Jeff Smith: To speak to the original business case as we said it, so carving out the military, pro sport team therapists and so on, but looking in that HR tech, it was fascinating, as we’re working with CEOs, CHROs, head of people, chief administrative officer, to really understand their charge inside of these organizations or agencies. And you’re right, the size of the company matters. We were going with enterprise companies, CHROs are saying “Thank God you didn’t come to me with solutions.

I have an entire library of solutions. I just need better insights. How do I deploy them better? Who do I serve [unintelligible 00:42:53.21] development, with perks, with executive coaching and so on?”

So you kind of have that, where they’re saying “We just want better diagnostic tools, better insights to work with.” One said “Hey, I spent last year $250,000 on a single survey that took 90 days to get the results back in nine months and talk about it, for a moment in time that we weren’t in anymore. The company had changed. And so this gives me regular insights that I can really look at.”

And then I can call up that manager, “Hey, this team is looking on the verge of burnout. What’s going on down there?”

Or “Your team’s crushing it. Hey, how can I provide you with more resources?” And they all segment differently. You’re right, it’s regional, it’s by business unit, different teams… So you figure out that segmentation, and then they run with it.

The smaller companies I was finding are saying “Great! Thank you for this information. But what do I do with it? Can you help us better understand what this means inside of our organization?” So what we did was we ended up with our data sciences, behavioral sciences, organizational health, sitting down with these companies saying “Hey, this is what I see.” I mean, already they’re looking at the data. Everything’s coming back, they’re looking at it, great, let me tweak this, let me change this… And before the reporting goes out, because they have a dashboard and [unintelligible 00:44:10.26] on teams… And then at 90 days we sit down with the client and with the behavioral scientist and say “Okay, this team is showing these signs of this. I would provide more fuel on the fire. Rewards. Perks. That’s really – honing on this. This is an area you might want to check on.”

[44:27] And so we’re in that phase right now to understand “Did the large enterprise companies say they had the tools, and just projecting that they had everything, but they really need help, or not?” And we don’t know yet. So I think it’s been fascinating to walk alongside them. We said “Hey, what is our role in remediation, intervention, solution?” Let’s first be the best in the world diagnostic tool, using music as a signal for emotional health and wellbeing. Great. Then, where do we step into the solutions and recommendations? And so we have a full report out, say “Hey, this team - motivation’s up, stress is up, balance is okay. Collaboration’s good. It means they’re crushing it. Let’s provide these.” “It looks like collaboration’s going down, stress is remaining high, balance is a little bit off… Okay, they’re on the verge of burnout.”

So to you understand behavior and what’s truly happening there, and how it influences and impacts companies is important. And so that’s why we have the team look at that, and we’re providing this HR intelligence center to these clients to say “Hey, do you want to tap into this? Can we help?” And I’m finding that they like that. They like being led to water. It’s not a reflection of whether they’re doing a good job or not, it’s just to say “Hey, here’s some resources that you might want to consider”, because we created it for that reason. They do care, and they’re throwing resources at employees… Let’s just help them better deploy those resources so they land. So people stay in their jobs, and companies are growing, and people feel excited to come to work every day.

Quincy Larson: Yeah. So you could lay off one of every five developers, maybe. One of every six, or something. I’m not sure exactly how the math shakes out. But yeah. And that’s basing it on the productivity increase… But we haven’t laid anybody off. We don’t believe that – and we’re using this stuff extensively, internally. There’s just been so much work to be done, we just shift people around. So I can talk about how we’re shifting the team around… But let me talk just a little bit about people who are on the job market. If you’re listening to this, if you’re on the job market, trying to get a job as a developer - the jobs are going to come back. Hang in there. There’s a very slow correction… It takes forever for these hiring cycles to happen. It takes forever for the Gartner hype cycle to ride the wave from the peak of inflated expectations to the trout of disillusionment, to the plateau of productivity. If you’re familiar with the Gartner hype cycle, it’s this phenomenon that pretty much every technology has gone through. And LLMs are going through that. And when people say AI, they’re talking about LLMs. That’s been the major breakthrough.

So I think that if you’re in a situation where you are trying to learn to code and you’re hoping to get a developer job, my advice would be don’t quit your day job. The same thing I’ve always been saying. Like, if you’re working at a Starbucks, or if you’ve got a job at some accounting consultancy… I worked as an accountant for a while; I did temp work, essentially, moving from company to company… Doing that sort of stuff, it sucked. Keep doing that work. Whatever’s paying the bills, whatever’s keeping food on the table, and keep paying down your debt, and doing all that stuff. But just plan long-term. Expect it to take a couple years.

If you look at the pre-2000 bubble, if you had basic HTML, CSS skills, if you knew how to run a web server, FTP some files or something like that, you could get a job as a web designer. And then in 2012-ish, I would say, the tools became so good that a lot of people could get jobs as WordPress developers, or doing basic Ruby on Rails work… That’s what my first job was, doing Rails dev on a small team, just maintaining a Rails codebase… And over time, it’s gotten a little bit harder, but the jobs were increasing, so we didn’t really think too much about it. It’s just “Oh, now we’ve got to learn React. Now we’ve got to learn about a whole lot of security considerations. We’ve got to think about accessibility…” There’s always been this layering of additional things you need to learn, and that’s not going to go away. It’s just going to get harder and harder in terms of the actual skills that you need to know to work as a software engineer.

But at the same time, the number of openings started to fall… So now there’s this bigger gap between what needs to be learned, and the rewards or the likeliness of being able to find a job. And I do think that if you just continue applying for jobs patiently as you continue to build your skills, your network, your reputation, then you will eventually get a developer job, but it may take a little bit longer now than it took in 2022. That is kind of my thinking on that.

[00:33:58.21] So my advice to people who are on the job market is just be patient and keep learning, and don’t give up. Because if you give up, you’re never going to get a developer job. But if you keep at it, you will eventually be able to build out, through your community, through people you know, through different social groups you’re a part of… Hopefully, you’re going out and putting yourself out there and trying to meet people, through building projects and publishing stuff on Twitter, Reddit, wherever you share your stuff… Eventually, you’re going to have a decent portfolio. Maybe you’re going to build a big, impressive app that ties a bunch of stuff together, and is impressive from an engineering standpoint, and you’re going to be able to get a job. It’s just going to be a grind.

Jimmy Miller: But the basic idea for that one, just to be clear, is salespeople – obviously, there’s the real accounting. I know there were some comments on Hacker News of like “That just sounds like fraud.” There’s like the real accounting system, and then there was the rewards system, where they’d get bonuses and stuff… This was the rewards system. And yeah, they were trying to get bonuses every month. And if they made a big sale at the end of the month, and they’d already gotten their max bonus, they would just move that to the next month, so they could get their bonus for next month. So they had a cap on commissions, and things like that, and this was moving it around…

Jerod Santo: The great npm garbage patch

The Phylum Research Team:

Like the island of discarded plastic twice the size of Texas floating in the North Pacific Ocean, npm has accrued an astonishing amount of spam packages over the past six months…

Our 95% confidence interval for the estimate of Tea protocol spam in new packages over the past six months jumped to between 68.66% and 74.67%, or somewhere between 613,000 and 667,000 packages. In other words, among all new packages published to npm in the past six months, about five out of every seven packages are Tea spam.

I first covered the unintended consequence of the Tea Protocol’s crypto rewards back in February (issue #83). It appears the damage is even worse than previously discovered! What a mess…

Joseph Jacks: Yeah, because there’s quite a few clauses in the LLaMA license that they can prosecute people on. But they just relaxed the license terms. I would not be surprised if they just made the whole thing like full Apache, or full some traditional, conventional open source license in the next couple of releases. I would not be surprised. The reason is - and Zuck literally just said this on Bloomberg yesterday. I believe Emily Chang interviewed him. He said they have no need to monetize these models. They have a very strong advertising business, and they’re not selling, they’re not licensing software… This is not a core business-differentiating product, and they can give it away completely. And as a result of giving it away, they reap the same rewards as they did open sourcing PyTorch, and open sourcing Open Compute, which is indirectly saving them - actually, pretty directly - many billions of dollars on building out their compute infrastructure for serving applications that support billions of people, which is what they do.

And so it’s actually quite a beautiful, very aligned company to be embracing and investing so much in open source and open platforms. And you kind of think, like, “Why isn’t Google doing the same thing?” They also print money with advertising. Why isn’t X doing the same thing? I mean, Elon - maybe he doesn’t quite as much print money with advertising; he’s lost a lot of advertisers. But you know, the point still stands. If your business model is not in diametric opposition to your innovation model, you should be innovating like bananas. Like crazy. And your business model should be reasonably sound, and protected, and you have different aspects that compound the value of an advertising business, which is collecting and generating more user data, which is what they do… And unfortunately, that’s the trade. We use these free tools for free, we don’t have to pay for them, and we happen to part ways with our very precious personal data. That kind of sucks. I’m actually very anti that. I deleted my Facebook account many years ago. But it is a very aligned kind of company in terms of acknowledging “Look, we can invent and create crazy powerful technology, give it away, no problem, because our business model doesn’t need us to charge money for that technology, even if it’s costing us billions or tens of billions of dollars to produce”, because we get rewards and benefits elsewhere, that show up elsewhere on the P&L… Which is like compute efficiency, improvement, talent acquisition, whatever.

So I think Meta is a really great example of – I think, frankly, a pretty incredible steward of this open source kind of philosophy in the last few years. I mean, it’s been pretty surprising, sort of rebranding almost from the previous years of lots of drama and blow-ups that Mark Zuckerberg has had to deal with.

Birk Jernström: So that’s certainly a helpful tactic, but you don’t have to deploy it. And I think this goes back, again, to – it’s just a key principle of ours… Like, this is not issue bounties for open source. I’m very much against them. We’ve designed issue funding and rewards with maintainers in mind, and putting them in the driver’s seat. So some maintainers, they don’t want a funding goal, because they don’t want to necessarily say that “By X amount I promise to do it.” They want to leave it open-ended, and so we give them that complete control and flexibility.

Jessica Sachs: So the one FAQ on the Kickstarter that makes me curious of the structure of the company is “What if I’m already a Font Awesome pro subscriber?” That FAQ. And it says “If you’re already a Font Awesome Pro subscriber, and if your rewards include a year of Font Awesome Pro, your subscription can be extended by a year.” Does that mean that the whole model of like every company that’s currently using Font Awesome is going to be affected by this awesome – oh God, I used “awesome” as the adjective…

Robert Ross: You know, make a rewards system for our most prolific incident commanders.

Karan Malhotra: I think that definitely just like straight up instruction tuning is great. There’s other ways to tune, like the eval instruct method. I would advise people to try to create new instruction methodologies, that allow us to make even better formatted data. People don’t spend enough time trying to create new instruct formats. And we’ve definitely been swamped with not doing that as well. So I think towards the general community, it’s a really easy place to get started; you don’t need to really know how to code, so much as think about how a human might more effectively phrase something, or format something, and kind of remix from there. I think that’s probably the easiest place to start.

Then there’s model merging. Model merging is great. You can just like take two models and Frankenstein them together to question-mark results. You’ve got to just try and see what happens, and feel it out. Then from there, I would say there’s stuff like DPO, there’s RLHF… DPO kind of rewards things; that can let you enable rejections, or create censorship, or put some kind of general concept or attitude towards a model. We’ve found that to be pretty effective with the latest News Hermes Mistral DPO. It seems like people really like it and prefer it over just SFT. So that’s another thing that I’d heavily recommend.

From there, we get a little more complex. We have some reward model stuff we’re working on that I won’t speak to just yet, outside of saying we’re working on it, that we think is going to be like pretty big for reasoning boosts. Of course, there’s techniques like chain of thought, and tree of thought, for like multi-step prompting. Creating datasets even out of that for any of these purposes I’ve already mentioned is going to be really effective.

Now, to stuff that maybe not everybody can – actually, a lot of people would already be able to do this. There’s something that we like to call over at Nous activations hacking, where you’re kind of messing with the way that a model – I’m trying to think about how to say this in like the most layman’s terms… You’re trying to mess with how a model generally vibes about something. [laughs] So rather than just doing a system prompt or something like that, you can actually change the model vectors to kind of be like more political about something, less political about something, more terse, or more specific… And it has far more effect and control over a model than a system prompt. It’s basically like a system prompt that tells it to embody certain characteristics, but it’s not something you can really jailbreak or get around, as far as my testing has shown. Certainly not as easily as a system prompt. We have no problem jailbreaking even the most censored closed models today. It can be done by anybody, with the right words. But this activation stuff, it really creates a bit more of a robustness and fidelity to the concepts that you’re trying to tell it to embody.

There’s a few more I’m trying to think of that would be useful for people… One thing is soft prompting. It’s not really around anymore. It used to be pretty big during the GPT-J, like pre-LLaMA days, and the Cobalt AI guys really pioneered the use of it in the open source community. But a soft prompt basically takes like a massive prompt and it compresses it down to like way less tokens. So you can give your model like a huge prompt, a huge system prompt, or a huge amount of information, and use like way less tokens. So soft prompting is cool. It’s not going to be too difficult to update it for like LLaMA, Mistral, today’s architectures. Nobody’s really done it that I’ve seen. So to the community, if you guys do that, please share. [laughs] That’s actually much easier than the activation stuff, I think.

And then finally, probably the hardest unsolved is sampling methods. Today we use like Top-K, Top-P, [unintelligible 00:39:52.10] sampling etc, whatever. There’s better ways to pick tokens, for sure. There’s better ways to judge the value of tokens, for sure. Everyone has been too concerned with higher levels to get that low, and do whatever the magic math is that I can’t do, that would enable some steering, and some – even beyond steering, like alternative sampling paradigms. And I think that would probably bring the biggest change and transformation to literally all models, regardless of the tune, regardless of the architecture etc. if it gets pulled off. So I’m really looking forward to something like that happening in the space.

Adam Stacoviak: [01:15:49.00] The most innovative content happening out there, in my opinion, the proving ground is happening in virality, as well as consumability on TikTok. Like, there is a lot of remix… I just think it’s a proving ground for a lot of great things. It’s not the only ground, but I think there is a lot of people on that platform, and it’s rewarding to those who can find ways to present certain things in a way that people want to consume them. I think the algorithm rewards you well as well. I think a hub and spoke, multifaceted approach is the right way; maybe long-form or clips on YouTube, maybe very personal things on TikTok… I mean, a newsletter is still very great… I think you’ll see some dividends be paid to you, if you – even in this case here, you’re exposing your idea to a larger audience. Like, do this again, but on a whole different podcast.

Jerod Santo: Sorry, digital rewards just don’t get us.

Haroon Meer: [01:34:17.27] Yeah. And again, I’m pretty convinced it’s also the path to winning. I’m pretty convinced… And again, we’ve been lucky so far, but I think the market rewards that stuff. You end up making a good product, and the market hopefully rewards it. [unintelligible 01:34:35.01] we just did the Black Hat conference, where we had a booth… And one of the things I was talking to someone about which was super-interesting is when we do a booth - like, we’ve got this really long blog post out on doing booths and why we think it’s actually good for people. Young me hates it, but booths are super-good for us. We do a booth at RSA, we do a booth at Black Hat, we get to meet all our customers, they come by and chat to us… People we’ve never seen before come and say nice things about our products, other people hear them… But at Black Hat this year something that occurred to me is we have this booth, and I’m there for the full two days… And Marco who’s our CTO is there, and Bradley, who’s one of our other founders… So literally, all of our original founders are there, plus some of our engineers. And so for two days, people are rolling up to us, talking to us, “Hey, I’ve been using you for six years. Hey, I did this”, but that’s surprisingly rare on the showroom floor. Because on the showroom floor, what lots of people have done is they’ve paid a whole bunch of young interns, a whole bunch of college students to say “Scan as many badges as you can. You scan the badge, you then get to spam all these people, trying to sell stuff”, and again, it’s horribly mixed incentives.

For us, the thing is we get to meet our customers, and we get to do demos with new people who might be interested in the product. And it’s so counterintuitive, because if you talk to – with any VC, one of the playbooks that they will tell you is the truth is not in your building. Go out and meet customers. If I told you “You’re on the showroom floor, and you’re going to meet 20,000 of your customers in two days”, why wouldn’t that advice mean that every CEO, CTO and chief product officer is the person on that floor? You’re gonna meet 20,000 of your customers or potential customers; you can talk to them about the product. But it’s just not done. Because fundamentally, what the execs are doing is they’re sitting in a suite somewhere, trying to arrange their next raise, or trying to talk to analysts, or trying to talk to the media. And again, the state of the products in our industry are a reflection of that; mostly, we build terrible products, because people just don’t care enough about them.

Cory Doctorow: So that’s chickenization. And if chickenization sounds familiar, it’s what Uber drivers have, right? It’s what DoorDash has. It’s what – there’s a company called Arise. They hire primarily black women to do work from home call center stuff, and each of these people are misclassified as contractors. They have to start an LLC. And then because they’re a company, they have to bid on the job. They work for like Carnival cruises, and Disney, and whatever - they have to pay to be trained to be the phone operator. And then they’re listened in on constantly, and if their kids cry in the background, they are struck off; they lose the job, but they still owe the money for the training. And there’s a penalty for cancellation on their side, where if they quit their jobs, they have to pay their boss to quit their jobs. So that’s a form of chickenization as well.

Chickenization is kind of working its way up the privilege gradient from the most blue collar workers now to a more middle-class group of workers. It’s your boss’s ideal, right? All the risk is pushed on to you, all the rewards go to them. Every business owner would love to have that as their arrangement, right?

[24:18] So what’s a centaur? Well, a centaur in AI research is someone who’s machine-assisted, right? I’m a chess grandmaster, you’ve got a chess playing software program. Neither of us are as good on our own as we are together. We can beat opponents together that we couldn’t beat on our own. So that’s a centaur. But a reverse center is when the machine is in charge, and you’re at support. So say you’re an Amazon warehouse worker who’s wearing haptic gloves that go “Bzz! You need to pick that up. Bzz! You need to turn your body. Bzz! You need to drop this thing in the box. Bzz! Bzz! Bzz!”

Or say you’re an Amazon driver, right? You’ve got a facial recognition camera that’s watching your eye movements and penalizing you if you look away from the road, you’ve got a clock that’s timing your drives, that’s telling you you have to turn left, even if turning left would result in your death, and then it docks you for doing it. Machine says no - that’s a reverse centaur, right? It’s when you are the disposable meat sack for the all-important algorithm.

Fred K. Schott: I would call out – I think the interesting part of React and Next sort of like merging a little bit right now is Next.js I don’t think is… That’s a strong project, that people are very happy with, and that’s starting to become – it’s gone from like “Wow, they’re doing a great job” to like “This is fairly abnormal.” I think it speaks to the fact that React is so big, and there’s not really a competitor to Next. You can think of SvelteKit and Nuxt on the Vue side, but everyone’s so framework-specific… No one really – Remix is I think the first time I’ve seen a second framework for an ecosystem come along and really challenge it… So I really wonder how much of that is like there just aren’t really other alternatives in the React ecosystem, and the fact that React is so big, it just kind of like by default is like the top of that food pyramid… But at the same time, it also speaks to just people are happy; like, they’re not really looking around for other things… Maybe because they don’t have other things in front of them, but also it’d be very different if it’s like “I use it, but I hate it.” If interest was low, there’d be a real risk there. But I think it speaks to the strength of that team continuing to innovate. There are huge risks to continuing to change things at this late stage, but there are big rewards to staying top of mind, and keeping your number one spot as the most popular – you know, you’re not the newcomer anymore.

Daniel Whitenack: [26:08] In a reinforcement learning loop you would have a kind of policy, which outputs like a “what you should do next” sort of thing. And then you have some type of reward system that rewards the agent for acting according to the policy or not. So in this case, the reward model is outputting that reward or that preference, and the language model is actually acting as the policy here.

So you have an original language model that is kind of your original policy, and isn’t fine-tuned yet, according to human feedback. Then you gather some human feedback, like actual human feedback, train a reward model to simulate that human feedback, and then you fine-tune a copy of your original language model or a copy of your policy with this reward model.

And so the pre-trained language model - it could be any language model, it doesn’t have to be GPT-3, but in the case of OpenAI, it was GPT-3. But you have an original language model, and that language model could just be a general pre-trained language model, or you could additionally fine-tune that model, maybe for a domain or a specific type of output you want. So that’s your pre-trained language model.

And then step two, to get the reward model, what you do is you start outputting data from your original policy, from your original language model, and you have humans rate it. Maybe you combine that with certain human output, or certain other outputs, and you have human rating. So that way you’re creating a training set for your reward model, which includes human labels of their preference. Also in this step two you then train a reward model using that data that you’ve gathered from humans to output the preference.

Now, to your point of “How does this scale?” Well, the fine-tuning of the policy is done kind of with this automated reinforcement learning loop… But you do need humans to generate enough data to train your reward model that’s used in that loop. And what’s interesting is – and the Hugging Face blog makes this point, is that different people or different groups that have applied this reinforcement learning from human feedback have used different-sized reward models. And obviously, as the size of your reward model increases, you need more data to train it. That would be a general rule. In the case of OpenAI, their main language model was like 175 billion parameters, and the reward model was much, much smaller. 6 billion parameters. In other cases, people have done similarly-sized models. And so I think that is an open question - how should these models size-wise be related to one another? What types of models should you use for your reward model, and how much human feedback do you need? To be honest, I think those are open research questions.

Nick Nisi: So tech lead is a really interesting thing, because different companies treat it differently. And even sometimes different teams within a company treat it differently. So some people treat it as a durable thing. Tech lead is a title, it’s a role, this is a thing you’re doing, that might come with rewards financially, or otherwise, or it may not.

Charles Martin: Look, what are you trying to do? Like, what does it mean to be optimal? If being optimal means that your alphas are close to two or three, then you should adjust your hyperparameters such that the alphas go down. That’s what you do. Now, doing that analytically, typically, doing what are called analytic derivatives, meaning you try to compute the gradient from that - that’s somewhat difficult (it could be done), because you have to compute the eigenvalue spectrum, and then you have to fit it, and then you have to figure out the derivative. And that’s a very complex, nonlinear calculation. It’s very iterative. It could be done numerically, or it could be done analytically with some work. It’s a lot of work… I would love to have VC funding like Hugging Face to do that… But I don’t. It’s just me. Me and you. So you just try to tune your parameters; if alpha goes up, go the other way. If you turn your learning rate up, and you find your alphas are going up, turn the learning rate the other way, and hopefully they’ll go down.

[34:06] Obviously, it’s a complex optimization problem, because you have 100 layers, you have 100 alphas, and so you’re trying to tune different layers, you’re trying to tune your layer learning rates, and your amount of dropout, and the amount of momentum… So in principle, you could try to do that algorithmically, in a way, using like a Bayesian type approach, where you try to get your alphas to go down on every layer. In principle, you could do that. It’s a complex optimization problem, but that’s what I would recommend. And I think it’s theoretically well-grounded. I mean, the point is that you want to learn more correlations.

Typically, what I’ve found is that it’s a good tool for newbies, because you get into a model, you start doing something, and things are totally wrong. And you can go in and fix some problems. Okay, now we’ve fixed it; we’ve found – like, what did we not do? Like, I didn’t put the proper regularization on these layers. Let me add regularization and try again. And you can see that, okay, that’s much better.

So from a newbie perspective, it’s a very good tool, because it helps you get started. Now, it does work – keep in mind, the tool works at the end of training, not in the early stages of training. You’ve got to let the thing bake for a while. And once it’s about halfway through training, then you can start looking at things… It’s got to have some correlations. But this is what it’s for.

Typically, trying to do large-scale meta learning would just mean you’d have to integrate the tool into some sort of process that allows you to look at the alphas, or look at more details in the layer - the shape of the spectral density, the number of spikes, the alphas, the volume of the spectral density - and figure out how to tune from that. I mean, this could even be used in a reinforcement learning situation, where instead of the reward being something that the agent takes, the reward is Oh I got smaller alpha, So I have rewards on every layer, and I sum the rewards in some average way to try to get the optimizer to work, even in situations where I don’t know what the reward is for a reinforcement learning situation.

Obviously, that would be nice in areas like you’re trying to trade in the markets, because you can’t take actions – that trade, you can’t trade on historical data and expect to learn from that. So this gives you a way of sort of doing things in a supervised or semi-supervised way, that doesn’t require peeking at the test data to optimize. And then that’s it… I hope that answers the question, but that’s sort of the idea. And there are lots of things people I think want to try. I think it’s great if you try them.

Adam Stacoviak: You’ve got Dropbox versus Box… You’re going to have competition. Somebody’s going to come into the space, be very similar or very same. And I think the market rewards those who are not straight-up copiers, but there’s like this, the best artist steal kind of aspect. Obviously, it’s a lucrative space. There’s going to be a lot of things happening here, a lot innovation. I’m just curious about the relationship, how it works out.

Jacob Kaplan-Moss: So I think if it’s going to be longer than eight hours total, I think compensation is appropriate. I wouldn’t consider asking for more than that without compensation. However, I’m starting to see a couple of positions that are offering compensation for anyone taking a work sample test, or sometimes finalists, or something like that as a way of indicating respect for their time and indicating some company values. The example that I can think of most clearly is the Software Freedom Law Center, which is a nonprofit involved in open source and free software law. It was conducting a search, I believe, for an executive director, and they compensated everyone who they selected to go to the interview stage. I believe they gave them $500 each.

This is an organization that is keenly aware of the difference between paid and volunteer labor. They’re in the open source world, where this is like a massive area of conversation right now… And I think - putting words in their mouth, to some degree; I don’t know why they did this, but I think this was a way of living their values of, “This work is important. The way we represent that in our society is we give people money for it.”

I would like to experiment with compensating everyone who gets to a work sample test phase. It’s not going to be entirely my decision, because I’m part of a company, and I can totally understand. I think reasonable people will differ on this. I don’t think it’s unreasonable to say, yeah, both candidates and companies should invest some time. And the norm is that interviews are not compensated. And I don’t know that it’s unfair, but I would like to try and see what impact that has on the candidate pool. And if that moves the candidate pool in a positive direction, I’m inclined to say that the cost would be minimal compared to potential rewards there.

Thomas Simonini: So to say it simply, deep reinforcement learning is to learn from action. So the idea is that you have an agent that will learn by interacting with its environment. If we take a simple example, you can make an agent that will learn to play SuperMario Bros, so to play videos. And how it does that, it just tries some action in the environment, so going left or right, and based on the rewards it gets - so for instance if he dies, it’s negative - it will improve its behavior to win the game.

Matthew Ahrens: [24:12] Yeah. And I think most people do. If you care about this data, then you need to be able to have some redundancy with it. So you need the functionality of a volume manager, where you can have multiple drives, and some of them can die, and you don’t lose all your data. And you need to know that the data that the drives give you is correct, so you need checksums. And even just those two basic requirements, there’s not– I mean, there are other technologies that do that. They are much harder to use, typically. And so I think the choices are - if you’re deploying it yourself, ZFS is just so much easier. If you’re making a product, then your customers might not know or care, but building your product on something that’s as capable as ZFS is going to long-term rewards. And the fact that ZFS is under continuing development and improvement, it means that your product has a solid foundation that’s going to keep up with what the future holds for software and hardware storage. That’s just the very basics, right? I think a lot of people even forget about that, because there’s all these other cool, great features of the ZFS that are very exciting, like snapshots. Being able to protect your data with snapshots.

A lot of people nowadays are thinking about ransomware, and “What if somebody has some virus or whatever they call it that encrypts my data or alters my data or deletes it? How do I recover that?” Well, ZFS has built-in snapshots, it takes snapshots every day, every hour. The storage cost of them is very low. You’re only paying for the data that’s different, each snapshot, and the performance impact is basically nonexistent. So you get a lot of protection from accidental or malicious changes to your data very easily, and very low cost in terms of the hardware that you have to pay for it.

Things like compression built-in - I think a lot of people nowadays take this for granted. At least ZFS users probably take it for granted. But it’s not present in all the competing technologies, being able to just turn on compression. We’re using LZ4 compression by default, which is very, very fast. It doesn’t give you the highest compression of ratios, but it means that you can just turn it on and not worry about it. You turn it on and typically performance improves, because you don’t need to read and write as much data to your disk.

And then in more complicated deployments, people look for things like replication. I have data that’s on this machine, I want to get it to this other machine. I could use rsync, and that would probably work just fine if I just need to take one copy of the data one time. But if I need to continually move the changes over, rsync is very, very slow, because it needs to check every file and every block of every file to see if it needs to be sent. And then if you’re using ZFS to begin with, obviously you want to preserve things like the snapshots that you have on the source system, the compression that you have on the source system. So ZFS has these built-in send and receive commands that let you serialize the contents of a snapshot, send it over to another machine, and it preserves all the complicated stuff like ACLs, access control lists. That might be sound files and other esoteric things that might not be used that often, but when they are used, you don’t want to have to worry about “Is rsync preserving them correctly?” or whatever.

Daniel Whitenack: Well, one cool thing is I think we do have scheduled to have Thomas from Hugging Face, who created this, in an upcoming episode, so we’ll dive into it more then… But I wanted to mention it here, because I think it’s cool that Hugging Face made this transition; I mean, at first, it really centered around chat interfaces, and then more broadly open source NLP, and then into open source and general-purpose AI tooling and hosting services. And now we see this reinforcement learning piece coming in, where the goal is that within Hugging Face itself, you’ll be able to build and share reinforcement learning environments.

Now, Chris, I know that we’ve talked about it a little bit in the past. Reinforcement learning, in general, is not a model. It’s a framework in which you can train agents or models. And if you think of a self-driving car or something like that, if you say, “I’m going to go from point A to point B in a self-driving car”, well, there’s a whole bunch of routes you could take, right? And it’s not really that there’s one perfect solution to that problem; it’s more about the decisions you make along your route, based on what you’ve done so far and the feedback that you’re getting from your environment. And so in order to train an agent to execute decisions in that environment or make actions in that environment, you need to have a simulated environment, that will allow an agent to navigate in that environment, get feedback and rewards, and then be trained accordingly to all operate in that environment, in this case, to win a snowball fight. So although the game is interesting, that’s ultimately the most interesting thing about this, is that it provides a route towards training reinforcement learning agents in environments that are shared on Hugging Face.

Ashley Willis: For me, I tend to project my motivations onto people. Say that somebody does not necessarily love rewards, they’re not necessarily motivated by promotion, or even more money. Maybe they’re actually more motivated by public recognition, and I’m like, “What do you mean, you don’t want more money?”

So I think sometimes I have a hard time understanding that we all have different motivations. To fix this, I have sent out surveys to all of my people, asking them various questions, like “How are you motivated? Do you like public recognition? What was the last time you were promoted? Is that even important to you? How do you like to receive feedback?” So these important questions that normally I think - at least in my experience - people don’t ask, I think that that is a common pitfall for managers to not really dig into how people are motivated, how they wanna receive feedback. Sometimes people want direct, sometimes people want you to be a bit softer… I think knowing those things are important.

Gergely Orosz: It’s so hard… And companies do wanna fix it. Google wants to fix it, for example. The interesting thing - if Google cannot fix it, is it really fixable? Because I think you have two forces that go against one another. One of them is the best companies know that if you don’t give people internal mobility, they’re gonna quit. So if it’s easier to change your job, you’re gonna leave. So a lot of the best companies - Google, Facebook - and I’m talking work culture; let’s not get into – like, some people have thoughts about these, Stripe as well for example, they wanna have a great working environment, so they empower people. They want people to make an impact, and they make really easy internal mobility.

Now, the downside of this is that – and they promote based on impact. Every company says “If you’re impactful, we’re gonna give you a fat bonus, and you’re gonna go ahead, because you’re helping the business.”

Now, the incentive that this creates is people do go chasing impact; the smarter people will chase easier impact, the stuff that in a year it shows… And as soon as they see that - there’s a lot of work that would take blood, sweat and tears, it will take multiple years, let’s say 3-4 years to get out there… Let’s say for example the Kindle at Amazon. It’s a project that if you kicked it off and the sales were initially low – it took years to pick up. If there was not a founder there, Jeff Bezos, who actually invested in this and said “We’re gonna do it regardless of the impact”, I don’t think it might have happened.

So to summarize your question - it is fixable by having a founder or someone with a very long-term incentive; ten years incentives. A CEO who does not give a 4-year grant but a 10-year grant. For example, like Tim Cook did at Apple. He had a 10-year grant and he made tons of money. I don’t know what other companies give 10-year grants… And have these long-term incentivized leaders kind of give air cover, rewards for things that might not be impactful in the short-term. I think from my observation this is missing at Google, for example. It’s missing at most tech companies as well. The ones that are doing well, there are one or more founders.

I’ll give you an example… At DoorDash, for example - they have multiple founders there. Their evaluation just surpassed Uber. I don’t really know exactly how or why, but I can tell you that the people I talk with, they felt a little bit more connected to the company and to the founders than, say, Uber, which does not have a founder anymore. The founder is not at the company. So it’s a different dynamic, and it’s a hard one.

Amal Hussein: [27:50] Yeah, on the flight, right before the hallway; you’re still editing… So you don’t get to riff on your talk in the way that you would normally. You don’t get to really connect with the audience and you don’t get the energy boost. So all of the rewards of doing a talk aren’t there as much, so you just have more work as a speaker, and then you don’t have the feedback reward as much.

The conferences I’ve seen go off really well are actually the ones on Twitch. I’ve participated – yeah, Progress DevReach was on Twitch, and I did a session with them, and I thought that was so much better than pre-recording, or this or that… Because it’s a format that just lends itself to people chiming in and having a lot of chatter, it’s all live… It’s as close as you can get to an in-person conference.

But this kind of like “I’m gonna play an 8-hour video track and hope that people around the world are tuning in” - it’s just less appealing for me personally. I’m just at the point where for me conferences are more about the hallway track, and content from conferences is a thing that I do at home. I’ll watch a talk at home, alone, where I’m able to digest it and take notes…