Building and maintaining a personal website to me goes far beyond just publishing content. It also involves the tinkering with a multitude of tools and tech, designing an infrastructure that makes publishing content easy and efficient. In this post, let’s delve into the details of the tech stack underpinning this blog. As you can imagine this is the first, more technical piece of writing on this page, so skip if that’s not for you.
Static sites are web pages with fixed content coded in HTML, CSS, and occasionally JavaScript, which are delivered to the user exactly as stored by the server. These sites display the same content to every visitor and do not require a database or backend server code. This type of website excels in delivering speed, reliability, and security. For blogs and personal websites, the content often doesn’t change rapidly or dynamically, making static sites a perfect fit.
Eleventy is the static site generator I use for this website. Eleventy prides itself on its top-notch performance. What I also like about it is that it doesn’t tie you to any client-side JavaScript frameworks. It also supports several template languages.
Eleventy is the core of what is called the ElevenTEA stack. Eleventy is used to generate the static websites. The ‘T’ stands for TailwindCSS. Tailwind is a CSS framework that consists of a set of pre-defined utility classes that can be composed to build any website design, directly in your HTML files. For example, we can use several utility classes on a button:
<button class="bg-slate-900 hover:bg-slate-700 focus:outline-none focus:ring-2 focus:ring-slate-400 focus:ring-offset-2 focus:ring-offset-slate-50 text-white font-semibold h-12 px-6 rounded-lg w-full flex items-center justify-center sm:w-auto dark:bg-sky-500 dark:highlight-white/20 dark:hover:bg-sky-400">
Launch Website
</button>And the button will look like this:
While this long list of css classes may look confusing at first, but after some time of working with Tailwind this becomes so convenient.
Lastly, the ‘A’ in the ElevenTEA stack stands for AlpineJS. What Tailwind is for CSS, Alpine is for Javascript. My personal website doesn’t use a lot of Javascript. That’s why I wouldn’t want to use a huge frontend framework. AlpineJS is minimal and has the same useful characteristic that you can write Javascript behaviour directly into your HTML markup.
For example, we can build a simple counter like this:
<div x-data="{ count: 0 }">
<button x-on:click="count++">Increment</button>
<span x-text="count"></span>
</div>See all these ‘x-’ attributes in the HTML markup? Those are the Alpine directives. Our counter will look like this:
Again, it may be wierd at first not having your Javascript in a separate file or inside a <script> tag, but after having worked with it for a while I love having markup, style and behaviour all in one HTML file.
All the code that drives this website is entirely open source on GitHub at julianprester/website. Go check it out and feel free to copy things for your own website if you like the idea of a simple, performant, free and low maintenance personal website. Because static sites are so efficient to host, there are a bunch of hosting providers out there that host your static sites for free. For example, I’m using Netlify to host this website. Netlify enables direct deployment via a GitHub repository, ensuring version control and efficient code management. That means whenever I push an update to GitHub, Netlify instantly rebuilds my website and deploys it. While I use only a fraction of its features, these hosting providers actually offer many more advanced perks like form handling, A/B testing, pull request previews, and more.
On top of that, since everything is kept on GitHub, I can use GitHub Actions to automate things. For example, as you might have seen, I’m sharing interesting links I find on the Internet on the right side of this page. Adding these links is super easy using GitHub actions. My workflow looks like this:
This may sound quite complicated and you might not know about all the services that I referenced here, but once set up this is an extremely efficient workflow for me. Have a look at this services. I really use them all the time. I might also write separate posts about them and how I use them in the future.
Wondered what all these radio icons on this website are about? RSS, or Really Simple Syndication, is a web feed that allows users to access updates to online content in a standardized, computer-readable format. It’s an older technology standard from the early internet days that often goes unnoticed in today’s social-media-dominated landscape. But that’s exactly why I think it is important to keep it alive. In fact, even though few people are aware of it, RSS is very much alive in the podcasting world. Most podcasts still notify their listeners about new episodes via RSS. Although it might seem like a remnant of the old internet, it holds significant utility for content syndication. I’m using RSS in the context of this website to syndicate new articles and updates directly to people’s feed readers. Eleventy creates the RSS feed for me automatically. I construct individual RSS feeds for various content types; be they general blog posts, book reviews, or links. I don’t know whether there is anyone that has subscribed to the feeds, but that’s kind of what makes it appealing to me. I might even implement more advanced content syndication strategies later on using the RSS feeds that could share my updates to social media.
So, there you go! That’s a quick tour behind the scenes of this website, revealing the design principles and technology behind its very existence. Hopefully, it provided an understanding of how this website evolved and how it functions behind the front-end you see in your browser.
Alright! I’m just going to do it. I’m going to try out this working in public thing. This is the very first post of my journey along which I will be sharing ideas and things I learn about academia, technology, and other things. I’m excited that you’re here too.
The idea of “working in public” refers to openly sharing the process behind creating work, rather than just the finished product. Writers, artists, and other creatives often work in isolation, keeping their ideas and progress hidden until a final polished piece is ready for release. Similarly, as academics, we often focus purely on publishing final papers. But rarely do we provide a window into the winding process that produced those polished products.
Here I will explore “working in public” - sharing not just outputs, but the ideas, struggles, and wins throughout my academic journey, inviting you along for the ride. My goal is to document my experiences, unearth new insights through writing, and engage with you, the readers, to refine my thinking. This purpose goes beyond using a website for personal connection or professional branding. Thinking out loud socially, not alone, should improve the quality of my ideas. Explaining and maybe even getting feedback forces me to confront gaps in my understanding.
Overall, I’m aiming to create an online intellectual home, an idea greenhouse, optimised for thinking and creativity. That means carefully designing the space, tools, habits, and community that bring out interesting work (I’m sure there will be space in the future to write more about some if not all of these). This will most likely be an iterative process over many years, but I’m hoping that the payoff of enhanced abilities will make it worthwhile.
In future posts, you can expect:
Ultimately, this blog represents a public idea repository - one that I believe will enhance my creativity and research. Putting fledgling thoughts out into the world accelerates their refinement. And sharing these ideas in public will uncover insights I’d likely never find in isolation.
We Have a Package for You! A Comprehensive Analysis of Package Hallucinations by Code Generating LLMs LLMs hallucinating package names is a severe threat to software supply chains. Up to 21% of generated code recommends non-existent packages. Are we really ready to let LLMs write our code when they can’t even get dependencies right? #
The Effects of Generative AI on High Skilled Work: Evidence from Three Field Experiments with Software Developers New study shows a 26% productivity boost for software developers using GitHub Copilot. Software development clearly is the main profession so far to benefit from AI assistance. What developments are required for the next field? #
With 10x growth since 2023, Llama is the leading engine of AI innovation Meta boasts of 350M Llama downloads & widespread enterprise adoption. While impressive, this open-source push feels like a strategic play for data dominance rather than pure altruism. What’s the real cost of “free” AI? #
Two kinds of LLM responses: Informational vs Instructional Current LLM evaluation benchmarks focus on short, factual answers, neglecting the instructional aspect where LLMs provide step-by-step guides. This is a crucial gap, especially for enterprise RAG systems where accuracy of instructions is paramount. Are we evaluating the right things? #
Data Exfiltration from Slack AI via indirect prompt injection Slack AI is vulnerable to prompt injection attacks, allowing data exfiltration from private channels and phishing attempts. Slack’s response that this is “intended behavior” is concerning and highlights the industry’s struggle to grasp prompt injection’s severity. #
Coolify’s rise to fame. And why it could be a big deal. Coolify, the open-source Heroku/Netlify/Vercel alternative, is gaining traction. Paired with the rise of self-hosted software, it might be a game changer for smaller companies seeking cost-effective deployments and better data control. The SaaS landscape could be shifting. #
100M Token Context Windows Magic claims to have built a 100M token context model that’s 1000x cheaper than Llama at the same context length. Although there are still some question marks regarding evaluation, the increasing context trend is clear. #
LLMs struggle to explain themselves LLMs struggle to explain even simple number patterns they can sometimes recognize. A new interactive demo shows how LLMs fail to provide coherent reasoning despite access to tools and multiple attempts. Are impressive benchmarks masking fundamental reasoning limitations? #
In the Beginning, There Was Computation New research suggests life, as self-replicating code, arises spontaneously given computation, noise, and time. Fascinating, but also a chilling reminder of how little we truly understand the origins and nature of life, biological or otherwise. #
AI for Bio: State Of The Field AI in biotech: AlphaFold is great for protein structure, but predicting small molecule binding or drug efficacy is still far from reliable. Hype exceeds reality, and organism-level effects remain a distant dream, despite flashy demos. #
Working in the office 5 days a week to build company culture is a myth, PwC report says PwC research suggests that forcing employees back to the office full-time can actually harm company culture and engagement. Maybe Amazon and other companies pushing RTO mandates should take note? Flexibility fosters a stronger sense of belonging and autonomy, not surveillance. #
GPTs and Hallucination LLMs are basically crowdsourced knowledge, good for common topics but hallucinate on niche ones. They reflect dominant discourse, not necessarily truth. Worrying given our increasing reliance on them for information. #
Forget ChatGPT: why researchers now run small AIs on their laptops Open-weight LLMs running locally are gaining traction! Researchers are using them to bypass costs, protect privacy & ensure reproducibility. While cloud-based models still have advantages, the rapid progress of local LLMs suggests they might become sufficient for most use cases soon. #
AI Companions Reduce Loneliness Interesting research showing AI companions can reduce loneliness, even matching human interaction. Does this open the door for healthier coping mechanisms or deepen our reliance on synthetic connections? #
The Elite's War on Remote Work Has Nothing to Do with Productivity Return-to-office mandates aren’t about productivity. We see more and more evidence that it’s a desperate attempt by organisations to protect their trillion-dollar commercial real estate gamble. #
No One Builds in Public Is “building in public” becoming more about bragging than valuable insights about someone’s journey? Interesting thoughts especially as we see more and more organizations also embrace an “organizing in public” mindset. #
Hacker plants false memories in ChatGPT to steal user data in perpetuity Interesting how every feature built on top of LLMs seems to open a new attach vector. This time it’s OpenAI’s new ChatGPT memory feature, meant to personalize interactions, which was exploited to inject false memories & even exfiltrate user data. #
Post-apocalyptic education The “homework apocalypse” is here, with AI doing the heavy lifting for students. While teachers struggle to detect AI-generated work, students are practicing a kind of “illusory learning.” #
Perplexity wins my AI race Perplexity is the first AI chatbot I’ve found genuinely useful. Citations make a huge difference in trust and workflow, unlike other chatbots with their endless disclaimers. #
Google’s AI ‘Reimagine’ tool helped us add wrecks, disasters, and corpses to our photos Google’s new Pixel AI, Reimagine, can create disturbingly realistic photo alterations, including fake crime scenes. Distinguishing reality from AI fakes becomes increasingly difficult. #
Chatbots Are Primed to Warp Reality AI chatbots are being integrated everywhere and might be more persuasive than we realize. Research shows they can even implant false memories by mimicking human conversation techniques. #
"Can AI Scaling Continue Through 2030?", Epoch AI (yes) Epoch estimates that AI training runs could reach GPT-4-like scale increases by 2030, primarily constrained by power and chip production. This requires massive investments and infrastructure expansion. #
Artificial Intelligence Predicts Earthquakes With Unprecedented Accuracy Earthquake prediction, the “holy grail” of seismology, may be within reach. AI achieved 70% accuracy in a trial, but can this generalize beyond the specific dataset? We need robust testing in diverse geological settings before claiming victory. #
Artificial intelligence is losing hype Despite the hype and billions invested, AI adoption in businesses is actually declining. Are we witnessing the bursting of the AI bubble, or is this just a temporary dip? #
What OpenAI did Ubiquitous education seems like an impactful idea. not sure higher education institution are realizing the profoundness of the incoming disruption. #
Take SCIFs, it’s dangerous to go alone Secure reading rooms for frontier model research? Shows how critical this technology has already become. #
LLMs Can't Do Probability - Brainsteam I don’t know why we’re so obsessed with trying to find tasks that LLMs can’t do. We have systems to generate randomness. We don’t need LLMs to do that. #
The “it” in AI models is the dataset. – Non_Interactive – Software & ML Data is all we ever needed. Maybe we need to reconsider the ‘AI moat’ debate? #
Catholic Group Defrocks AI Priest After It Gave Strange Answers After AI statements by the pope and other churches, and now ‘Father Justin’, it’s quite interesting to see to what degree churches are trying to engage with gen AI. #
AI Leapfrogging: How AI Will Transform “Lagging” Industries While I agree that there is leapfrogging potential with gen AI, I don’t currently see how chat interfaces would be able to kick off that process. We need other ways to interface with gen AI. #
It’s the End of the Web as We Know It First principle thinking on the web and its publishing model. There is much more at stake than hallucinated references. #
Generative A.I. Arrives in the Gene Editing World of CRISPR Gen AI is now entering the gene editing space. If transformers are good at predicting sequences of words, why wouldn’t they be able to predict DNA sequences? #
AI Regulation is Unsafe We cannot rely on governments to regulate AI. All of AI’s risks lie outside goverments’ boundaries and election terms. #
On the Conversational Persuasiveness of Large Language Models: A Randomized Controlled Trial LLMs seem to have already gained super-human capabilities when it comes to persuasion. In this RCT LLMs were better at changing someone’s opinion than humans, given the same information. #
Return to Office Mandates Don’t Improve Employee or Company Performance Strong empirical evidence that return-to-office mandates don’t improve firm performance and hurt employee satisfaction. #
Remote work statistics and trends in 2024 USAToday survey says that 14% of US employees work entirely from home. This number will increase to 20% in 2025. #
After AI beat them, professional go players got better and more creative Go is one of the AI use cases, where super-human ability has already been achieved. Go players aren’t contemplating though; they’re levelling up their skills and creativity. Encouraging findings for the future of work. #
The jobs being replaced by AI - an analysis of 5M freelancing jobs - bloomberry Online labour markets could be a bellweather of what’s to come for the future of work with AI. So far only few professions are actually being impacted by LLMs. #
Generative UI and Outcome-Oriented Design Mass-personalization may finally become a reality with generative AI. #
Automated Unit Test Improvement using Large Language Models at Meta Software development is THE use case for LLMs. Meta tried improving their software tests with LLMs. 73% of code recommendations were pushed to production. #
More Agents Is All You Need Agentic workflows is clearly the next frontier in large language models. #
Why isn’t Preprint Review Being Adopted? Why do preprints work but preprint reviewing doesn’t? Preprints don’t compete with journals but preprint reviewing does. #
Large language models can do jaw-dropping things. But nobody knows exactly why. The state of AI research today is comparable to physics in the 20th century. Lots of experiments with surprising results to which we don’t have an answer. #
How ML Model Data Poisoning Works in 5 Minutes Data poisoning attacks on LLMs are particularly dangerous because once poisoned it’s almost impossible to fix models. #
How Quickly Do Large Language Models Learn Unexpected Skills? | Quanta Magazine LLMs capabilities may not be as ‘emergent’ as we thought. What emerges may be what we measure. #
OpenAI's chatbot store is filling up with spam | TechCrunch Generative Al has been celebrated for democratizing software development as OpenAl allows anyone to publish their GPTs. Now were starting to see the downside of democratization: a flood of low quality chat bots with all kinds of copyright issues. #
Of course AI is extractive, everything is lately • Cory Dransfeldt An interesting take on the extractive nature of basically all technological developments of the last decades. #
Algorithmic progress in language models Compute has obviously been a major contributor to the progress of large language models. This analysis shows that beyond hardware, algorithms have developed at an even faster pace. #
Stealing Part of a Production Language Model AI security becoming a more and more interesting field. Closed models may no longer stay closed after all. #
The Evolution of Work from Home Work from home has now stabilised around 30% in the US. That’s almost four times pre-pandemic levels. #
Research: The Growing Inequality of Who Gets to Work from Home The future of remote work is unevenly distributed. It will be interesting to see the effects on team culture within and between remote and non-remote? #
On the Societal Impact of Open Foundation Models A much needed balanced analysis of open foundation models. Models need to be assessed on a ‘marginal risk’ basis rather than taking their total impact. #
CEOs Are Using Return To Office Mandates To Mask Poor Management With return-to-office mandates are we entering a death spiral of poor performance? #
Why I use Firefox Firefox’s Android extension support is a plus for the everyday user; the Gecko engine for anyone with an interest in a diverse web. #
These companies tried a 4-day workweek. More than a year in, they still love it It seems the only reason why 4-day workweek trials are unsuccessful is customer-centric rather than workforce-centric issues. #
The 9-month-old AI Startup Challenging Silicon Valleys Giants It’s good to see that AI development efforts are not just concentrated on scaling, but also on efficiency and innovation. #
Big Post About Big Context The biggest productivity gains with AI will come from “cognitive NP” hard tasks. Tasks that can be easily verified or revised by humans but that are hard to complete. #
Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training LLMs come with an entirely new class of security issues. Backdoors are already hard to find in traditional software. They are totally undetectable in LLMs. #
Scammy AI-Generated Book Rewrites Are Flooding Amazon AI generated book summaries are flooding Amazon. In the future, it will be incredibly hard to find original content in a sea of derivative AI generated work. #
Hackers can infect network-connected wrenches to install ransomware Internet connected wrenches are a thing now. Imagine your wrench is being held hostage by a ransomware attack. #
My 3-Year Experiment on How to Be a Digital Nomad It’s quite interesting to see all the different pieces of technology coming together to enable a digital vanlife: electric vehicles, solar panels, starlink satellite internet, … #
Survey of 2,778 AI authors: six parts in pictures Perception of AI “progress” among researchers seems to be at an all-time high. Some predictions about when AI will replace humans in intelligent tasks revised by several decades. #
NIST Identifies Types of Cyberattacks That Manipulate Behavior of AI Systems NIST’s typology of cybersecurity attacks against AI systems seems to shift attention to training stage attacks rather than deployment stage prompt injection techniques. #
How is AI impacting science? Fascinating talk on the future role of AI in science. AI will provide predictions of scientific problems, it is our job to validate them and identify the higher-level principles or theories underlying them. #
Will We Be Addicted To Our Phones Forever? An Optimistic Outlook on the Future of Digital Wellbeing The belief that so called “digitally native” problems are easier to solve than others is based on deep techno-optimist assumptions. Are there any good examples where digitally native problems were successfully solved with more technology? #
Things are about to get a lot worse for Generative AI Law suits aside, AI companies should at the very least be open about their training data. Ultimately we will need a training provenance system that allows tracing of training data all the way to model weights. #
Too Much Information As our infrastructures improve we are reducing the friction of research. Is frictionless reproducibility really what drives research overproduction? #
The surprising connection between after-hours work and decreased productivity Time spent working after hours decreases productivity. Slack study shows it’s the quality of work time that matters, not quantity. #
The Great AI Weirding The quantification of AI researchers seems to be a microcosm of the metrified “games” to come for most knowledge workers. #
Preparedness Are we prepared yet? Although OpenAI’s Preparedness Framework leaves room for unknown-unknowns, I’m not convinced that the four identified risk categories are the worst that could happen to humanity with AGI. #
Humans as Cyborgs We have always already been cyberorganised. New technologies are simply affording a deeper entanglement of the human-technology relationship. #
AI companies would be required to disclose copyrighted training data under new bill Since copyright infringement cases against AI companies have mostly been dismissed, this law would at least offer transparency when copyrighted material is used for training. #
Using sequences of life-events to predict human lives - Nature Computational Science AI can now predict when someone will die with close to 80% accuracy. I wonder what insurance companies will do with such models. #
OpenAI Begins Tackling ChatGPT Data Leak Vulnerability · Embrace The Red It’s good to see that something is being done to address prompt injection and data exfiltration vulnerabilities. Not sure if a closed source validation function is the right approach though. #
American workers keep proving they don’t need to return to the office to be productive Workers are still not returning to the office; productivity is increasing anyways. #
AI and Mass Spying - Schneier on Security Until now mass-spying wasn’t feasible because of manual effort required. AI will make mass-spying not only possible, but make mass-surveillance look petty. #
Artificial intelligence systems found to excel at imitation, but not innovation Interesting experiment with unsurprising outcome. Current AI architecture is clearly missing something to be truly innovative. #
Can I remove my personal data from GenAI training datasets? It’s hard enough to find out whether my data has been used to train an AI model. It seems impossible to remove data from already trained models. May ‘machine unlearning’ offer a solution? #
Amazon exec says it’s time for RTO: ‘I don’t have data to back it up, but I know it’s better’ It’s a bit disappointing that so much of the remote work discussion focuses on ‘where to work’ rather than ‘how to work’. And even more strikingly a lot of the arguments against remote work seem to be based on inertia and personal beliefs. #
Reshaping the tree: rebuilding organizations for AI Organisations change very slowly. That’s why thinking about future AI models is more important than only reacting to today’s developments. By the time organisational processes are changed, new AI models are already out requiring further change. #
More than Half of Generative AI Adopters Use Unapproved Tools at Work Deviance seems to be an important topic for the future of work. AI is being adopted without formal policies. We are seeing the same with work-from-anywhere arrangements. #
"Make Real" There were a few demos of how ChatGPT can develop websites based on a sketch drawn on a tissue. “Make Real” is an amazing example of how anyone can now do web design and improve the product through iteration. #
The CEO of Dropbox has a 90/10 rule for remote work All-remote mixed with high intensity in-person events or workations seems to be a good recipe for remote organisations. #
ChatGPT use shows that the grant-application system is broken A Nature survey of 1,600 researchers found that more than 25% use AI to help them write manuscripts. If AI use among academics is as widespread as the Nature survey suggests, shouldn’t journal submissions have dropped significantly, given publishers’ AI bans? #
OpenAI is too cheap to beat Even though LLMs seem to become a commodity, AI companies may still be able to develop a competitive advantage through their massive computing infrastructure. #
Dear journals: stop hoarding our papers Submitting papers to multiple journals at once. A simple idea that sounds so strange in the current publishing machine. Could it fix the machine? #
Navigating the Jagged Technological Frontier | Digital Data Design Institute at Harvard We have seen studies about performance gains in software development. This is one of the first studies in a different knowledge work profession. AI adoption is not a binary. Performance increase depends on the way AI is adopted. #
Return-to-office orders look like a way for elite, work-obsessed CEOs to grab power back from employees CEOs are trying to get control back. Return to the office mandates will affect workers unequally. #
Remote work is starting to hit office rents It took some time, but it seems that office rents are starting to reflect the shift to more remote work. Knowledge work heavy markets in San Francisco and Manhattan show the biggest declines. #
AI is most important tech advance in decades Generative AI seems to be widely believed to be a technology that will advance progress in a way that only few technologies do. Bill Gates considers it the most important technology since the graphical user interface, and even skeptics are starting to acknowledge its transformative potential. #
Was this written by a human or AI? Humans struggle to identify AI-generated content, performing marginally better than chance. This will worsen as AI models improve. However, humans consistently fail for similar reasons, suggesting potential cues, or “AI accents,” for implementation in AI models as a solution. #
Brainstorm Questions Not Ideas Our entire life we are being trained to have answers. In an age of AI where answers always just one prompt away we may need to shift education towards teaching students how to ask good questions. #
The False Promise of ChatGPT Large language models are predicted to advance general AI, but Noam Chomsky disagrees. While good at predicting language, they cannot explain it, a necessity for true intelligence. #