Vlad Iliescu

AI can already do most of the work here.

The boring stuff, the hard bits, or the things that have disastrous failure modes are the things we should collaborate on or employ specialists to handle.

Present-day models tend to produce code that is too defensive, too complex, too local in its reasoning.

More generally, liability concerns could mean that many current use cases for agents won’t be commercially viable.

putting your art, writing, expression out to be judged by others is an act of bravery as much as talent, and a lot of people lack bravery. -- Tom MacWright

The tooling as it exists today promotes absolutely nothing like the focus required to apply it judiciously. -- David Wilson

moral status does not appear just because the machine can emit text in the first person. -- Armin Ronacher

Overall I think we’re going to see a much higher quality of software, ironically around the same level than before 2000 when the net became usable by everyone to download fixes.

The real danger is that people stop thinking.

I’ve been using Codex (I know, I actually prefer it over Claude Code) more lately and one of the issues I’d run into from time to time was quota.

Your code is a strategy memo to strangers who will maintain it at 2am during an outage. -- Addy Osmani

[Sandboxing ]isn’t just about LLMs, but it becomes even more important now there are so many more people writing code often without knowing what they’re doing. … Being able to read a detailed specification and transform it into lines of code is the thing…

I’ve just released a new version of Grabit Clipit, my little command line app for saving full-text copies of webpages.

I very often use AI for inspiration.

If we don’t think hard about WHY we are doing work, and what work should look like, we are all going to drown in a wave of AI content. -- Ethan Mollick

For the longest time I’ve used iTerm2 as a replacement for Terminal.

I’ve been thinking about the best ways to use a model router ever since Azure AI Foundry added one in preview coupled with last week’s GPT-5 announcement, and…I have some thoughts on this.

About Last week I showed a roomful of GenAI enthusiasts some obvious and some not-so-obvious ways to, basically, make GitHub Copilot do their bidding.

I’ve recently configured my two favorite tools for LLM-assisted development (aider and Continue) to use o3-mini and DeepSeek-R1, with both models deployed in Azure AI Foundry.

In the meantime, I’ve renamed Grabit to Clipit to prevent PyPI naming collisions – see more details here.

Update: My PR adding built-in support for Azure OpenAI has been merged in smolagents v1.5.0 so all you need to do now is from smolagents.models import AzureOpenAIServerModel.

Microsoft has recently introduced the ability to cache Azure OpenAI prompts.

In the meantime, I’ve renamed Grabit to Clipit to prevent PyPI naming collisions – see more details here.

To tell you the truth, I’m a big fan of dependency injection.

I’ve (finally) decided to start writing regularly about the interesting things I’m learning every week.

Base Language Models This is a (discontinued) non-exhaustive list of large language models that I find interesting.

Update 2024-07-01: Microsoft have updated their billing to bill based on the number of tokens in your training file, so the info in this guide doesn’t apply anymore.

Update 2024-02-10 A few days ago Microsoft have announced that OpenAI’s text-to-speech voices were (finally) available on Azure OpenAI Service.

About Orca-2 The fine folk at Microsoft Research have recently published Orca 2, a new small large language model and apparently, it’s quite good!

These are the slides and notebook I’ve used during my talk on how to build an Internet-connected search assistant almost from scratch.

Recently, I was curious to see how easy it would be to run run Llama2 on my MacBook Pro M2, given the impressive amount of memory it makes available to both CPU and GPU.

Update: The Bing team fixed this behavior.

Ever since my article on running Stable Diffusion on Azure ML, I’ve been thinking about how to make the experience of using compute instances more enjoyable.

One of my favorite ways to deploy machine learning models in production is by using Azure Machine Learning, especially their new managed online endpoints feature.

About Ever since I had read Andy Salerno’s post on How to Draw Anything I was fascinated by the idea of using Stable Diffusion to, well, draw anything.

When compared to data scientists, traditional software developers have it easy.

GitHub Copilot is a tool that helps you write better, faster, and most importantly, more code.

The issue with machine learning pipelines is that they need to pass state from one step to another.

I had been unhappy with my website’s speed for quite some time.

Azure Automated ML offers a quick and easy way to train baseline models for all sorts of machine learning tasks such as regression, classification, and time series forecasting.

Ever since I wrote my very visual guide on predicting house prices using automated ml, I’ve been thinking about developing a more realistic use case.

Machine learning pipelines are a way to describe your machine learning process as a series of steps such as data extraction and preprocessing, but also training, deploying, and running models.

The Case for Automated Machine Learning I’m gonna start with an inconvenient truth: Machine learning is hard.

As you probably know, in less than three months’ time, Bitbucket will stop supporting Mercurial repositories.

This is a more detailed version of my Boy meets Girl talk, created specially for Microsoft Ignite | The Tour Amsterdam 2019.

Long title, I know 🤫.

This was a fun talk to write :).