Meta’s open-source Llama 3 model puts Meta’s AI assistant head-to-head with ChatGPT.
Stability AI laid off 10% of its workforce, the first major AI foundation model to reduce its workforce since the advent of generative AI.
Is the dot-com bubble a cautionary tale for AI enthusiasts? Listen to the episode of Marketplace from NPR.
Do LLMs support Wittgenstein’s position that “meaning is use”? That depends whether you’re talking pre or post Tractatus, of course.
TikTok wouldn’t lie to you: you really can make friends with your local crows. Eira’s proven method for befriending crows: Feed them dry cat food on a consistent schedule, so they learn you’re reliable, and watch them eat, so they grok that you are feeding them intentionally.
Stack Overflow user Arman Ordookhani received a well-deserved Lifeboat badge for telling nearly 30,000 people How to free memory in go.
TRANSCRIPT
[intro music plays]
Ryan Donovan If you're building an app, listen up. With Nylas, integrating with email and calendars has never been easier. Discover how 250,000 developers launch 40 times faster without all the headache. Start building for free at nylas.cc/stackoverflow to start building for free today.
Ben Popper Hello, everybody. Welcome back to the Stack Overflow Podcast, nerd gamer edition. I am Ben Popper, Director of Content here at Stack Overflow, winner of a local in-store gaming tournament– actually, I took second place. But nevertheless, I'm here on this podcast to talk about software technology, and I'm joined by my daring duo from the content team, Ryan Donovan and Eira May. What's going on, y'all?
RD Not much. How’re you doing, Ben?
BP Well, like I said, I won a Magic the Gathering tournament and I received a small pin, so I'm feeling on top of the world.
RD You're not wearing the pin today?
BP I can find my cactus pin. I can hook it up.
Eira May Yeah, we need a picture.
BP Speak amongst yourselves.
RD Pins or it didn't happen.
EM Pins or it didn't happen, that’s pretty good.
BP All right, shots fired. I'll find it. But we've been soliciting listener input and we've gotten about two or three so far. It's exciting. We're going to have one listener onto the show, and today, I would love to talk a little bit about a question which was raised by a listener– “I want to hear more about the software job market.” And so I did a quick Google search prior to the show, and the number one article that comes up is from January. And in January, the sentiment was clear– there is a slump in the job market, software engineers are finding it's much harder to get a job than it used to be. January 2024, not so good. We're now in April of 2024, and if I search for news articles from the last month, things look a little bit different. There's a lot of articles about what is the new competitive landscape. There's a lot of articles about how you get the skills you need to become an AI engineer because that job is in demand. And there's stories about areas like agribusiness and cybersecurity where the hiring has continued unabated. So I would love to hear your thoughts. I can dig into some of the details of these articles along with you, but what have you been hearing or seeing? Talk to me while I find my pin.
RD Just from seeing my LinkedIn feed, I still see a lot of people looking for new roles. People are getting new roles, so there is hiring going on, but it does feel like there is a little bit of a slump, a little bit of gun shyness about what's going to shake out. Eira, seen anything?
EM I think I've been hearing the same thing. It seems like there's not quite so much industry-wide angst it feels like as there was when we had sort of the waves of layoffs last year, but still a little bit of a sense of instability. Speaking of stability, one thing I wanted to mention to y'all is just that Stability AI has laid off about 10% of its workforce, which is about 20 people. So not a hugely unique event to be happening in tech, but interesting because this is really the first major AI foundation that's really had to cut its workforce since generative AI became the kind of tip of the tongue topic for everybody. So whether this is part of a bigger market correction around AI hype versus AI reality, or whether this is just that this particular org has had a rocky few months, I don't know. But that's just something that was on my mind today because I don't think we're used to seeing the layoffs and AI firms in the same headline yet.
BP Right, it is really interesting. I think that one of the things to think about here is that there was a period of time where the big players were not really doing anything too public or commercial in this space. They had big research arms, but they weren't commercializing it. And when ChatGPT came out, everyone said, “Oh, boy.” This caught on with consumers, it got to 100 million users, fastest growing app. It just became that that was the media throughline– this thing is a rocket ship. And so then Stability AI, which has had Stable Diffusion for a long time, and that has been one of the cutting edge models, I think was seen as, “Okay, this is going to be another one of these upstart challengers that can dominate the space.” But OpenAI did something very different, which is they cozied up right away to Microsoft, which took 49% of the company and provided them with a ton of compute and I think probably helped to establish them in the eyes of potential enterprise customers as very reliable, whereas the CEO who left here is going to pursue decentralized AI and was not interested in that increasingly corporate version, but was interested in the Wild West open source decentralized version. And I think it's hard for startups to compete because this business is extremely capital intensive. You have to buy a ton of hardware, you have to pay for all the compute and electricity, you have to fight over great staffers who are getting million dollar salaries, and so tough for a startup unless you pick a big corporate partner.
EM I think that has been kind of the takeaway so far.
BP I think the reality that AI would come along and replace software engineers has not materialized at all. And when you look at the high profile examples of where AI has replaced jobs, I can think of two. There's Klarna, where basically they replaced the very first layer of what was already outsourced customer service. This does not seem like a revelatory thing. In 2015/2016 people were pitching chatbots that would replace call center workers. Maybe the latest large language models are better at it, but it's not like this was a job considered to be at the top of the white collar food chain. And then the other one that Ryan and I discussed was seasonal hires of graders for these exams in Texas, where again, you're hiring 4,000 people for three months to churn through a million high school essays. It's pretty brutal work and the essays are not at the highest level, and maybe this is something an LLM should be doing with the right amount of human oversight or whatever.
EM Having worked as a writing tutor and an English language instructor in college and graduate school, that's an area that I'm pretty comfortable outsourcing to AI just out of a wish to not have anyone's life suffered.
BP I think if AI can write the SEO content and grade the student essays and be the first line of defense on a call center, that's a boost to global happiness in all directions, in all directions. That's a win.
RD There was a lot of period with AI and AI companies sort of excited doing new stuff and then a lot of difficulty finding market fits and a lot of people sort of being like, “What's the business case for all this stuff?” And Stability seemed like it was always sort of like, “Let's make cool things and put them out for free,” so maybe they are reconsidering trying to find market fit.
BP I think it also depends on how much money you raise and then how much pressure you get from investors. If you raise a ton of money then suddenly there's people breathing down your throat, “What does your business model look like and how's it working?” Midjourney is very different, for example. I think they have like 10 employees total. They've been charging a fee to users for a while that hopefully is more than break even for the cost of generating the images and they continue to do it in a community-driven fashion. So I guess Midjourney will be an interesting sort of counterpoint to Stable Diffusion. Those two were very similar in my mind, emerged sort of at the same time with these breakthrough text to image offerings, and so interesting to see what happens.
EM For sure.
BP I'm going to add one more data point here. This is from Marketplace, which is part of National Public Radio, which I consider to be decently objective and they at least do their research news organization. Once upon a time, working in tech was the place to be. Salaries were high, demand was booming, careers felt secure. Now the story is changing. The tech sector added 260,000 jobs in 2022. Last year, 2023, it added 700 and that number is expected to shrink even more in 2024. I don't know if I really expect it to shrink. I think that might be inaccurate. So I think we'll have to see if 2023 was an anomaly or not. But obviously going from adding 200,000+ jobs to adding just 700 feels like a huge shift for people who've worked in the industry. And there was a boom from 2010 through 2020, pandemic was crazy, 2022 most hiring we've ever seen, and then a brutal correction. So we'll see how it goes from here. I guess one other thing that I would say is that two of my friends who were laid off or left work because they felt like it was no longer tenable are now doing contract work. And I don't know if that's a bigger trend, but contract work that is basically full time, like 30 hours a week, not benefits, but seeing if the company is right for them or still playing the market, stuff like that.
RD Maybe this is all a sort of general part of the correction after now quarantine and the pandemic where everybody thought they were just going to be remote and everything was going to be cloud forever and people hired up a lot, and now that correction has been happening for a while. AI has been a different part of that correction. Hopefully we're getting to a more stable part of the job market, but we'll see if there's another bubble bursting.
BP And Eira, to get back to the discussion of Stability for a second, the story about them doing layoffs was from today. Timestamping this April 18th, 2024. So the top story on Hacker News is, “Facebook announces Meta LLaMa 3, the most capable open available LLM.” And so for free, you can use completely state-of-the-art, cutting-edge stuff done by Facebook. It's already going to be built into AWS, Google Cloud, Hugging Face, Kaggle, IBM, Microsoft, NVIDIA, Snowflake, blah, blah, blah, and they're adding tools to it– CodeShield, cybersecurity eval. So Facebook has given away top of the line stuff for free and so that's a tough market to compete in. Stable Diffusion, in some ways its competitive edge, Ryan, was like, “There are no restrictions here. We're not going to tell you what kind of images you can create or what kind of prompts you can write. This is open source software. You do with it what you will.” But I guess maybe they didn't build a business model around it or the CEO didn't, I don't know. We don't know why a CEO left. We can never really say.
EM Sure.
RD Building a cool open source model or software, as we've talked about before, there's a funding problem in open source software. Everybody loves using it, nobody likes paying for it or maintaining it. So the Stability folks, like you said, we don't know the business model. We don't know if they even wanted a business model. Maybe they just wanted to do the research.
BP Exactly. If that was the case, then they should not have raised money. That's the open AI story. They just wanted to do research and then they raised the money and they said, “Oops, now we have to be a capped for-profit. Excuse us while we figure out the profit thing.”
EM Probably something that you think would have come up during one of the funding rounds, but what do I know.
BP They were an eight year old company when they broke out. For the first five years, it didn't come up. It was just like, “We're just chugging away. We're not doing anything that's really making news and a few kooky billionaires support our mission of creating AGI for the betterment of the world.” They got pretty deep before it became an issue.
RD And ChatGPT came out and it was a cool toy. And unlike Midjourney, they decided to move past the toy stage.
BP Cool toy, sir. Sir, a cool toy?
EM Fighting words, fighting words.
BP Stability AI raised 126 million over two rounds. That's not a ton of money, but it's a chunk of change. October 2022, Stability AI has raised 101 million. The round was oversubscribed, everybody wanted in, and now they're saying, “We wanted in. Where's your business at?” Investors are not patient, they're not patient. But as always, when a new big model is released from one of the big companies, Meta 3 outperforms the existing open source alternatives, Google's Gemma and Mistral, on the various MMLU and human eval and math blah, blah, blahs. I don't really put too much stock in these anymore because the tests have been out there for a while and I don't think that they translate necessarily to user success and value or to business application and value. But I do think it's clear that everybody's making progress, and so every three or six months when you release the latest greatest from a cutting edge shop, it's going to boost on these evals and that does add up to something. I think I mentioned before, my AI skeptic friend was finally won over by Claude Opus, Claude 3, and just sort of its facility with language and reasoning.
EM What was it that really tipped the balance? Just that he felt like there was more facility there than there had been with previous models or what?
BP He felt like the critiques it gave of essays and law briefs were good enough to be useful. I'll show it a piece of writing, it'll give me a critique, I will make the writing better. That's a simple application that I could use every day and I actually think there's utility there. So that was his breakthrough moment.
EM Totally.
RD And honestly, I think all the people who are pushing AGI sort of missed the point of what's good about these large language models. It's that it makes things a little better. You can get the key points out of these research papers. You can simplify whatever law brief is in front of you. You can get these quick ways to understand it instead of spending two-three hours.
BP You can think of it as a universal librarian that can do great summary or bring you great knowledge. It can also be a good explainer. If it gets to the level of being sort of a critic and an interlocutor and a conversation partner, I think that's where it starts to tip the scales for me to be like, “Well, this is something like an AGI.” Its intelligence is boosting my intelligence via this discourse, not just looking things up for me or summarizing things for me.
EM Condensing information for me, like “Pull the high points out of this report for me,” versus “Help me make this essay stronger, more persuasive, better organized,” those seem like different order levels of functionality for sure.
BP And not to say that it hasn't seen essays and law briefs before, but what he's giving it is stuff that doesn't exist online. It's not stuff that it's trained on. It's some new case and some relatively novel scenario. So Ryan, I think your point is right that most of the utility for most people lies in the simple stuff, the low hanging fruit, but it's getting into a higher order. And then there is another piece of it, which is for me, I'm a terrible coder, but now I can make a few things. And if I was a slightly better coder, I could build bigger things or better toys. I do think on the coding side, it unlocks a certain level of the ability to use new languages or acquire new skills or sort of venture a little bit out of your comfort zone.
RD And I think that's exciting, and being able to use natural language was sort of always a goal of some of the early programming languages, instead of just doing machine language.
BP Here comes the old papers for computer science guy.
RD But I think that'd be exciting. Instead of having to figure out the esoteric priest language to understand the machine, you could just be like, “Hey, I'm bored. Give me a game that lets me punch dragons for money or whatever.”
EM You guys talked about kissing dragons on the last episode. Kissing, punching dragons.
RD The old dragon puncher 5000. We're all about the dragons here.
BP And then I will say we just had our company all-hands and a lot of discussion about this stuff. And one of the ideas that was floated across, Ryan, to your point, was somebody said, “Let's not get ahead of ourselves. At this point, I think of AI as just being statistics with language that can add value, but isn’t AGI, so don't fool yourself.” But statistics with language turns out to be really powerful, or statistics with any kind of framework, not just language, when AI can solve a math proof that went unsolved for hundreds of years or come up with a new drug idea that beats out everything we've seen at the trial. So generative AI, even if it doesn't know what it's doing, is powerful. Its brute force ability to spit out a thousand ideas and one of them turns out to be gold is powerful, even if the machine itself could care less about what it's saying.
EM Back to the monkeys with the typewriters there.
BP Yes. To harness the million monkeys with the typewriters is to harness great power.
RD I think there's a question, Eira, you added to the newsletter for this week about whether large language models are something about Wittgenstein and the use of language. And I thought that was an interesting one, but I think ultimately the question is, “What do you mean ‘use’?”
EM Isn't that always the question? So the question was, does the use of AI large language models support Wittgenstein's position that ‘meaning is use’? So I am not somebody who knows really anything about Wittgenstein. My brother wrote his undergraduate thesis on Wittgenstein, so I am going to be harshly fact checked if I say anything. But the idea is basically, if an AI, if a large language model is producing text based on statistical analysis, and that the text seemed to us to be meaningful, then is the upshot of that that meaning really just equates to use?
RD If it's able to produce a coherent sentence, it's using language. Is that the essence of it?
EM Yeah, I think that was the essence of this person's question. So he talks about words being meaning families, so this is Wittgenstein’s formulation. The specific meaning of the word is determined by, or maybe even is, its use in context. So once we get an AI that's able to engage in that same kind of context association, that same kind of production of language games, what are the sort of philosophical implications of that for language?
BP I have to say that I increasingly side with this concept. I forget, Ryan, who was on the podcast with us and was like, “Well, I guess thoughts are just a random collection of numbers moved into electrical signals. It's all just chaos that we make order out of.” But how do people come up with slang and new words? They say something in a certain context, it feels like it has a certain meaning, other people pick it up and repeat it, and now a piece of nonsense sound has a meaning. There was nothing inherent to it before, and so the use is the meaning.
EM Exactly. That's a better way of describing it. A phrase doesn't inherently of itself mean anything until we assign it that meaning and agree on it.
BP As the father of some tweens, they come home and they're like, “Dad, that was so rizz.” And I'm like, “If you say so,” and then I have to look it up on Urban Dictionary because ‘rizz’ means something now.
[music plays]
BP All right, y'all. I hope you enjoyed the conversation. We had a few laughs. Like we said, we'd love to hear from you. We spent the first third of this podcast not talking about AI. Why? Because a listener wrote in and asked us an interesting question. And if you don't, we'll talk about AI the entire time every time, and we'll eventually end up talking about philosophy and nobody wants to hear that. So write in and tell us what you do want to hear. Come on the show and be a participant because we enjoy it. Before signing off, I want to share a link in the show notes. “Why feathers are one of evolution's cleverest inventions.” It's a dope story about why feathers are so amazing. And I have to share this story– my wife has been obsessed for years with the idea of becoming friends with a crow because of various Instagram videos and because crows are so smart and we've read various books. And so for the last six months, three months, we've been leaving shiny things and a little bit of food out on the table of our back deck near our porch. And over the last week we have made a crow friend and now he comes every morning and he caws and then we go outside and put the food and then he eats and then we stand next to him. It's very fun.
RD That's the life goals.
EM That is amazing. I actually had a little crow friend that I fed at my old house.
BP Oh, that's so cool.
EM So I had to kind of wean him off. I stopped putting food out every day, but he would perch on the porch outside my office window and just caw until I came out there with dry cat food, which it turns out is not what some crows like a lot but something that is good for them to eat.
BP Yeah, we’ve been using cat food.
EM I'm excited to see your crow friend.
BP I've also seen the crow bringing the food to another crow, so I feel like I'm supporting a whole family here.
EM Oh! I love that.
BP All right, y'all. Enough about crows. Let's talk about Stack Overflow users. Y'all are the best. You contribute knowledge to a platform. You don't get paid for it. You get some free internet points, and you're happy with that. Awarded April 14th to Arman Ordookhani. Arman Ordookhani, thank you. A Lifeboat Badge, which means you've helped save this question with a great answer. “How to free memory in Go. I started to learn Go, but I don't have enough free memory.” Well, Arman has got a solution for you. It's got 20 points and an accepted answer, so Arman gets a Lifeboat Badge, and has helped 30,000 people. We appreciate it, Armand. Congrats on your badge. As always, I am Ben Popper. I'm the Director of Content here at Stack Overflow. Find me on X @BenPopper. Email us with questions or suggestions: podcast@stackoverflow.com. Like I said, we're bringing on guests and doing show topics based on your contributions. We've gotten two or three so far, we want some more. And another great way to contribute if you're so inclined is to leave us a rating and a review, because it really helps.
RD I'm Ryan Donovan. I edit the blog here at Stack Overflow. You can find it at stackoverflow.blog. If you have an idea for something we should cover, please reach out to me on X @RThorDonovan.
EM And I am @EiraMaybe. I'm also on the Editorial Team at Stack Overflow. And if you want to talk about corvids or anything else, you can find me on X at @EiraMaybe, or you can email us some suggestions at podcast@stackoverflow.com.
BP All right. #CorvidChat, #CorvidGang. Thanks for listening, everybody, and we will talk to you soon.
RD See you at the crowbar.
BP See you at the crowbar.
[outro music plays]