I believe we are at the threshold of a new programming paradigm. As the latest advancements in AI make it more accessible and closer to a self-hosted utility, we are entering a world in which developers can articulate what they want to achieve in simple natural language terms. I call this paradigm semantic programming.
No one can deny that LLMs have disrupted the way developers code. By July '23, Github reported that 92% of all polled devs were using AI in their work. By November, Snyk reported it was already 96%. The exact figure may vary, but I think it's safe to say most developers are already using AI in their day to day.
I have seen two prominent ways of integrating AI into the workflow. The first is using chatbots like ChatGPT or Bard as a Q&A oracle to which you send your questions or ask for code. The second is as a linter on steroids that you install in your IDE and constantly gives you suggestions coming from a model trained for code completion.
In both scenarios, the workflow involves sending a request to a server—often a supercomputer—that hosts a humungous model trained on vast amounts of data. While there are smaller, self-hostable models, they perform poorly on most AI leaderboards, despite being quite resource-intensive. This is a grim reality, as only big players are able to offer useful AI these days, since the cost of running inference is too high for domestic computers.
It's hard to determine when it will be reasonable to run a good enough pre-trained model locally, because of the constant pace of breakthroughs we're seeing, such as quantization, mixture of experts, LoRAs or distillation. But even if we just consider Moore's Law, it seems it will be a reality soon enough. And when that happens, maybe semantic programming becomes the new normal:
I know, using a trillion-parameter neural network to add three plus five seems cumbersome, even triggering. But so does shipping Chromium with every desktop app just to ignore platform compatibility, yet today it's standard practice with frameworks like Electron. Computer science is a tale of programmers embracing lazy abstractions whenever hardware gets faster.
The add example is an overkill for illustration, and I hope we don't do basic arithmetic this way anytime soon. But methods like to_html would require much more time to handcraft, if that's even possible. Maybe semantic programming becomes simply another tool in the set, same as other niche paradigms like constraint or symbolic programming.
Quality-wise, the main problem is how unreliable the output is. We could have next-token limitations tailored to the problem. For instance, we limit tokens for is_kids_safe output to be either 1 or 0, or dynamically constrain the next token for to_html to adjust to some regex for valid HTML. But these ideas won't get us any further in having a formal understanding of the reasoning behind each answer, nor will it give us mathematical certainity that the algorithm is correct.
Performance-wise, it's easy to see its limitations. Running this tiny example available here requires sending 109 tokens and getting 33 back, which costs $0.000104 with GPT-3.5. This is not a huge price for complex operations with short outputs like is_kids_safe, but longer texts or frequent calls could make the costs add up. Plus, server round trips take ~100ms, which is less than ideal for some seamless code integrations.
Despite all these problems, I'm really excited about this new way of coding. It enables functionalities that were plainly impossible before, like this anything-to-HTML converter. It democratizes coding, allowing people with no previous experience to craft on their own solutions. It shines in contexts where we can be tolerant to errors but can also work in critical contexts, such as law or medicine, by transforming human labor into supervision tasks. And most importantly, it enables, for the first time in history, a way to embed human intuition into code.
7 comments
I've become a huge an of a platformer game called Noita, where every pixel is a dynamic part of a simulation. These pixels can interact with each other and the player in complex and often unpredictable ways. You're plunged into a procedurally generated cave, that you have to explore and descend into its depths. It seems like the game was governed by realistic physics material interactions: oil ignites, ice melts, metal rusts, gases explode, and acid dissolves almost anything, including the player. And mostly everything kills you.
This intricate simulation is powered by a custom-built game engine fittingly named Falling Everything. I was curious about how it would handle such a vast array of interactions, and I was surprised by the elegant simplicity of its design. Basic particle interactions, governed by a few rules, result in rich, emergent behaviors. This inspired me to create a simplified version of the engine.
Drag to add sand
Consider sand in our simulator: it moves to an empty pixel below or, if blocked, to the diagonals. This basic rule simulates gravity and creates piles of sand in which new grains slide down. Since sand always goes down and never up, you can just scan the map from top to bottom and move each grain down until it's blocked.
Drag to add water
Water follows a similar logic: it tries to move down like sand, but if blocked, it tries to move sideways; this is, swapping places with a pixel to the left or right if it's empty. This creates a fluid that flows downwards and sideways, stabilizing into a flat surface. We can allow sand to swap places with water but not the other way around, so sand can't sink into water. This is how the game handles liquids and their interactions with solids.
Choose particle types with the left palette
Expanding this system is equally rewarding. Gases are just water that flows upwards. Walls don't interact with anything. Fire just disappears if random() > 0.8, spreads through gas, and when touching water both become gas particles. It's not hard to imagine how this could be extended to other elements that follow simple rules like grass, lava, ice, or electricity.
Feel free to check out the fullscreen version of this tiny simulator. It only contains these six basic particles but it's already fun to play with. You can also extend it if you want, the whole thing is a single HTML file, but don't expect the cleanest code ever as it was a 3 hour hack.
In general, I think this is an interesting approach to world designs with so much untapped potential for videogames, probably due to its unpredictability. And it would be interesting to port this to a 3D world, probably using higher-level memoizations akin to Hashlife to process more than one pixel at a time.
No comments
It's no secret that authenticating into services is an unresolved topic. With time, we have managed to make them more secure, but that was at the expense of user experience. The new generation of mail codes and authenticator apps has moved us from the ease of one-click browser autocomplete to complex ordeals involving multiple steps and sometimes multiple devices.
Last month, I was logging into Notion after it automatically logged me out, and I couldn't help but think "It feels like I'm logging in here every second week; maybe I'm doing something wrong." After a long examination of the settings, I decided to open a ticket asking if the session length was indeed that short. The response from Notion's team was prompt and specific, a great example of customer service. However, the content of the answer was less pleasing.
Notion is not alone in this; many other services enforce similarly short sessions and uncomfortable methods. This has me pondering the evolution of our authentication methods, from their ancient beginnings to modern complexities. Let's take a look at the history of authentication methods and rate them on two scales: user experience and security.
The first recorded password in western history is the book of Judges. Within the text, Gileadite soldiers used the word "shibboleth" to detect their enemies, the Ephraimites. The Ephraimites spoke in a different dialect so that they would say "sibboleth" instead. Experience ★★★★★: you just had to say a word. Security ☆☆☆☆☆: there's a single word to authenticate multiple users and it can be cracked by learning how to spell it.
Ancient Romans also relied on passwords in a similar manner called them "watchwords". Every night, roman military guards would pass around a wooden tablet with the watchword inscribed and every military man would pass the tablet around until every encampment marked their initials. During night patrols, soldiers would whisper the watchword to identify allies. Experience ★★★☆☆: you just had to say a word but you have to memorize it every day. Security ★☆☆☆☆: it changes every day, but it's still a single word, and without a "forgot password" button, a wrong answer would mean a spear in the gut.
Fast forward to the '20s, alcohol became illegal in the US, and speakeasies (illegal drinking establishments) were born. To enter the speakeasy, people had to quietly whisper a code word to keep law enforcement from finding out. Code words were ridiculous, to say the least: coffin varnish, monkey rum, panther sweat, and tarantula juice, to name a few. Experience ★★★★☆: you just had to say a word, and they were made to be memorable. Security ★☆☆☆☆: it's a single word, and it's not even a secret, but at least you don't get stabbed for getting it wrong.
The first recorded usage of a password in the digital age is attributed to Dr. Fernando Corbató. In the 60's, monolithic machines could only work on one problem at a time, which meant that the queue of jobs waiting to be processed was huge and a lot of processing time was lost. He developed an operating system called the Compatible Time-Sharing System (CTSS) that broke large processing tasks into smaller components and gave small slices of time to each task. Since multiple users were sharing one computer, files had to be assigned to individual researchers and available only to them, so he gave every user a unique name and password to access their files stored in the database. However, these passwords were stored in a plaintext file in the computer and there were a few cases of accidental and intentional password leaks. Experience ★★★☆☆: you have to remember a user and password. Security ★★☆☆☆: it's one per user, but they're stored in plaintext.
To prevent the problem of plaintext passwords, Robert Morris and Ken Thompson developed a simulation of a World War 2 crypto machine that scrambled the password before storing it into the system. This way, the system could ask for the password, scramble it, and compare it to the scrambled password stored in the system, a process called one-way hashing. This simulation was included in 6th Edition Unix in 1974, and got several improvements up to our days, but the basic idea remains the same. Experience ★★★☆☆: you have to remember a user and password. Security ★★★☆☆: it's no longer plaintext, but stealing it would still give you access to the system.
Over time, many different problems arised from the fact that people use the same password for multiple services, so the industry started to push for unique passwords for each service. This was a problem for users, since they had to remember a lot of passwords, and password managers were borned. The first password manager was developed by Bruce Schneier in 1997, and currently every major browser comes with a built-in one, often with an option to generate strong passwords and store them for you. Experience ★★★★☆: you have to remember a master password, but the browser remembers the rest. Security ★★★★☆: it's no longer plaintext, but the master password is the weakest link in the chain.
Phishing attacks and data breaches have made passwords a liability, so the industry has been pushing for multiple-factor authentication (MFA) for a while now. 2FA is a method of authentication that requires two different factors to verify your identity. The first factor is usually something you know, like a password, and the second factor is something you have, like a phone. This way, even if someone steals your password, they still need your phone to log in. There is a myriad of ways to implement 2FA, but the most common ones are SMS codes, authenticator apps, and mail codes. It is often used in conjunction with very short session lengths. Experience ☆☆☆☆☆: you have to remember something, have a phone or mail app, and it requires multiple steps. Security ★★★★☆: it's no longer a single factor, but it's still vulnerable to phishing attacks.
I, like most people, hate passwords and all means of authentication bureaucracy. And it looks like we're now at the lowest point in history in terms of UX. There is still hope with the rise of Single Sign-On (SSO) and biometrics. And certainly passkeys, which are getting a lot of traction lately, are a step in the right direction. But only time will tell if their adoption will be widespread enough to make a difference or if we'll be stuck in this dark age of authentication experience for a while.
Related posts:
13 comments
Una práctica que se ha adoptado últimamente en algunas ciudades es el llamado bono comercio. Es un bono con unidades limitadas que el ayuntamiento vende, y pueden usarse como forma de pago en algunos pequeños comercios físicos para incentivar la compra local.
El esquema actual en Sevilla ofrece estos bonos a 15€ cada uno, que pueden usarse para hacer compras por un valor de 25€. Debido a que sólo hay 55,000 unidades disponibles, se agotaron en cuestión de horas, una tendencia común con este tipo de bonos. La dinámica que esto genera es bastante interesante: el ayuntamiento, y por extensión todos los ciudadanos, está básicamente regalando 0.55 millones de euros para fomentar un gasto de 8.25 millones en comercios locales. Este beneficio se dirige especialmente a aquellos ciudadanos que han sido lo suficientemente rápidos para comprar los bonos en línea, con un límite de cinco por persona (o lo que es lo mismo, 125€ en valor de compra por 75€ invertidos). Es una manera peculiar de promover el comercio local.
Tras haberlos comprado, mi única queja es que el buscador web de comercios es bastante limitado, y la app tiene un mapa pero no ayuda mucho. Cuando quiero gastarlos en algo, o bien conozco el comercio (hay un buscador textual de comercios), o bien la categoría de lo que quiero (haría falta ver el mapa filtrado por categorías), o estoy explorando comercios (haría falta poder ver los establecimientos con niveles de zoom menores).
He notado en Twitter que la mayoría de usuarios que mencionan su bono lo han usado principalmente para adquirir libros, cómics y juegos. Esto podría ser porque los más tecnológicamente adeptos (por lo habitual más geeks) fueron los más rápidos en conseguirlos, o tal vez porque el buscador actual no haga fácil encontrar nuevos comercios.
Por eso, decidí extraer los datos de los comercios de la web y convertirlos en un archivo KML, disponible a través del enlace en la imagen de arriba. Este formato facilita cargar la información en plataformas como Google Earth Online, simplemente arrastrando el archivo. Al final, un pequeño toque de tecnología puede abrir nuevas puertas a los rincones comerciales de nuestra ciudad.
No comments
I'm taking the Advent of Code again this year. This time, I'm using it as an excuse to dive into Rust. Rust is a modern general-purpose programming language that focuses on performance and type safety. It ensures that the memory pointers are safe without relying on a garbage collector, a key feature contributing to its popularity.
However, the adjective modern is what sounds most interesting to me from all this. Until now, C++ has been my go-to compiled language when I need performance. But, having become accustomed to Python being my default choice for general purpose stuff, returning to C++ feels increasingly tedious.
C++ itself isn't the issue, in fact I think of it as the vanilla flavor of programming languages. It's the surrounding ecosystem that feels outdated: scrolling endlessly through 90's sites from when we used to have long attention spans, every dependency installation that is not just one command line away, every unintuitive build chain… all of this made me look for a fresh compiled language.
Being realistic, one can only excel in so many programming languages. I used to be a fairly good Java programmer, but I barely remember how to read from a file. Which is why I'm taking very seriously this search for a new default compiled language that I can use for the next 10 years without having to worry about the next new kid in the neighborhood.
When looking for alternatives, I started by filtering out the top languages from TIOBE — I know it's not a very scientific source, but it gives a good sense of what languages are at the top — because a language can be as cool as you want, that if I cannot find an easy answer to some obscure error or there are no connectors for a less known database I'd like to use, I'm out. And when looking here, Rust and Go where the only real contenders.
Initially I started using Go, as the syntax looked much simpler. I believe a programmer's efficiency is inversely proportional to how many things they have to remember. With good reason, a language ~64 times slower than C is the #1 simply because syntax is as important as semantics. Just as rookies often focus solely on the former, seasoned programmers sometimes overlook the latter. And in this sense, I decided to ignore Rust because I deemed the syntax as too alien. I mean, this is normal Rust code:
fn main() {
(1..=5).filter(|&x| x % 2 != 0)
.for_each(|x| println!("{} is odd", x));
}
However, I'm intrigued about why Rust consistently ranks as the most loved language on the yearly Stack Overflow Survey. I really enjoy coding in Go, how unified all the tooling is and how readable everything ends up being (iota aside), but I'm still not convinced on committing to it for the long term.
Hence, I'll be doing the Advent of Code using Rust with Copilot disabled, to also assess how reliant I've become on AI over the last two years. There is something about the Christmas spirit in doing things the old-fashioned, human way for a change.
1 comment