on the same account, it's practically free of any other cost. I wouldn't bother doing it outside AWS unless it's a one-time thing.
Running a simple analytics query on ~4B rows across 6.6K parquet files in S3 on an m6a.xl takes around 7 minutes. And you can "index" these queries somewhat by adding dimensions in the path (s3://my-data/category=transactions/month=2024-05/rows1.parquet) which DuckDB will happily query on. So yeah, fairly expensive in time (but cheap for storage!). If you're just firehosing data into S3 and can add somewhat descriptive dimensions to your paths, you can optimize it a bit.
If the parquet file includes any row group stats, then I imagine DuckDB might be able to use those to avoid scanning the entire file. It's definitely possible to request specific sections of a blob stored in S3. But I'm not familiar enough with DuckDB to know whether or not it does this.
I asked this some time ago on their Discord in relation to AWS lambda and the Python client and the answer was that you need to handle caching on your own but it is easy to do with fsspec. I haven’t tried it yet though.
Duckdb over vanilla S3 has latency issues because S3 is optimized for bulk transfers, not random reads. The new AWS S3 Express Zone supports low-latency but there's a cost.
Caching Parquet reads from vanilla S3 sounds like a good intermediate solution. Most of the time, Parquet files are Hive-partitioned, so it would only entail caching several smaller Parquet files on-demand and not the entire dataset.
Stuff that works well in homes often is a lot more complicated to implement in restaurants, where you're: a) constantly fighting grease buildup and hard-to-remove dust that clings to greasy or damp surfaces, b) often have a profit margin of like 2% if you're one of the successful ones, c) aside from looking clean, you have to worry about pest control, fire codes, health codes (you can't have built-up dust falling in people's food, d) etc etc etc etc. Also, how restaurants look is as, or in some cases more important than the quality of the food. A good, attractive, practical restaurant design is one of the things that can steer you towards success or failure. Much to many chefs chagrin, hip and attractive restaurants with shitty boring food are often more profitable than ones that only focus on the food. Marketing is annoyingly important.
With, floors hardwood is a hard surface (so only mildly sound damping) so they're not too bad for cleaning and health stuff, but are expensive to install and take a lot to maintain if the worn-in look doesn't fit the aesthetic. Low-pile carpets can be shampood inexpensively for medium-term maintenance and replaced comparatively cheaply in the long run, but take a lot more effort to keep clean when someone drops a catering tray full of crème caramel and something with a port wine reduction.
Artwork: anything that you'd want hanging on your walls is either going to need to be a print or covered with glass or plastic because it will get ruined otherwise.
Acoustic panels are usually pretty ugly, difficult to clean, not resistant to pests, are a fire liability if coated in grease, etc.
Curtains definitely are definitely viable, but if you've got enough of them to really impact the sound level, they probably need to be expensive ones, and expensive curtains can't just be tossed in the wash and pressed on an ironing board.
It's not like they aren't effective, they're just not nearly as easy to deploy or maintain as they are in homes or offices.
Unrelated blathering because a lot of folks in tech don't have much exposure to this stuff and I always enjoy seeing a slice of someone else's life: In general a lot of people are understandably perplexed by seemingly simple, avoidable problems that they encounter in restaurants-- you can chalk almost all of them up to misinformation, or deliberately obfuscated factors. Firstly, there's a ton of inaccurate folk knowledge about the way restaurants work... (most infuriatingly to me is the food safety stuff. Look up the incubation time for most foodborne illnesses and consider how many people blame some lower GI symptoms the meal that met their stomach lining 3 hours earlier.) Also, a big part of the restaurant mystique is making it all seem sort of easy, uncomplicated, and fun, even for regulars and the 'friends and family' crowd; underneath that thin veneer, it's absolute insanity. I've worked in tech and the restaurant business extensively. Most days, the pressure level is "we just discovered a possible active intruder in our production systems" for at least a few hours. It's exhausting, and one of the reasons drug and alcohol addiction is so prevalent. Knowing that an entire staff is breaking their back so you can have a fun cozy bite to eat makes the experience palpably worse, but it's true. That's why you'll usually find people who've worked in the service industry are serious over-tippers. You have to give up a lot of your humanity to do that work, and a lot of people you encounter respect you less instead of more for having made that sacrifice.
I've proudly convinced so many people to not go into that business, though I've also convinced a few people to give it a shot. It's not a good choice for most people, but some people can't really do much else and be happy. In many ways, its especially tolerant to neurodivergent folks with different skillsets being downright useful in different roles. It's hard as hell though. There's a good reason that CIA (the school, not the spies) requires 6 months of full-time back-of-the-house restaurant work to get admitted to their degree program.
There definitely are but, perhaps by definition, items soft enough to dampen sound are often easily damaged so they aren’t great fits for most commercial locations.
They are also out of vogue as was mentioned, unless you’re a coffee shop then these “cozy” items just aren’t as common right now.
Honestly the cheapest way to muffle sound is to not create it in the first place. Guests make noise to hear themselves over other guests and the din of the room, the quieter the room, the quieter the guests, etc.
Essentially, the louder the noise floor, the louder the signal has to get to be intelligible at every table, which raises the noise floor, creating a feedback loop. Good acoustic design in a space accounts for this by minimizing how much acoustic energy is present in the room - both by removing it (with acoustic treatment), spreading it away from sources (by isolating tables/booths, using hard surfaces to reflect sound away, etc), and preventing it from being created in the first place. For example, keeping bus stations behind galley doors and training staff not to clink silverware/glasses/dishes when filling bus bins and avoid playing loud music, etc.
In my experience, most restaurants fail at this because all the people who do it well are in the high-end restaurant business, which most restaurants are not. If the key to a space that isn't too loud is to limit the number of patrons, have dining room space allocated to treatment between tables, have highly trained staff with consistent management, and a big enough kitchen space with heavy enough doors to isolate the sound within - your only option is to be a high end restaurant.
But the high end places fail at it because they don't care and want to maximize the guest throughput because their margins still suck.
High-end places not only still have bad margins, they're quite often worse!
Low-end places are often even more carefully designed, though, but they're designed for different things: high turnover and low staff wages, meaning simple, flat, easy-to-clean (and sanitize), nearly-zero-maintenance surfaces like bare laminated tabletops and quarry tiles. Especially once you start moving into fast casual, they want their diners comfortable enough to enjoy their meal, but not comfortable enough to linger, which is often a tough balance to strike.
The look of high-end places is like putting your sales people in nice suits. I don't think most places are trying to maximize throughput-- if they're not completely booked at least a few nights per week they're probably not staying open-- I think they're trying to maximize check averages. Nothing inspires "maybe I'll get the dry aged wagyu app and flip to the expensive page in the wine list tonight" thoughts like a luxe dining room.
Yeah, by integrating art with sound-absorbing materials, it's possible to enhance the acoustic environment and the interior design. I wonder if there is a place where foam art is on display?
