Simon Willison: The Future of Open Source and AI
54:46 Share

I don't think i've seen any open models yet. None of them are are doing the things that we've traditionally associated with, like well governed open source project.

I love that coming at IT from the government angles. I hadn't at all that hints, developed experience people actually getting on board IT still isn't good. If somebody trying to lead vice and they want to get IT set up on the laptop.

that's still too difficult. Is still this perception around A I as far as you know, it's going to take people's .

job is going to take over. There are people whose whose jobs are threatened by A I if you are in who does, who works on commission, then the threat of people going to A I models is a very real threat. And IT really sucks because the models were trained on your work without information and now they're competing with you in the market place.

You have you have like truly one of the most uh unique backgrounds from helping create uh the super popular Price on development frame, jane. Uh now data set uh and you've also become sort of one of the most promotive voices in the AI equals this. And i'm i'm actually very curious, just like for your journey perspective, like what was this like some sort of transition point point like everything has been happy in A I is just like an extremely natural extension of the of the work that you've been doing uh for the last you know ten years.

So to me feels like an actual extension because the theme of my entry career has always been, how do I build more stuff more quickly, right? One of the things that I can use that will let me punch above my weight, like have more than will accelerate my development. And if you look at jane, jane, was that, like we built jane twenty one years ago now, for the idea was that was web development on journalists and deadlines.

We wanted to be able to produced the web application support things in the news where you have to turn around within a couple of days because he takes you six months. The new story isn't reliant anymore. So that was really important to us.

And and then just open source in general. I think open source has been the single biggest productivity boost to soft engineering in anyone's lifetime. I felt today that's not really acknowledge to much because developers, they have grown up with open source as a sort of default thing twenty years ago.

You have to convince companies to allow open source within their within their walls. You know, there are companies who have policies, say, absolutely no open source code under any circumstances today. Try writing any java script at all with that policy in place. And that just won't work.

So the thing for me has always been, what are tools that can add my tool belt, that let me do more, that let me produce more, produce more and Better code with the time that I have? And I got into large language models, originally with GPT three, when I realized that there were things that they could do, but just massively accelerated me already, like the using just the opener playgrounds interface I was doing things like getting IT to write me little snippy of jq code. And so for and IT worked really well, and so that I traded jq from at all that I didn't use because I wasn't fluting IT to at all that I was using a few times a week because I didn't have to learn the same tax anymore.

That's just kept on accelerating with past two and half years. Like I write apple script, I deployed code written go with unit tests and continuous deployment and continuous integration, despite not be influence and go. That's quite a shocking thing to be able to do.

And I love that. I love AI as a accelleration in a sort of personal productivity improvement. I hate A I as a replacement for humans. I love IT as a sort of cyber enhancement for humans, and i've been leaning William harder to that. And the ethics of the thing have always been so complicated.

The ethics of how the training is so murky that so much about this and that feels he feels a little bit like cheating when you're using IT as well. Just kind of interesting. So my wear around that, I said, okay, well, i'm going to help other people use IT too.

I'm going to be completely transparency how i'm using these tools. I will write about them. I will share my prompt.

And I already had a blog. I've had a blog for twenty years. So just leaning into that, saying, okay, well, I should start writing about this.

And the weird thing is, very few people write about things on the internet in long form these days, so you can have an outsized impera influence on the world just by being the person that write about a thing. Now, i've been writing my I started writing quite a lot about launching language models again maybe two years ago. And it's got me a long way, and it's got me invitations to all sorts of interesting things.

I met lots of people. I have developed a sort of I found myself in a semi influential position in this industry just from sitting home typing things up once, twice a week. That's that's prety amazing. I would encourage other people to do that too. Don't want to be the only person who's writing about this stuff.

I love that. And can you actually just back through this thread of, like, how you are seeing the world and what you're doing? Like, how do you actually spend your time? Is IT like, are you I remember when we chat last back in twenty, twenty two, you are like in the process of, uh, creating a start up and doing itself. Is that is that like how you spend most of your time as IT really spread out across bunch of itself?

It's seven spread across bunch of different stuff. The problem I had is a few years ago, I got some option to do the stanford journalism fellowship, where it's prom called J S K. fellowships.

They basically take medical real alist. And they pay them to spend a year at stanford on campus, hanging out with people and working on problems that are relevant to the news industries. Deliciously vague.

And i'm not a journalist, but i've done a lot of work at newspapers like with journalist. So I as to reply as the sort of wild card in the batch. And IT completely ruined me because they paid me to spend a year working on whatever I wanted to work on.

And once you've done that, very difficult to go back to other people telling you what to work on. And so what i've been trying to do, but the goal is I want to turn north my open source work into a sustainable full time thing that employees, not just me, but other people to work with me as well. And the the business model, i'm going off that when i'm doing the paid hosting things.

So data is open source. You can run IT yourself if you can install on the server or i'm building out data set cloud, which is the pay me and our amount of money a month. I will run IT for your team with the plugins that you need, a private space, all of that kind stuff and that i've been sort of preventing IT and trying to figure out the right product mix for quite a while now, getting very close this week.

I've been hacking on strike integration to like far up a like a payment page and all of that kind of stuff. But yeah, that's that's the the goal is effectively it's the word press model, right? Word press has built a very successful business by being the best paid hosting for word press.

Despite that, word press is open source and freely available. And I want I want to do effectively the same thing. But for data set, which is almost like a word, press the data, right? Word press will solve any content publishing problem that you have. I want to build open source software with plugins that can solve any data analysis, vision, ization and publication problem that you have. So if you want to share data with the world, I want those set to be the default all for doing that.

I guess you know a question kind of into that. And I guess I kind of also goes into potentially some use cases. But I guess you know from data that being or how is that being utilized in a AI projects if IT is? And then I guess you know what unique advantage does the offer from data exploration but also sharing?

So data set of relationship with A I has been interesting is some they started as that the two worlds didn't really overlap, right? Data sets. The project I started around about six years ago, and the idea was take data that fits in the equal database, which is anything the equality, you know, it's a database that has jacked on support and binary objects, so and make IT see you got a weapons face that lets you explore that data, publish that data online for all the people, and they're also interact with that data via on API.

So effectively, i'd say out of the box, Jason API, for anything that you can put in secret like database over time, as i'd built up with plugin system that lets plugins ad extra features, I started growing more of the data manipulation features. So I thin editing data for geo coding things, for changing table schemers kind of thing. And that made, made me realized that IT was effectively the problem when going after is data that's too big for excel and that you want to share with other people but that you don't really need to justify postgrads.

Or like a data engineer working on a full time, there's people talk about bigger data lot. I'm interested in small data, which I defined as anything that fits on my phone. And my phones got a terrible age of this space.

So small data, most companies don't actually have a terrible of data, you know. So I think that something really interesting about building out a sort of almost like the open source version of microsoft s cell crossed with a table, which is an interesting field to be playing in. It's definitely, it's its ambitious, you know.

And then the A I stuff, my earliest st applications of data step to A, I were just expLoring training, training data. So when stable, the fusion was raised a couple of years ago, they that was based on the lay on M L A I O N dataset. And I put a dataset instance with all of the land data so that people could search and see. I'll look at all came from pentreath and that so being able to explore those kinds of things, anyone who works in able to you, you have to get your hands dirty and look at the training data and no giver wants to do IT. So having an interface that helps people actually dig around in there is is just inherent use of anyway.

But then over the past six months, i've actually finally started bringing the two interested together and saying, okay, what are the AI driven features that I can add a data set? And i've got a three of those that i've been experimenting with so far, all of them targeting this world of effectively data journalism as the frame that i'm looking at this so that effectively its data science, but the kind that journalists do when they're working on investigative stories. And it's this wonderful field of this weird, the niche field of computer science, which I have been involved with since the juno days.

And so the three feature I ve got, firstly, there's the really obvious one. There's A I sequel assistance. You you type in the ink a question, quite that question. English, I look at the schema I turned out to seek and I I won. For you. The way i've been thinking about that, though, is in all of my experiments, I find that if you build the obvious version of that, like take the question and run the seek of query, use that to generate the answer, he gets IT right eighty percent of the time and twenty percent of the time. There's some kind of mistake in there, like maybe I thought that the state column was sea, but IT was actually california, little, little things like that.

So what i'm doing, instead, I want to build into face, which shows its working so you ask a question and instead of just getting an answer, you get bounced to a page with a sek query er the secret query is heavily commented and started experimenting with like what if I show you diagrams to help you understand which tables who joins together, pull out fragments of the tables so that you can see them, things like that. That's all quite early stages at the moment. But that the key thing for me is in journalism.

The last thing you want is somebody publishing a story that's just got untrue things about the world in because they're tooling, let them down. Journalists are actually very good at this kind of thing. Like journalists, good news organizations have fact checking.

They have editorial teams. The best data teams i've talked to actually like they. They do code review on the data as part of that fact checking process, so you can't publish something.

And I think IT was right as if they put out a data driven story. Two separate data engineers have independently worked on those numbers, and they have cross check their work. He's great, right?

That's responsible way i've doing this. So I want to make sure that if you're a journalist who doesn't speak equal and you use this tool, you get a result. You can send a link to your friend who does speak equal, and they can help you review what happened, like things like that, really, really important.

So feature is ask questions of your data with secret, which is sort of the obvious one. The one that i'm most excited about is structure data extraction. So if you're journalist and you spend so much time trying to gather data for your stories and ed garbage, try its government websites where you download P D F. And the P, D F is a monkey sand hand written documents like five going, all of that kind of thing.

And something that all of the these models are getting really good at is you give them unstructured day to IT can be tax or IT can be images and you give them the schema and say, you know, I want the name, address, the phone number that represents the and and they had, and they pulled those out. So I built a feature of the dataset again as a plug in, called them dataset extract and IT. Let you do exactly that.

You can define A A table. And because everything's a sequel like table in the end that has named, addressed and so forth, you can not hints like, I want the dates to be in year, year, year, year, year, month, month, day, day. Then you pass data into IT or you drag, drop an image onto IT, you click a button and you will extract that data you take in the table for you.

This is phenomenally effective. I've so far we've mainly been doing with them GPT for vision and GPT four O I started playing jeep for a mini. I want the next step is get IT working against claude and gami because both of those, especially german, I with its long context, but particularly in in, in, in, in, like exciting for this.

This is so cool. And i've talked to, I have friends who are like ceos of of A I companies who talked to lots of like fortune five hundred people. They he he told me structure, date, attraction is actually the thing they care most about, like these really big companies IT.

Turns out everyone has this problem. Everyone has garbage on drupal data when they need to make sense of IT being able to pipe that through a language model where the chance of things like fluctuates is a lot lower. It's not it's not zero, but but and that you still have to do that sort of frustrating spot checking process, but it's such a huge timesaver that's feature number two.

And then the third feature is something i'm calling enrichment. And this is more of a general purpose idea where the this started actually, with geo coding, right? You ve got a table full of addresses, and you want the last tude and longitude columns.

You can blood IT on the map. I built a enrichment feature where you can say, enrich this table, take this address column and turn them into a last few. But that felt as plugin.

So the first plugin was a geocodes plugin. The second plugging, plugging in was the language mother plugin. He can, okay, given this database table, take each row template, ted, in this way into this prompt, run the prompt and put the output of the prompt back in this column.

And now you can do all sorts of things. This is like a human language data clean up, to which I think is really exciting like this. There's there's so much potential for that sort of thing.

I'm working on versions of that where instead of the language model processing, the rose IT writes code for you. So you can say I need to double of the from this, this this number in turn in inter year and the language model, and then IT write a line of java script that does that. The javascript then gets executed across ten million rows.

So there's lots of potential for that kind of thing. Just code generation in general, I still feel, is the most exciting, like the most valuable applications of things. Weirdly, like nobody expected that when this technology first came out, that writing computer code that was going to be one of its most powerful.

Nothing but the idea that I can make those capabilities available to journalists to have never learned to program but our data literate enough that they understand that you'd want to have a function that does x one that's really cool. And that's one of things i've been most excited about in the field of journalism, is how do we take the kind of reporting that can only be done by the new york times in the washington post because they can afford software engineers in their news room. How do we make them available to smaller publications like your local newspaper? Should be able to do data reporting against the latest like government salary date base, without having to tap, without having to bring the programme room to do that, an analysis for them.

I think that such a very interesting concept, this data journalism and and I might again don't want to offend anyone here too, but I never really thought of journalists as people who are the weeds by on data using some code, you know these different A I offerings. But that sounds like, again, your your position is to provide that offering two individuals who might not have that technical experience.

but they say that data journalism is IT is a vibe. It's a lish field. IT is very viBrant as an annual conference called nia, which stands for the national institute of computer assisted reporting.

Because that's what they called date to journalists back in the seventies when they started doing IT right with they. There are people at this conference who remember main frames because they're y've been doing this kind of stuff for that long. And that conference I go over year, there's like a thousand people.

They are all nerds like me, basically like all sorts of different backgrounds and interest and geography, but fundamental that people who want to dig into data and find stories in IT. And I love that, right? That so again, janoo has been a big part of that over the over the years as well.

But that idea that you can, because there are stories in the data, like my favorite example and the city of san Frances o has an open data portal where you can download, like city data, one of the files on that is a csb file of every tree in safran ces. go. And there are one hundred and twenty five thousands trees, and they updated every day of the week. You can go, and you'll see last up data today, because they added three trees and remove five other trees. And they've got, the last year long, chew the address, who owns IT, who looks after IT, what species of trade is when I was planted, really rich data.

There are stories in that, right? Which neighbor ods get the most trees? Which neighbor od some? Which kinds of trees are more likely to to be removed after a set man time? There's so much, and of course nobody's lls those stories because what the hell is supposed to do with the csb flow of one hundred five thousand trees that that changes every day? So that's what I want to solve. I want to build software that means that you, that a reporter who again, is data literate, they know they were around excel, but they not nearly like writing piton code. I want them to be able to tell a story based on that weird three database.

So I am, why why frame this and and maybe my this is just my assumed framing, but like why build, you know, like an open source tool for this use case versus like I more like a consumer application sort of you abstract away all the technical detail and it's really sort of, you know throwing the data and ask questions.

So the original reason is that I was building this a side project. And I know from past experience that if your side project has user accounts, that's not a side project. That's like an unpaid job.

So so originally when I build this that i'm like it's going to be open source, people going have to install IT themselves. I am not responsible for anyone resetting anyone's password, anything like that. And that worked for the first couple of years when I was just a sign project.

But IT was a massive barrier to the people I most wanted to use. IT, like newsrooms telling me, use, I was easy, just get new urban to V, P, S, and then act, get, install python and three, and then do this, and then do that. You've lost them, right?

So the people I most wanted to use my software well. So eventually I had to admit that i'm going to have to have user. I need to host this for other people.

And I also it's a it's a great it's a proven business model for other sources charging people money per month for the last version. So once I was full time on IT, that became a lot more so of feasible to do. I tried a bunch other things.

First, like I built a desktop application version of dataset using electron APP so that people wouldn't have to install python and figure out how pip install works in all of those kinds of things I built. Imagine get IT running in web assemblies. There's a version of dates that went entirely and in a browser.

Using python, pia died. Python weather somebody thing is super cool. I'd built IT basically as a sort of experiment as possible, thinking, nobody will ever use this because I asked the outlaws, like five megabits of stuff, just to boot up twenty two thousand and twenty four, five, six weeks is fine.

That's like most react home pages. So the fact that you're loading an entire python interpreters to this Price isn't the block I thought I was gonna be. And so actually, quite a lot of people are using that. The web browse a based version of my service side, python h is quite amazing. yes.

So and then the other thing is that has to be open source, because the news industries and trouble already, if I get them hooked on a property tool and then shut IT down, i've cause I just caused even more harm, right? That's that's been very important to me from the very start is I want this project to have longevity, and I want people to be able to trust IT. When is initially was just me hacking on my last, I think, open source of amazing way to build that trust.

I i've had situations in my career where I I picked a commercial open source thing, the company and out of business. And we kept on using the software because we could because we had that, that escape clause and I wanted to make and also in use, there are news is an awful industry to build a start up in, because the newspapers, I have money. So and my plan is not to, I do not intend to make money from the news industry the long term.

My theory is any software I build that helps journalists tell stories with data is valuable to everyone else on the planet who has a data and wants to tell stories with IT, like there is nothing I could build for journalists that wouldn't also be useful for a corporate strategy department or quants. So all of these different countries of people. So i'm not worried about sort of pigeon holding myself from to the street that doesn't make any money because I think the stuff on building is is applicable way wider than that.

I love that perspective. I am also curious like IT seems like again, I think you're united suited to answer this question given your perspective on unsound of how developers have been building with a different tools over the last twenty years.

But do you have a sense of specifically on the A I front like IT seems like every time there's a new model or a new thing, like the developer willingness right now to try that thing or built with IT or adopting new tool is incredibly high. And it's not like my instinct a year ago is that like we basically have like a year worth of attention where people are just eventually gna burn out, they're not really interested and try the new thing. But IT feels like somehow that hasn't slow down and anything is accelerated and maybe the ecosystem has just expanded.

But do you think we i'm curious if there's like a time bound for one. People are really like, okay, you know what? This model was really good.

This tool was really good. I'm done trying all the new stuff like just give me the thing that works. I don't think about this anymore.

That's a really interesting question. I mean, i'm probably not qualified to this because I ve always been an earlier adopter like I will try new things that's that's again, I ort of interviewing things to add them into my my tall box overall.

And I mean, the valuation model is so difficult, right? Like a new model comes out, how am I supposed to tell if IT is incrementally Better for the things that i'm doing? Then the model was using previously.

And i've got a another sort of weird little open so side project that I might start pointing at that as an eval's toll. I rather not build one. I'm hoping somebody else saw that one really well before I have to take look count.

But this is so my other big open source projects. I've got two major projects right now. I've got dataset, which had proved up to my the one is this package.

Could this toko LLM, and it's a command line python tol for working with language models, which started initially, was just for working with OpenAI. I wanted to be able to basically pipe files on my terminal into opener, into into GPT for and say, do this thing with them. And so I built that not really fun.

And then because I built legal systems before the dataset, I thought, well, would would be interesting if you could install pluggin S S. To this thing at extra models. And I got that working.

And now there's like two hundred plus models that are available through this single tool. If you install plugins, some of them, the API models like OpenAI and anthropic, some of the local models using like lambda C P P. So you can actually pick you can almost IT install lama until laptop.

Now with my talk is really fun, like there's there's a few little education around that. But that's been really interesting because the other feature of this tool is that anything you do that gets logged to a city light database, so simple light is kind of the substrates that all of my tools into into, Operate through. And that means that I can do research.

Just my mucking round in my, anything that I try against any model gets logged in my second light database, which now like ninety megabits of recorded prompts and responses across a hundred different models. And so theoretically, I can then analyze that. So okay, given this prom against these different models, have that work out, haven't actually been doing that all we've been holding the data without doing the analysis step, but that's always on the cards.

So yeah, so this also gives me an abstraction layer against different models, which is great, right? I can run the same prom against GPT four and then against pull three point five minute and against germany one point five pro, just by changing the dash and flag in the thing that i'm typing on my in the command line so that my evs project is a plug in for ella, that a new commands to LLM, which lets you, at the moment, the idea is that you can run a whole bunch process cept defining a yambo file. So you will have an eval, which is a yano file full of prompts and little assertion things like I should have the word paris in the in the thing that came back to you contested what's the capital france gave back the correct answer.

I haven't actually touched that project in a couple of months. It's i'm still i'm finding time for IT amongst everything else. But I like the idea that I want to be able to hold my own events and quickly run them against new models.

I want to be able to share them with other people. So I can download your yo file if like your favorite math question prompts and quickly execute. Doesn't see how those go. That that the sort of ambitious version of that, I honestly don't with the other things i've got going going on, I don't know if have time to put more effort into that before somebody else ves the problem Better. And that be great.

Simon. I guess as someone who is is not only building, but I was very active in the open source ecosystem. I was going through some of your tweet tired hopping on. And I think there was one that was kind of, you know, some raising markets like on burgs ah something around a medas open source model.

You know i'm trying to build open source and you know I think they come at they made as how do we get organizations to stop misusing a the the word open source ah so okay, I guess kind of from your perspective, you know what are you seeing in that space? You know how do we combat that? You know, again, building in that space would love to get your thought specifically on. This is a huge buzzword. Everyone's talking about up and source.

All these models are coming out like this is a tightly panic thing for maybe IT does matter to me. The problem is if you tell me something in is open source, if you're being if you're sticking to the open source start or definition of that that's used for information for me like, okay, I know what I I know my rights with regards the software.

And I can build on IT without having to go in lawyer and get them to review a custom license, all of that kind of thing most a lot. And some of the open source models, like the most most of the mystery models are on doing a country two license. That's fine. I know what I can do with IT things like the lama three and lama three point one license.

They're not open source because they have those additional clauses and the additional clauses seen relatively inoffensive, like one of them is you can use this model and without license, if you had less, if you had more than seven hundred million active users the day the model was released, which sounds like I say that just means no alibaba a, no apple, no microsoft, whatever. Except the question I have is, if I do to start up and it's successful and I want to sell my start up to apple, does that license clause suddenly kick in? And now metric get to veto my start of acquisition? Because apple have more than ten, seven hundred million uses.

Now the answer to that question might be no. It's completely irrelevant. But I have to get a lawyer to help me understand that.

So so that my aversion to having to ask layers questions is what one of the reasons this masses to me so much personally and the and then that's the larger argument that people make is it's not open source if you can't don't get the source code such that you can rebuild IT. And for a model, the source code is the training data. And obviously, you can't open source of the training data because it's all a license ff that's been scraped off for the web.

That one. I respect that opinion. I don't use that as my own personal definition because my definition is I don't have to call my lawyer and asking questions of our license.

So that's my side of selfish take on that. There's an interesting conspiracy theory about why matter slap the label open. Saw some things that that aren't, which is that if you look at the european unions, A I X, it's specifically cast things out for open source models. But IT doesn't say a definition of what open source model is.

So there may be an incentive right now for anyone who's got something even remote events state is definitely open source because there's legislation being written that uses that time inaccurate way and that again, that's it's a theory i've heard that sounds a right to me, but i'm not a an international lawyers, so I could not tell you for sure if that's what's going on. But yeah so that that's why I care. I care because if you tell me something open sils and IT really is, I can skip a whole bunch of like concern over that license. And most of these these open source license that aren't those, I still have to think about them.

Have you been engaging in all with um or or felt keeping an ye on the conversation ah that osi has been around trying to what open source A I mean, I know they're been putting a ton of effort to this, but they're still early.

I had a conversation with them in january just about my very loosely form opinions, pretty much what I just said you now and i've not followed since then. So I do know that because they remember for you to seeing the results of IT.

But I closely, yeah I am really optimistically that they will help this. I think my my personal opinion and I think like an in both work at an focus and spend a lot of time thinking about the sustainability of open source IT feels like the open source A I ecosystem has actually been like a net negative to the broad movement of open source because people are conflicting. You know what people have history ally called like source available. Like I have an open model that I can use that source available, like the final dot E X E file um vers, like in actual sustainable distributed governance, distributed you know community of people who are built around an open source project. It's funny that I don't think i've seen any open source projects or any open models yet that would sort of meet the threshold of like actually becoming an unfocused project because of the fact that none of them are are doing the things that we be traditionally associated with, like a real well governed open source project.

I love that coming at IT from the governance angles. I hadn't thought about that. All talking interesting.

You again, I think, is someone to logan point disadvantage. You know, you have all these large oranienburg are talking about open source in importance of open source. And then again, for my my perspective, which is test was trying to get funds back to the tools and which they use.

No, no one never has money allocated to support in the open source eco system, which is ironic. And even from past craters or founders that we spoke to on previous episodes, you know, it's interesting to see that it's been stated that open source of the competitive advantage. So being able to, you know, put your stuff out there, learn from IT, obviously, number of eyes and and resources allocated to these open source tools is is interesting.

Again, I think we're in a very interesting space. And I think as IT continues to get pushed out there, hopefully the definition becomes more clear for folks. But I definitely still little bit sticky as far as what that looks like.

Also frustration that you now have, it's getting political in the legislature saying we should ban open source and you like all that is so not how this was supposed to play out. You know this is the abuse of determines is getting in some very scary rtp ical arguments these days.

And leave is australia right? There was a country someone who basically Mandated that all uh organizations, I think I was, have open source uh, in their code. Ah sorry for beaching, but I know australia did something.

If my tax pay your money plays for code stick IT on get hub. That's so that quite I like I like that that's beginning to at least be talked about examples. Es.

yeah. So I mean, i'm curious from the one of the hat that you where is the work that you do with the python software foundation? I'm curious um what your reaction to this is. But IT feels like for a very long time and maybe this my bias from being in the Julia community and and sort of some other trying to push a bunch of languages that I weren't a python I think python an incredible ecosystem but IT feels like alums have actually banned, um perhaps .

unintentional.

timely to me, sort of an accelleration or the sort of semmens layer of a python as sort of the canonical language that everyone is going to use and build with.

Mostly because like a lot of the tools and the model is so good, the model are so good at generating python killed, but also like code interpreter and a bunch of these other things are actually like running python behind the scenes um which is sort of further accelleration ating the distribution of python to people who previously weren't using python. I'm curious what your action to that is like is this something positive for the ecosystem? Yeah I mean.

there's also this pitch. Pitch is the default, the of these moves being built in that that pyne system stuff. Yet the it's interesting like like I M A python javascript programmer and the training data for python java script massively outweight of the else. So these modes are incredibly good at python java scrip.

So as I post the users as a language, i'd get a massive boost myself, like bigger than if I was using rust primarily or or even go as just as just less that let the those ecosystems aren't as known by the elms. And then IT seems of teaching people to program. Python has always been, I pithum started out and evolved out of education based language many, many years ago.

It's always had a very strong helping people to learn programme thing. And that sort of bit, yeah, that definitely being accelerated by things like code interpreter, by a lot of these tools will like they will ever noticed some of these models are default to answering in bitha when he asked a generate code question, even though you didn't say what language you wanted. The flip side is pythons to develop experience for people actually getting on boarded still isn't good enough.

I think that's the single biggest weakness. Weakness of the language is if I if somebody is trying to learn by then they want to get IT set up on their laptop, that's still too difficult. And that and figuring out dependencies and all of that other stuff, the sitting works fantastically once you've claimed that initial arn curve.

But the initial learning curve is a little bit steep per it's quite a bit steep. But then I wanted IT to be. And so the P S.

F, i'm on the board of directors of the python software foundation. One of the initiatives that we're Sparking up at the moment is around that. It's around the on boarding and the the education.

And just help making sure that when you go to python dot ogg, you get the best possible set of instructions on how to get started in this language. That really matters to me. One means I love that code conceptor, is that you don't have to store python to use IT, right? That's that's amazing.

You know that something that that keeps on happening up. I mentioned I died earlier, the python way assembly thing, again, phenomenal, right? You can fire a full jupiter like data science environment in your browser on a phone even.

I mean that U I doesn't quite work on the phone, but he does. But but this stuff is possible. There's a lot of, there's a lot of great things around that. But yeah, I I think the python communities grist responsibility now is to is to carve down that initial learning to make sure that people who want to learn to programme and cannot can get started this business as possible.

Yeah, I too really quick questions. One, do you think that because of alums, IT will actually sort of on one hand, pipes on is getting really good because are so much data around IT.

But on the other hand, the company itself could be a really interesting accelerate for new languages or things like that because you can actually sort of like auto create, you know, all of the blog post all of the possible questions that one could have and like preemptively put that content out in the world where like, again, this was the chAllenge for us in the july. The system was like, know, there is a lot of questions that people to know the answer to. And just like the act of someone asking the first question could often times to be really, really difficult. I'm curious, you think that alone could sort of help. And like data IDE as an example, I have you thought about like how to get more content out there are so that the elms are Better added and stuff like that.

There's this idea that keeps some cooking out, the idea that you can take a program language library and try and create the best possible fifty thousand token prompt that teaches the little enough about that library that I can then answer questions using IT i've not actually seen.

And a really great inflections that idea yet, which surprises me because IT IT feels like some like i'd love to have the copy in pace this chunk into or or here's a closed project to a GPT a GPT or something that had not been preceded with this um so that I find really exciting. I love your idea that that maybe this does lower the barrier to launching new language because launch new program language is so hard, right? Like like you said, you need first.

You need that the standard library for IT can take ten years to evolve before you can do all of the things you need language do out of the box. And then on top of that, as the educational material and maybe, maybe this stuff make not faster, maybe find vented the program is tomorrow. I could knock out the URL passing like in like five minutes, because the language model wants to figure out enough of that language could could help me, help me do that.

So i'd love to see IT. I'd love the idea that there's just generally the idea that somebody might build a programme language, which is specifically good for language models like IT uses less tokens. It's it's it's easier for the models to get there.

Their matrix brains around that. I think that, that more exciting a couple of years ago today, the models are so powerful that IT doesn't feel like you need to optimize your language to work Better with the model, little figure and pretty much anything. But you know, that's interesting. I'd love to see a new programme language that use language models to to accelerate its adoption and its the development of its sounded library so that that could be really cool.

So I feel there's anyone whose position to do that, that sounds like this is a good cake for years.

Not i'm not going to do A P H D. I'm not going to invent a programing language like those. Those are off the table.

Yeah my my sort of um side comment to this is and this is a hopefully as shameless of a shameless plug for for at least thinking about long context to be the solution to this problem. In the gni one point five technical report, the deep mind team shared an example of how you can take. Low resource language that essentially is not represented on the internet or in training data and put the dictionary, and like not many shot example, but like hundred thousand shot example, of what I might look like to use that language and actually get human level performance on that language just through fully in context learning.

That's a human language.

Try yeah a human language.

Wow, that I can believe IT, that's fascinating.

So sounds like you have your third open source project.

Yeah, i'm genuinely i'm maintaining over two hundred and fifty projects at the moment in terms of if somebody reports a bug, I will fix that bug in there, will ship release most of the meta y they're like little plugins for data that add one one little like the on the the.

Having a pluggin ecosystem for at all is such a great way of doing open source because somebody who wants to feature in one of my tools, can you just build IT as a pluggin? And they don't even have to talk to me, they don't have to go review IT and they can ship IT to pipi and people can start using IT. Now I was completely hands off on the whole thing, and that's amazing, right? You can wake up in the morning in your software, grow a new feature overnight.

And if that feature sucks, IT doesn't matter because people have just not install that plugin. So like I built, a lot of my plugin is, quite Frankly, terrible ideas that i'd built as a plug in this experiment because there's no risk to the integrity of the core projects. If if I, if I should dumb featuring the core project, i've to maintain IT the next forever.

That's bad, right? And if it's a little plug in, who cares? It's fine.

Simon, one of the things I love about the conversations we have on this podcast for you and other people in the space, you know, everyone is creating and building these these phenomenal tools that seem to be extremely helpful to develop individuals, whatever the case might be. But I think that again, there's still this perception around A I. As far as you know, it's going to take people's jobs is going to take over.

And I saw one your tweet, which I think that was quoted as A I assistance makes programing knowledge and skill more valuable, not less. yes. So yeah, we love to kind of hear your thoughts on that and also how you kind of come back and tree talk to you, you know, hundreds of hundreds of individuals around these tools and how to implement them and kind of, you know calm them down around the fact of, again, it's not here to replace you, here to help you. And this is exactly how so I think I think IT obvious ly comes back to like that educational resources that you provide, but if you are to get your thoughts on that and share.

So this is a complex issue because there are people whose whose jobs are threatens by AI because but and it's weird like like if you're an illustrator who does who works on commission, then the threat of people going to AI models to like stable fusion stuff. And so that is a very real threat. And IT sut really sucks because the models on your work without your permission and they're competing with you in the marketplace like I I think we'll probably end up shaking out as as fair use and legal, but is still suck from a moral point of view if something if if that happens to you. So I won't say university AI is not a threat to to people.

My hope is that in all, in almost all of these cases, there is a thing you can move into, which is taking all of the skills and experience you had beforehand and using and and your ability to use these models, which means you Better at them, like, I would much rather work with a professional illustrator who uses language, who uses image generation models to produce greater than a rank amateur, who's using the same models and same thing, and the program is the same thing, right? Somebody who doesn't know how to program can use cloth three point five artifacts to produce something useful. Somebody who does not need know the program will do IT Better and faster and y'll ask Better questions of IT.

And they will produce a Better result and that. So I don't feel threatened by as a professional programmer, I don't feel threatened about about by these things. Now I can use them Better than almost anyone else based on my twenty years a programme experience.

No, I will like I can absolutely fly with code interpreter and with laud out facts. And because i've got twenty years of experience in what's possible, what isn't possible, what tracks to avoid all of that sort of stuff. And so you hit people saying, i've quit my computer science degree is no point me learning programing.

This career is dead and it's all gonna replaced. That really upsets me because I genuinely feel like like as a, as a sort of senior software engineer, the time I spends typing code into a computer is about ten percent of my work. And that work is like two to five times for more productive.

Now with the language models, which is great, no, I massively accelerated that one tenth of my work that actually the bit where I typed the code, but everything else, the figure out what needs to be, build, researching what the options are, talking to, talking to users, talking to state, all communicating things to help developers. I get little bits and pieces of boosts from A I there as well. But that's, again, based on my existing experience, knowing that these are the kinds of things that the models are useful for, these kinds of things that aren't model useful for something.

I keep on coming back to is that I think language models incredibly difficult to use as a piece of technology, and I don't think that's a known wodge nearly enough. Like how hard is IT to type something at to ChatGPT? Surely anyone can do that.

The problem is that the to get the really good results them, you have to have quite deep understanding of them, you have to understand training cuts of dates, and why can't a language model the number of letters in a word? And all there are all of these very uninsured or things that they can't do well, like they can't count and they can't they can't do mats. It's a computer system that can't multiply two large numbers together.

That's inherently absurd to people. And but IT can do all of this other stuff. And so building up that intuition as to this is appropriate to ask for language model.

This isn't takes a long time like i'm still learning new tricks might been using them on a daily basis to two years now. So I feel like, firstly, developing expertise and language models is not as easy as people think IT is. And second, if you're going to use them to build great software, there's way more to great software than just typing in the prompt and copy paste coda.

You have to be you have to have all an so if you want to learn to program right now, I think language models accelerate you enormously, like that learning curve I talked about with python. ChatGPT will help you get over that. If you get an error message from a missing semi colon and paste to charge p, it'll tell you what to do.

Where's previously you could have been stuff like an Alice scratched me ahead over what this this weird thing meant so that I love that I think that there there has never been a Better time to start learning to program because these tools are available to shave down that learning to b to help you move so much faster and the value that you could provide us so much more, right? Like me, as a experience program with eem assistance, I can build stuff I couldn't have built before. If you hire me, you will get more for your money than if, if, if I didn't have access to those tools.

My hope is that that then increases the IT, increases the size of the market programing skills. And this is where I am uncertain. Like if we make every programmer two to five times more productive, does that mean we need half the number of programmer? Or does IT mean that companies that previously would not have commissioned customer software because they couldn't afford to hire a team of five to work for six months on IT, they can now hire a team of two to work for one month on IT.

Does that mean that now all of these companies are saying, you know what, we will commission programmers to build that software focus? My optimistic hope is that it's the second one like the the pool of like what companies ever said, oh, well, we we've got enough program is now we can solve all of our problems, you know. So ideally, I would like to see this open up so many more opportunities for people who use computer, who automate computers, to solve problems in people's lives.

Like that's fundamentally what we doing is programmers. That's never gna go a way as a thing that we need to do. That's so yeah, that's my optimistic version of that.

I love that perspective, Simon. I completely grew through. I'm really IT seems like the more developed, there's an infinite amount of software that needs to be created in the world then like people are always going to yeah it's it's it's an exciting time to be a developer um on this thread of like you mentioned, how much more productive the models are making you today um which I think is like A I I um I feel that as well and I think about of people who feel that I do think there's a ton of rough edge. So with models and uncurious, I give you had your your magic Simon Wanda, you can sort of waive IT and you know the top three problems that you would solve from like a core Allan perspective.

So my biggest one, and this is that from the data generous mango, is I need a version that has the safety filters turned off.

And the reason I need that is that if you're journalists like you work with horrible sources of information, right? You you're tracking fash down your reading fashious message pods, trying to figure out the up to you're am dealing with reports of police violence, all of these different kinds of things and right now a lot sometimes you will feel some stuff like like like um you will feed material for report reporting into a language model. And I will refuse to answer your question.

I had the best version of this happened to me doing that. I was doing a live demo in other generis m conference in front of one hundred generalist. And I had a campaign finance report this like crp eight like scam document with handlin numbers about what somebody y's political donations.

I fed IT into cloud three ops and said, extract the data from this. As Jason and claud three ops said, I am not comfortable doing this. This campaign finance report has personal, private financial information that I don't want to I will not do this for you.

Live on stage in front of journalist IT was a brilliant moment, right? It's like roles of laughter, really, really funny. But IT emphasizes that if we've got this technology and we can use IT to extract campaign finance reports, which is a really valid thing to do.

And sometimes I just says no to you. That's really frustrating. But that gets in the way I talk to a journalist who'd been analyzing footage from police body cameras using.

IT might have been using german, I actually. And they said that one of the things they had to do is reminds the model that dogs bite people, because police dogs sometimes bite people. But models are kind of like tuned on the internet idea that no g all dogs are good dogs and no dog have a bit anyone is just very entertaining.

But this is problem, right? And so but so with my journalism, have some I i've had conversation to people like anthropic open on google saying, hey, how about the fills of model? They don't say that that.

But I think what goes to the head is why would we give journied access to an unfilled model? They're going to write stories about IT writing nc manifest and step afraid like but it's difficult to make the case that news organza to get privileged access to an unfilled models and what we sort of pinky promise that we won't to write written new stories about IT. Yeah, it's complicated. Yeah, that's the thing I really want. I'm kind of hoping that like long the three point one, somebody will put up a fine tune hugging face with the safety filter, turn off that will turn out to be released ful.

For journalism, I love to be respectful. Your time always like to close with two questions. So i'll take up the first one that no. And up the second one tells us about like what your personal A I productivity stack looks like today. Like what told you're using anything custom, you built all that stuff.

So right now my personal stack is I spend time in chat P T for code in interpreter. I spend time in law claude treatment. Five time is currently my default model for most things.

And I use IT for artifacts which I used to generate. I like building way more h mm javascript un. I used to, because I can have IT knocked out artifact prototype and then copy pace that into my work. That's kind of fun. I use my L M comline tall on a daily basis for things like cats, this py and violin and act, take meditations, all of that kind stuff. Amusing VS code would get of co pilots and that basically at its chapt, claude L M and copilot, i've run this out called mlc chat on my phone, which runs models on my phone, mainly for demos to show people. But i've used IT on an airplane occasionally, because IT IT knows enough python that IT can answer, like questions about things I forgotten.

And then the last question we like to ask is, is crazy arty August. So we got five months left in twenty twenty four um but I guess you know and the remainder of twenty twenty four may be a little bit twenty twenty five um something that you're excited or hopeful about um and then also something that potentially hope gets push or does not come to foist based on any any reasons you might have yourself.

So I don't want A G I frustrated that all of the big A I labs, like at least at least sixty percent of them, appear to be very much a pilled. And you know, we're going to build digital god. And I don't really care IT.

Like i'm scared not of a gi. I'm scared of what happens to society when so much disruption happens all in one go. I have trouble imagining that like universal basic income or something will come along and save us from the society upheaves from that, so that that are not particularly interested.

And also like I get frustrated about the amount of discussion in the space that around the sort of science fiction ideas, the distracts from the sort of practices of what you can do. My focuses all been what in the models I have my hand right now do that's useful because that way I can just sort of I can duck out of all of the the sort of futuristic conversations. That thing i'm not excited about, the thing I am excited about honestly, just I let the incremental improvement on these models is so exciting.

The fact that like we did do seem to be even off a little bit and this kind of GPT for capable, but still like three point five sonic came along and was Better like notably Better for my stuff apparently. Need to check out the below of german. I one point five from from a couple of days ago.

I'm so excited about claude three point five high school because i'm assuming it's going to be Priced the same as cloud high and is going to be outrageously useful. I love these. I love these fast and cheap models that are increasingly coming out.

And then the other thing is that the local models, I ruin my laptop, like i've been running local laptops on my laptop, local models every year now, and they're never been very good, like they sort of begin to get sort of chat g GPT. Three point five level. They have just had a massively ford in the past few weeks.

Like the past month, IT feels like people have found optimize zing techniques where you can have an ab model that rs on my laptop that's actually beating GPT three point five. And i'm really interested to see where that goes. Like maybe with a few more the binary quantization techniques or something, do I get to a point where my next mac can run a GPT for class model locally like that we might get there and that would be really exciting.

I love that we just released gema to the tub version um which out outscores three point five turbo on a bunch of stuff and you can run IT on your phone locally and i'm like this is crazy like this is like if you would ask people how long IT will take for some of those original three point five global models till I actually make IT onto the device. I think most people would have guessed that.

that was going be a really long time. But because you did, I didn't think you could track them. I thought GPT three point five model was too big to run on a consumer laptop.

And IT turns out with the optimization from jet and such like IT, that's just not sure you can do IT. Ah I I love that. I also I don't care, I don't care for my model knows who the king of france was in seventeen eighty two.

I want my model to be able to summarize text and generate code and know when to use tools and all of those kinds of things. And that's it's interesting in like I think five, three, the papers specifically said IT doesn't know that much about the world. Like IT doesn't know that world, but I can do these other things.

That's what I want. I want a model on my computer that knows when to run a search against being or whatever and that knows when to um but otherwise can just do tall usage and summize ation. And that feels like it's going to close to as well. I think we're almost we basically there now I can get a model on my laptop that can do useful things and doesn't know everything that's that's known about the world and wikipedia. And that's fine.

Yeah, time. I love this. Hearing your perspective is incredibly interesting. I'm gay. We got a chance to have you on. Thank you for taking the time and hopefully hopefull folks enjoy this conversation.

Thank you. Great at Simon.