Jerod Santo: Grepping logs remains terrible

algernon first stated his public opinion that grepping logs is terrible a decade ago. The good news is he’s back, the bad news is his opinion hasn’t changed. After a quick comparison of running alike queries against 4.4 GB of JSON logs on his modern M2 SSD desktop and loading the same data into VictoriaLogs on an old Mac Mini, he finds:

What we saw here is that a computer with much more RAM, and computing power, and faster storage got absolutely obliterated by an overloaded Mac Mini that is ten years its senior. And not only in speed, but resource use too: I haven’t shown it, but working with raw text used a lot more CPU time, disk IO, and memory; meanwhile the Mac Mini barely blinked.

And not because it was all cached in memory! These are all cold queries. Cached queries are much, much faster. But how can it be that such an underpowered device runs circles around a powerful desktop?

Purpose-built software, dear reader. Purpose built software.

The moral of the story: databases are good!

Don’t grep your logs. Shove them in a log-ready database and query that. Not only will it require considerably less disk space, it will require less computing power, less RAM, and as a result, less time to perform the queries. And if that’s not enough, you can give the logs structure! You no longer need to remember the order of fields in a httpd’s Common Log Format, you can query against a named field instead. You can run aggregations, computations, build stats, correlate, join, and do all kinds of other database-y stuff. Not only does it perform better, you can do more, and do so more conveniently.

Solomon Hykes: I think it’s a combination of things. One is containers had to exist and be ubiquitous, because you needed a way to make all that stuff reproducible; otherwise it’s not worth it. There’s a family of tech called DAG tech that just had to mature. The ability to execute tasks in parallel, with dependencies modeled as a graph… I mean, make has been doing that forever, but it’s been kind of doing it on its own. But recently, in the last 10 years, you’ve seen things like Bazel, Buck, Nix… Everyone’s sort of independently exploring this model. And it’s really matured, and one of the big benefits, honestly, is caching. So everything’s got to be cached, or cacheable, for the thing to make sense. I think also just the amount of computing power that’s available, I guess… I mean, we joke that the biggest – you know, we ask teams, “In your organization what’s your number one provider of CI compute?” People are like “Well, we did the migration from Jenkins to GitHub Actions, so I guess maybe Azure now… No, I guess it’s still the original…” But the answer usually is Apple. So you’ve got these beasts on every desk. That can run all that stuff just really well. And people don’t realize it, because there’s this software blockage. You’ve got this platform there, in the sky, and you can’t run it locally, so no one’s trying… But if you try, and you add caching, and you add containerization, you’re like “Wow, this is actually not as big as I thought it was. Why am I paying $5,000 a month to run this, and it just completed in three, four seconds on my Mac?”

Chris Benson: Yeah, they do attempt to kind of address that. There’s a section, 4.2, “Ensuring safe and reliable AI”, and they talk about a timeframe… And this is one of those that calls out the Defense Production Act specifically. And within 90 days there’s a set of things that are being required… Which is a very short timeline, when you think about it. Part of that is they have a set of criteria, which is certainly not comprehensive. To your point just a moment ago, it’s better than I expected, but it’s not entirely sufficient. There’s a lot of nuanced questions, as you just raised.

I noticed that one of those, they talked about the quantity of computing power used for training, and they have picked, interestingly, 10 to the 26th integer or floating point operations for kind of as a general point of computation; it’s like a threshold, and if you’re above that, you’re kind of in that large range that they are particularly focusing on. They have reduced that down to 10 to the 23rd integer or floating point operations for models that are based primarily for biological sequencing, and things like that. They have some other aspects in, but they have an interesting threshold that they call out in the executive order.

Gabriel Ortiz: Yeah. I think, of course, you have to bear in mind that we have limitations. What happens to me also happens in teams in India, or in the US, that I am always seeing what they are doing. I would like to point out two limitations. One is the computing power, and another thing is the limitations of CNNs, which is the technology that we are using right now, convolutional neural networks. We can talk a little bit about model architectures and things; in terms of computing power, I think it’s worth delving into the role of GPUs, because in the geospatial realm it is not well understood why do we need a GPU. And it’s something – I don’t know if it happens in other markets, but in our industry, when you talk to somebody about a GPU, normally, my fellows and mates, I don’t know, they try to say it’s something related with the IT department; “I don’t want to be in charge of that.” But it’s not at all. You have to be aware of what technologies do you have for calculation; the hardware is so important, and you have to speak the same language as a data scientist, that the rest of the community speaks.

And that is very, very important to understand that it’s not the same as the GPU in your laptop, the DGX, or talking about the [unintelligible 00:37:46.23] We are talking about NVIDIA hardware… It’s not the same. And it’s everything related with the amount of data that you want to put into the train, the quality of your training, the level of the convergence that you are going to get if you are going to stay in a local minimum in the convergence, or you are going to reach on a source the possibility of the level that you are ingesting into the model… Everything is related with the hardware.

[38:16] I think that Bill Dally and Anima Anandkumar, in many of their talks, always talk about the trinity of AI. One of them is the data, another is the software, the algorithms, and many of them have been with us for a while… Back-propagation, many of those algorithms have been from the ‘80s, if not before… But the hardware is the third part. Bill Dally always says that it’s the spike that starts this engine of creativity of AI. And I think it’s true. You have to pay a lot of attention to computer power.

And there is salary another limitation that is ingrained in the DNA of the CNNs. As far as I know from my experience, you cannot expect to perform exactly as a human being, and sometimes in spite that you curate very well your labels and your [unintelligible 00:39:11.12] and your data, the models do not learn as good as you expect. But somewhere in between, you could have a reasonable amount of success in that. What we do to overcome is combining different model architectures; it’s something very useful and widespread in the geospatial industry. For instance, we combine models at two different levels. From the architectural levels it’s quite common to see the combination of ResNet with, for example U-Net. In ResNet you remove the last part, the fully connected layers, and connect the remaining part with U-Net. So you are using Res-Net for feature extraction, and then the rest of the decoder happens during the rest of the U-Net architecture. It also happens with Mask R-CNN. We use constantly Res-Net as the backbone, but then the rest of the model goes.

And there is a second point, which is combining the results of the inference when you have inference from two different model architectures. For example, talking about vegetation - imagine that you have one model that detects very well the big areas of vegetation, but fails in the smaller spots. And you have another model that works very well for smaller spots, but fails detecting the big areas, because the big areas creates artificial holes, and mistakes… You can combine the results of the outcomes of those modern architectures with traditional GIS techniques to mash all the results together, and obtain a bigger, the best quality of the layer that you want to infer. That has worked for me, and it’s one of the ways that we are trying to overcome the limitations of artificial intelligence.

Erin Mikail Staples: But he runs a SQL database. It is actually very good – he’s very awesome updating it; it’s available on Reddit. I can share it with you, and you can post it in the links. But I’m just playing around with the dataset for fun. But in this context, I’m playing around, building demos, and just having some fun, teaching myself some new skills. I don’t need a large foundational model for that. And I think going back to your original question of “Well, all these models are getting smaller, and more accessible, and we can run it in a notebook…” We don’t need the high-powered computer models, everything single times. And if we stop and think about the context of the problem that we’re trying to solve, it can give us a lot of answers, and it can save us time, energy and computing power. That’s why I get really excited about being on the data labeling side. Again, I have a background in humanities, I’m a self-taught programmer, but I think – I don’t wanna be like “We need more people like me in data science”, but we need more of the humanities in data science, because we’re missing the context.

David Heinemeier Hansson: It’s not only absurd, it’s actually revolting. I think we’re spoiling one of the greatest achievements of humankind, the internet. And we’re willfully regressing on all of the key components that made it so unique, and we’re handing it over to a bunch of monopolists to run as they see fit. That’s just dystopian. And I feel that because I’ve literally been working with the internet since the very first browsers came out. And I owe my entire career to the internet. I owe much of my life to the internet. So to see us spoil it in this slow-boil way is just heartbreaking.

Now, I understand why. It’s not like there aren’t reasons why people enjoy, or perhaps even prefer the cloud… But a) I think those reasons are overstated. I used to assign to those reasons. This whole moniker of like, “Well, you wouldn’t create your own power either. What, are you going to run a power plant in your backyard?” That was very compelling, I think, in the early days of cloud, and now it’s very uncompelling, as people have realized that it takes just as much f****n’ effort to run in the cloud as it does to run on-prem. In many ways, it actually takes more effort. And in many ways, the complexity of racking a server and connecting some Ethernet cables and so on is dwarfed by trying to understand the 150 services AWS puts out, or how you do permissions, or how you do cross-routing; all the other complexity of the cloud is there, it’s at least as much.

[38:24] Now, that is not to say it does not have distinct advantages, and we enjoyed some of those advantages very much when we launched Hey. When we launched Hey, we had scoped and set up the infrastructure to accommodate I think 30,000 people trying out and using the system over the course of several months. Well, we had over 300,000 people show up in three weeks. On the cloud, we could push a few buttons - it wasn’t quite that simple, but at least the mental image of it is simple, where we go like “Boom! Now we have like 10 times the capacity.” Now we can deal with 300,000 people showing up in a few weeks, instead of 30,000 people showing up in a few months. That is a difficult proposition if you had been on-prem. So that’s what the cloud does really well, it does spikes very well. So if you have this incredibly unpredictable business, the cloud does that better than anyone else. Now, is that generally our business? No. I can predict the amount of storage, for example, we need for Basecamp a year-and-a-half in advance, because it’s such a stable, predictable business. The vast, vast majority of SaaS software, and certainly all of b2b SaaS software is incredibly predictable, which is why in part it’s such a good business. But it also means that the key advantage of the cloud is not that pertinent. And in fact, it is almost sort of – it’s the downside, right? Like, you rent your computers.

Now, what do you think is cheaper - to rent your computers or to buy your computers? Now, it is cheaper upfront to rent the computer. So if you have no money at all - great. Okay, fine; you don’t have to buy the machines. But if you have to rent the same computer for five years in a row, you’re going to pay for that computer five times over, at least; in some cases 10 times over. Right? The expenses involved with the cloud are evident in the profits that AWS posts every quarter to its parent company. AWS is responsible for something like three quarters of all profit of Amazon. Now, imagine that - Amazon sells billions and billions and billions of dollars worth of merchandise every day on their website, and it’s barely profitable. They rent out a bunch of infrastructure to tech companies, and that is wildly, obscenely profitable. What do you think that money comes from? It comes from your f****n’ pockets, right? Now, there are some economies of scale, so they can obviously buy some things… But those economies of scales are not like 10 to 1. It’s not like Amazon could buy the same CPU for a 10th of the cost that you can buy that CPU. I don’t know, do they get 20% discount, 30% discount? Whatever it is, the math doesn’t work on the long term if you have a stable workload.

And that’s partly where we have been with our business since the beginning. Now, Hey, was different, but now it’s not so different. Now it’s a far more predictable thing. And when I see the amount of money we send to AWS every month from Hey, I cry a little, because it just looks like, “Geez, if we take that money and we invest it in our own software, a) we get out from under the big tech monopoly of this software, b) we’re not contributing to the death of the internet, where we reduce all the computers that run it to these three data centers, and c) we’ll actually make a bunch of money. We’ll own a bunch of software.

And what I’ve also found, particularly right now, we’re in this fascinating phase of hardware infrastructure where advances are actually happening again. Now, for the longest time, 2010 through ’18, or something, how much extra power they would get out of those Intel CPUs? Bubkes, Nada, Niche. You could barely tell the difference.

[42:07] Suddenly, TSMC and Apple have turned up the heat on the manufacturing processes, and we have AMD putting out five nanometer chips, and you’re suddenly seeing like 20%, 25% performance jumps generation to generation. And Apple has been doing this for multiple years running straight. It’s why those M1 and M2 chips are so ludicrously more powerful, in particular on the laptops, than any of the competition, because there’s finally stuff happening in the hardware realm again. And you get to be involved with that if you buy your own *bleep* computers. That’s fun. Fun in the same way as that example you refer to where we laid out all the RAM we bought for the machines back in 2012.

And the second thing is, computers are getting ludicrously powerful, particularly in terms of storage. And that is one of the things that just fascinate me. What changes, at which point do we have tipping points where - for example, Redis is a tool we’ve used for the longest time, and we used it for a bunch of storage, and that obviously runs all in RAM, but now you have NVMe cards that can do 10 gigabytes a second with a PCIe 5 and you’re going like “Whoa, maybe we don’t need to use RAM for caching anymore. Maybe we don’t need to–” It just opens a whole new domain; it feels like we’re a little bit back at the frontier, in much the same way as I recall like the Pentium days; we went from a 486 to a Pentium and you’re like “Holy s**t, we can do new things in new ways.” There’s some of that going on right now, that I don’t even think the industry at large have internalized, and part of the reason they haven’t is because so much goes straight into this amorphous cloud. If you now knew how large of a site or an app that you could run on a single damn computer…

I’ll give you an example. We just bought some new NVMe cards for Basecamp. 12 terabytes worth of storage - $3,000. You’re like “That’s so wild.” In 2012, when we were spec’ing out the hardware for Basecamp 3, we were buying these exotic, not available to mainstream Fusion IO NVMe cards that were like - I think it was $30,000 for like two terabytes, or something… Like, orders of magnitude less. They were orders of magnitude slower. I don’t think most people realize just how much computing power, how much storage performance you can get in a single box. I would say that probably 95% of – I mean, I’m just f****n’ shooting in dark here, but for comedic effect, 95% of all b2b SaaS software could run on a single box under the sysadmin’s table at the company. And there’s so many companies spending 20k, 30k, 100k, a million dollars a month renting a bunch of s**t from Amazon or Google or Microsoft, when they could just buy some of this new, amazing tech and stick it under the desk.

So I think we’re due to a pendulum swinging back here, where people realize, “Hey, you know what? The cloud was not actually that great.” And the reason it wasn’t that great was it was no less work; it has just the same amount of complexity.” Now, that’s not true in all regards. If you use something like Heroku, or Render, or some of these other fully – there, you can get away with an actual reducement in complexity. But if you are trying to set up Kubernetes, or run that yourself, or run that through Amazon, or whatever, the complexity and the expertise needed is just as high as it was on-prem, and the costs are increasingly out of sync, in my opinion, with what you can do yourself.

Theresa Kushner: That’s an internal discussion always. And every customer’s environment is different, too. A lot of customers’ environments are not set up to do this kind of data exchange. And so that’s one of the reasons why NTT Data Services looks at that, is because that’s what we do, is that we help you create those kinds of environments that are data-rich and responsive to all kinds of applications. And so we’re in the troves of trying to make sure that when we release these technologies, that we get companies that have the backbone in order to do it to begin with. Because it’s just frustrating to companies. Like you said, the computing power, the capability… A lot of what we do is in the cloud, but then again, you have to push our friends from the hyperscalers to do high-performance computing, at a cost that’s available to ordinary humans. So that’s another area that we look at constantly.

Adam Stacoviak: This is like high-performance computing power, or is that the kind of machine to run this kind of stuff on?

Nicholas Harris: Well, I would say the biggest deployments of AI are in hyperscaler environments and in cloud environments. To play in those spaces, you need to avoid doing anything weird. A big dilution refrigerator or a vacuum - that’s going to be a problem. And if you look at photonic computers, the things that we’re developing, they just look like a normal silicon chip. They do have optical fibers that come out. You do need to get light into the processors, so that it can do the calculations, but it’s really just a standard computer chip that looks like a card sitting in a server, but there’s a little addition of having a laser in there.

Maybe just on the larger picture here - there are so many different ways that people are looking at powering AI computation, because the market opportunity is so big… And we’re all trying to power these 500 billion weight neural networks. It’s going to take a lot of computing power, and that power is the principal problem. Computer chips are getting way too hot. That’s really one of the fundamental things that we’re interested in trying to help with.

Jerod Santo: So before we continue down the technical road of what this all looks like, I wanna take you back to that conversation you had, because I think that was a pretty good analogy with regards to the streaming of games or that direction, and the streaming of the application, kind of its final state down the wire, versus making the client do a whole bunch of computing… And I wonder if there’s a fly in the ointment… Some of these things in the industry - you just don’t see them coming. I think back to the iPhone; prior to the iPhone, I remember John Siracusa was making these blog posts about how Apple had to replace Objective-C, like, yesterday. They needed a higher-level language in order to compete with other platforms.

Then the iPhone dropped, and the App Store dropped, and the iPhone’s computing power compared to desktops was just minuscule… And Objective-C, because of its low foundations and it’s manual memory management and all these things that make it not a very desirable high-level language actually gave Apple this advantage for a few years, where their phones and the apps on their phones could outperform a lot of the competitors, because they didn’t have GC and these other things that Java has on Android, for example. And that bought Apple probably five years before they had to come out with Swift… So that’s just an example of like “Well, nobody saw that–” or maybe Apple saw it coming, but those of us watching said “You really need to do this”, and it turned out something changed… They didn’t really have to do it for a while at least, and then Swift came. I’m not sure how many years later Swift came; at least five, maybe ten years later… And now they’re competing at that higher level.

Well, one thing that’s changed with regards to apps is privacy and security… And I think about this with regards to the client-server relationship. I think today we are less likely to trust the server than we are our own clients. We see pushes for offline apps, we see pushes for phone-only apps, where the database is only local; and these are things that people are trying to build. And we see the ad blockers, the tracker blocking - this whole movement towards privacy, because our privacy has been abused by people running servers for a while now; we’re waking up to it. It has perhaps pushed people away from the idea of like “Well, I’ll just stream everything down from this server that I can trust”, because I’m not sure I can trust that server. I just wanna know what your thoughts on that – maybe a fly in the ointment of the idea of as much server-side as possible.

Will Ramey: It’s been a wild ride, both for me personally and for the company coming up on two decades here… When I joined the company, we were viewed as a chip company; we designed chips, that’s kind of our business. And I had the opportunity to work as the product manager for some of our developer tools, and then do some program management of large software projects we were working on with Microsoft and others, and work a little bit in our embedded business… And slowly but surely, the company started to climb up the value chain, from chips to boards and embedded systems, kind of system-level products… And then in 2009 we introduced this new technology called CUDA. CUDA, as many people know and love it today, is a parallel computing platform that gives developers access to the parallel computing power, the accelerated parallelism of NVIDIA’s GPU processors. And that was exciting.

Scientists, researchers were starting to figure out new and innovative ways to take advantage of this computing capability… And I moved into a role as the product manager for CUDA. So introducing this new technology, talking with people who were trying to make sense of it and apply it to some of their most challenging problems… And that gave us the opportunity to go and really start working with people who we’d never worked with before. Working with Fortran programmers, working on climate and weather simulations, working with people in the energy industry, trying to figure out what does the subsurface of the Earth look like. People who were worknig at the very smallest scales of physics, trying to figure out molecular dynamics and how things work at the quantum levels. Really big, sophisticated simulations.

Ben Johnson: I find it crazy - with Moore’s Law we’ve just expanded computing power at a crazy rate, but somehow I feel like our applications are slower, because we just keep adding all these little bits of cruft here and there… It just makes me sad.

Jaana Dogan: That’s a really good question. I think there’s one more thing that we should consider - the world was not disconnected… So I think in the last ten years people are really looking for faster results; they’re looking for more complicated queries. We just realized that this is a trend at Google especially. In the beginning I think most of our systems were much simpler.

Over time, people’s tolerance to latency was going low; at the same time, applications want to provide more interesting stuff, so they have to run more complicated queries, join more stuff together… And especially, think about a social network - there’s all these different data models and different databases you need, because the data is becoming way too connected. You have your friends, and friends of friends, different schemas, connections and filtering requirements and so on. Everything is becoming more complicated, and at the same time, our networking for example is getting better, our computing power is getting cheaper… So there are definitely different trade-offs nowadays.

Catherine Breslin: Yeah. One thing I should mention is that there’s what I’m talking about in terms of breaking it down into acoustic and language model and lexicon is a technology which is very common right now, but in the research community there’s a lot more effort to come up with new neural network approaches to build one speech recognition system which isn’t decomposed into these three parts… But I’ll carry on talking about the three parts, because that’s how a lot of our commercial speech recognition systems are built right now.

So the lexicon, which is basically a phonetic pronunciation of lots of words - often we have really good, high-quality handcrafted lexicons. They’re written by phoneticians, and you can imagine for a language like English we have good lexicons, which have been developed over many years.

Now, obviously, language is always changing and new words come up, and new ways of using language come up… So we still have to be able to predict pronunciations of new words, but we have very good lexicons to tell us pronunciations of the vast majority of English words already… So predicting the pronunciation of new words is a smaller task.

The acoustic model and the language model then end up being statistical, machine-learned models… And the language model is trained on text, and the acoustic model is trained on audio. So two different models, trained on two different types of data.

The acoustic model, for instance - we’ve had great success in recent years… If you think back like 10-15 years ago, speech recognition systems didn’t really work all that well, and they were using often for the acoustic models Gaussian mixture models… And one of the things that’s really happened in recent years, in the past 10 years or so, is the switch from Gaussian mixture models to neural network acoustic models. That’s had a really big impact on the performance of speech recognition systems. It’s made them significantly more accurate over the past few years. That’s come in tandem with other advances in computing power, and storage, and memory, and computation, and all these other things… But the actual switch from Gaussian mixture models (GMM) to neural networks on the acoustic modeling side has really made speech recognition systems much more accurate.

Chris Benson: [12:24] Sure. And even going on to the next point, where they talk about making the federal data and computing power available for AI purpose, as you mentioned before, it’s very generic talk about sharing data models and computing resources with researchers in the private sector, and it notes that agencies are expected to help those researchers access those resources, but kind of stops at that point. So it kind of states the obvious on what we need to do in the background, without making any kind of leap or strong directive in a detailed sense.

Corey Sanders: It’s fairly simple. It effectively runs containers and manages the health of the containers running inside that edge device. In the example today, the camera - it could run containers inside the camera, and the camera had enough computing power to be able to sort of do that, and do that work. That included cognitive services, it included functions, it included those aspects, and then the IoT Edge was basically taking the containers that were built in public Azure and accepting them and deploying them, and sort of managing them so they could take updates, and they could communicate back, and they could communicate the health of the containers, and so on. So it ends up being sort of a hosting platform for the actual containers running on it.

Todd Gamblin: Well, because the computing power is what enables you to do more and more with your computer. You can do many more computations, the thing gets faster, it’s good. I mean, I think the main consequence of the breakdown there is that you don’t get as much single-thread performance as you used to. That’s kind of kept out. So if you’re using Microsoft Word, you’re typing, or something that has to execute sequentially, it’s not going to go any faster. But if you can get parallelism out of your workload, then you can actually harness all the power that’s on your chip.

The difference is that if you increase this clock frequency, that just means that everything on your chip is happening faster, so that’s effectively free. If you had a program that ran on an older chip, it would run just the same on a newer chip, just faster, whereas with parallelism, if you wanna harness that performance, you actually have to rework your program.

Luis Villa: So I’m a former programmer… I got a CS degree in the last millennium, and worked in open source for a while. I got involved in open source while I was in college… Actually, originally, hacking on the LEGO Mindstorms, the very first generation of LEGO Mindstorms… And also just as a Linux user. This idea that people were building an entire operating system together on the internet was like… But I was also a political science major, and so I was very interested in this overlap of politics and power and computing.

It started off as just like I was interested in politics and I was interested in computers, and I really thought those two weren’t related… And then by the end of the ’90s it was like “Oh yeah, actually these are super-related.”

[04:04] So I worked at a startup called Ximian, which worked on the GNOME Linux desktop. Then, after that I was like “Actually, law school sounds like it would be fun.” Pro tip, kids - law school, not fun.

Since then I’ve worked at a series of Mozilla as an attorney, where I worked on the Mozilla public license, revision version 2.0, I worked at a big law firm for a while, working among other things on the Google Oracle lawsuit for Google, where fair use came up quite a bit, which is something we’ll talk about today. Then I worked at the Wikimedia Foundation, and now I’m the co-founder of a startup called Tidelift, where we are trying to make open source better for everyone by helping build a sort of economic and payment loop so that maintainers get paid to do all the sort of not fun parts of maintenance. They’re not fun, but they’re really important for businesses and enterprises, so we’re trying to close that loop as a business. But yeah, I’m a copyright nerd at heart, so I think that’s sort of why I’m here today.

Luis Villa: And also just as a Linux user… This idea that people were building an entire operating system together on the internet was like… But I was also a political science major, so I was very interested in this overlap of politics and power and computing. It started off as just like – I was interested in politics and I was interested in computers, and I really thought those two weren’t related, and then by the end of the ‘90s it was like “Oh yeah, actually they’re certainly super-related.”

I worked at a startup called Ximian, which worked on the Linux desktop… And basically then after that, I was like “Actually, law school sounds like it would be fun.” Pro tip, kids - law school, not fun.

[04:09] And since then I’ve worked at a series of Mozilla as an attorney, where I worked on the Mozilla public license, revisioned version 2.0. I worked at a big law firm for a while, working among other things on the Google/Oracle lawsuit for Google… Where fair use came up quite a bit, which is something we’ll talk about today.

Then I worked at the WikiMedia Foundation, and now I’m the co-founder of a startup called Tidelift, where we are trying to make open source better for everyone by helping build a sort of economic and payment loop so that maintainers get paid to do all the sort of not-fun parts of maintenance… They’re not fun, but they’re really important for businesses and enterprises, so we’re trying to close that loop for the businesses.

But yeah, I’m a copyright nerd at heart, and so I think that’s sort why I’m here today.

Chris Benson: [15:50] So I ran across an article this past week from MIT - and it was actually on news.mit.edu - and they have announced that they are now going to build the new Stephen A. Schwarzman College of Computing. Initially, just hearing that, you may say “Okay, well, that’s what MIT does - they do computing”, but this is a little bit of a different approach to setting up a new college. They received a 350 million dollar foundational gift from Mr. Schwarzman, who happens to be the chairman and CEO of BlackStone. Then on top of that, they have what amounts to a one billion dollar commitment.

What they’re trying to do here with this new college of computing is to build an interdisciplinary school around AI and the various fields that it touches on, to try to kind of drive things into where we’re seeing the future go as we’re living it right now. The AI world has really revolutionized education in that space, and some of the things that MIT notes is that they want to reorient - and I’m reading from the article when I say this - MIT to bring the power of computing and AI to all fields of study at MIT, allow the future of computing and AI to be shaped by the insights from all disciplines. They have 50 new faculty positions to support it, there’s a shared structure with other schools that they’re doing, and they’re really looking at trying to produce students that can operate in this interdisciplinary approach.

Personally, I think that’s fantastic, because in my own experience, AI touches on other fields almost every time you use it. It’s never a standalone thing by itself, it always intersects other areas that you’re using it on. What do you think, Daniel?

Bryan Cantrill: Yeah, well, you should talk to some of those folks, because those folks are in excruciating pain, for a bunch of reasons. First of all, they are all generally cloud-aware. Sometimes people are like “Oh, I love that Oxide – you’re a contrarian cloud play.” It’s like, no, no, we are not anti-cloud. We are very, very pro-cloud. Elastic infrastructure is an incredibly important development, and we are very, very pro Elastic infrastructure, to be clear; very pro cloud computing. We just believe that you shouldn’t have to rent all of it. We believe that you should be able to buy some of it, because when you’re renting compute, the Moore’s Law dividend is going to your service provider, not to you. And we believe that there are folks who got compute at a scale where it makes sense to own a computer, and own a rack of computers.

So the folks that we are targeting are folks that are already on-prem; they are on-prem for good reasons. Those reasons - they may be security, maybe they’re latency… They’re very often economic. AWS is expensive. The public cloud is expensive, especially when you’re at a certain scale. And there’s a certain scale you get to where actually the public cloud is debilitatingly expensive. I can’t even contemplate the cloud of this workload. So those folks - they are cloud-aware, they’re running their own hardware, and they are stuck on this personal computer that hasn’t meaningfully moved in 20 years. And when they buy something from Dell, or from HP, or from Supermicro they then are responsible for putting the software on top of it. It’s like, you’re not running Dell on it’s own, you’re running Dell + VMware; you’re running Dell, plus VMware, plus Cisco, plus software to manage the network. And plus your distributed storage system, whatever that is. And whenever anything goes wrong - well, you assembled this thing, so this is on you. And every vendor points at everyone else. And boy, I lived this; we were Dell customers.

[01:04:19.09] I just remember – I mean, I told Dell a lot of things over the years. One of them is “Never tell me you’ve never heard of this problem before. I’m not interested. All that is telling me is that your most technical customers are leaving Dell”, which is actually what’s happening. And so when you are having a problem, it’s like “Yeah, Dell was telling me I’m the only one seeing this.” It’s like no, Dell was trying to gaslight you into believing you’re the only one seeing this, because Dell itself does not have the depth of competence to actually understand what’s going on, on their systems. Because Dell - and Supermicro is taking it to an even greater extreme, of… There’s this self fulfilling prophecy that this is commoditized garbage, so we’re going to treat it as commoditize garbage. But it’s actually really expensive. I mean, if you look at a racked-out Dell 2U server, it’s expensive. And we haven’t even got the software on there yet. We haven’t gotten VMware, or OpenStack, or whatever you end up putting on this thing to manage it.

For Oxide, not only are we taking the fresh approach to hardware, but also software co-design. We are actually developing the Hypervisor, the control plane. What you get is an actual cloud in that rack. You hit API endpoints, you provision. You don’t buy VMware to put it on this rack.

So for our target demographic, they have been grossly underserved. In fact, the existing infrastructure providers have denied they even exist. They’re like, “Well, every on-prem use case is going to the public cloud.” It’s like, *bleep*. I mean, I know about the public cloud, I understand the public cloud. This workload needs to run on-prem for economic reasons. But this is not a legacy workload; this is not going anywhere. And to be told that you effectively don’t exist, and then be more or less treated that way - it’s pretty aggravating.

So you know, our target demographic is kind of worked up. They’ve kind of had enough. Actually, when we were doing due diligence for the initial VC investment, and having VCs talk to some of our first prospective customers… They’re angry. So one of the feedback – we got some feedback from some of the VCs, like “Your early customers - they’re agitated.” I’m like, “Yeah, they’re pissed. They’re pissed. They’re definitely pissed. And don’t try to tell them that they don’t exist. That’s a mistake.” Because what they have seen is all of this innovation happening in all of these dimensions around them. And all of the innovation around Elastic Compute happening on the public cloud, and where they need to run it, they have been entirely deprived. So what they see in Oxide is oxygen. “Someone who understands what I’m trying –” And yes, I mean, the aesthetics are extraordinary, and the rack really is beautiful, I’ve gotta say. The rack - it is so good-looking; it is a really good-looking rack. But what we’re doing is much deeper than that. It’s much deeper than the aesthetics. It is a true first-principles approach that allows them to actually appreciate some of those advantages that folks have in the public cloud. And then they can go focus their energies on their business, and on building the software and supporting the software, supporting the developers that they need to go build a better product, build a better service.

And what we see is, you know, our target demographic is really started on those excellent enterprises, excellent Dell customers. The customers that we see on the horizon are those folks that are born in the cloud, and then wake up to a business that is actually not sustainable, because their margin is actually going to Amazon.

[01:08:00.27] And over the years, there have been many of these folks who have come up to the edge and like, “We’re going to build our own data centers”, and they’re naive about how bad the market is, so they go in like, “I don’t know, I’m gonna buy some Dell, or whatever.” It’s like, “Oh, God… Yeah, you’re gonna learn just how awful it is.” And they ended up vacillating; you know, going on-prem a little bit, but then staying on the cloud. And we believe that if you give those folks a real off-ramp where they can have the power of cloud computing, but with the economics of an on-prem data center, there’s real demand for that on these cloud-born SaaS companies. So that’s the future that we see.

Johan Vos: That is one of the areas where I think that quantum computing will - not in the very short term, but in the longer term - be really beneficial, and that’s machine learning.

I think we made great progress in machine learning over the past decade or so, but still, if you look at how difficult it is for – I mean, if you ask a computer to calculate the square root of a five-digit number, it will do it immediately, while a human needs more time for this. But if you ask a computer to detect a rabbit from a dog in a picture, you need to feed it with thousands of images before it can really do that… And the algorithms are getting much better, but for example, if you do deep learning, this requires lots of linear algebra and Matrix calculations.

Now, if you can leverage the power of quantum computing, where basically those things can be mainly done in parallel, then there’s gonna be a tremendous speed-up in this. And I think actually that many of the hard problems in AI can only be solved if you also apply quantum computing to really increase the power of the mathematical algorithms that need to be performed in order to do this.

Brian Scott: Well, think about it… Soon, everything can be compute, and you can start literally leasing out your own computing to anyone who needs that time or that power to actually run functions. Now with the whole cryptocurrency coming out, it becomes more of a thing, right? Like, “Hey, let me go ahead and lease this out for some type of actual currency.” I think that’s even gonna drive down the cost of compute very much.

Evan Conrad: They’re definitely described within the app. There’s a help screen that you can go through and check through that… And I definitely recommend to anybody who’s trying it out to do that. But all of these distortions - they’re effectively logical fallacies. One of the problems that I had when I was making these is that people kind of forget which ones are which… But they are a set of things that psychiatrists have determined that these are the ones that most likely cause people’s problems… And it’s often that people don’t realize when they have this sort of logical fallacy of cognitive distortion in their thoughts that’s causing that problem.

One of the things Quirk tries to do is it’s very intentional about the branding of cognitive distortions. Every cognitive distortion has an emoji, and every emoji is the same emoji for wherever you go. On the website, that emoji follows every usage of the term of any cognitive distortion. Catastrophizing is the head-exploding emoji, and on the app it’s the same way; in every header, it’s the same way. It’s there to try and give you a visual understanding of it, and to try and make sure that you remember it… Because that’s one of the biggest problems with cognitive behavioral therapy, at least in Quirk, that people don’t realize this..

[28:04] There’s a lot of terminology and stuff that you have to do to kind of get into this and to understand it… But it’s all this stuff that you have to do, at the same time that you’re going to the app with a serious issue - like, your brain power is already at its maximum capacity, because by definition, you’re using this app at a really difficult time. So part of the app’s job is to get you to the place where you need to be to understand just enough, so that you can record your thought and get to a more mentally-healthy spot… But it has to do that at the same time that it’s competing with other parts of your brain power. So Quirk has a really big focus on being very lightweight… Because if it’s too heavy and tries to do too much, you won’t record anything. It will have the same problem that worksheets have, which is people pull them out and then don’t actually do anything; or they don’t pull them out at all, because it’s very “a thing” that you have to do.

Mikeal Rogers: Well, I also think when you’re comparing – the trade-off of [unintelligible 00:57:30.16] is that it’s a lower file size, but then you have to take some CPU cycles to decompress it. It matters if you have a fast CPU versus a slow disk, or a slow disk versus a fast CPU, or if you just have a fast both, right?

Zac Smith: There was no true need. You played the TCO on power, and speed, and density, but the Moore’s law curve had slowed. We’re just at the beginning of that curve right now in accelerated computing. So I think that’s going to represent an opportunity for a more deliberate model, versus “Here, buy the kit. Call your Dell rep.”