In other words: a curated list of news 🗞

🔍 Explainable Optimization (XOpt) brings transparency and trust to powerful models. Discover how it works inside expopt. Via Ilker Birbil.

🌍 GenAI and optimization are finally closing the gap between proven power and real-world adoption, and experts will remain vital for building, auditing, and governing these systems. Read it here. Via Tinglong Dai, David Simchi-Levi, and two more colleagues.

✨ Structured reasoning for LLMs is closer than you think thanks to this new framework. Via Sai Rajeswar.

⚠️ A reality check on NVIDIA cuOpt: demo technology or a true alternative to established algorithms? Via Vasiliy Zamolotov.

👥 A hand-picked list of optimization leaders worth following; honored to be included. Thanks Jerome!

Found something I should see? Send it my way and I’ll credit you next week.

In the meantime…

Let’s keep optimizing,

Borja.

Consider these new posts on Saturdays your dose of OR discoveries without the time investment of finding them yourself.

In summary: a curated list of news.

Today’s a special edition. We’re going back to school, so now’s the time to get inspiration and learn new concepts!

🎨 Learning with visualizations is super powerful, especially when you’re on the way to learn complex algorithms. And if it’s in a beautiful manner, even better. I stumbled upon Learn Graph Theory by chance, and I see it as a great resource for learning graph theory (bet you didn’t see that coming). Thanks, Mohammed.

🎲 Speaking about beautiful visualizations, what about learning optimization with games? 13 games for 13 optimization problems that you will compare against the machine. Incredible work from Jonas Velasco. You can also subscribe to his newsletter!

🏭 Learning the fundamentals, the underlying mathematics, and practical coding examples of the industrial applications of mathematical optimization is great, and if it comes from a Decision Scientist at Amazon, even better. Donato Maragno just started a blog about that.

⭕️ To close the circle, let’s finish with visualizations that I found by chance too. Constant, logarithmic, linear, quadratic… What are they? You get it right: some categories of Big O notation. Click here to visualize them.

Found something I should see? Send it my way and I’ll credit you next week.

In the meantime…

Let’s keep optimizing,

Borja.

Consider these new posts on Saturdays your dose of OR discoveries without the time investment of finding them yourself.

In summary: a curated list of news.

🎆 The signal you need to know: optimization is getting pushed by several fields. Next NeurIPS conference in San Diego is hosting a workshop on GPU-accelerated and scalable optimization. Exciting times ahead! Via Tobia Marcucci.

🕹 In most real-world applications (especially in supply chain, logistics, and planning) MILP models and heuristics dominate, but how to structure and tune them for sequential, uncertain environments? Rethinking MILP as a Policy is a practical implementation of the Sequential Decision Analytics framework using Gurobi. By Adam DeJans Jr.

🏎 Accelerating MIP solvers is becoming a trend. You’ll do it with a new framework called Kernel Pump, a new heuristic framework designed to find an initial feasible solution to any generic MILP problem. Improves success rate compared to Feasibility Pump, reduces average running times, and find feasible solutions in cases where CPLEX fails within one hour. Via Lucas Assunçao de Almeida.

🎧 Understand the way in which data is transformed into better business outcomes through improved operational decision making: listen to The Decision Intelligence Lab, a podcast that explores the many diverse links in this information value chain, which is now being referred to as Decision Intelligence. Hosted by and , former OR entrepreneurs who launched successful businesses before becoming full-time university faculty members.

⚡️ Energy-based Transformers are more powerful than the current feed-forward propagation methods, or so that say Yann LeCun. You can read one interesting paper here by Alexi Gladstone.

🌍 Speaking of which, researchers from Harvard and MIT have published the most interesting study on the abilities of AI models today: can they develop internal world models, or do they stop at being great predictors of token sequences? Read the full write-up, via Alberto Romero García.

Found something I should see? Send it my way and I’ll credit you next week.

In the meantime…

Let’s keep optimizing,

Borja.

You know that I recently started this new section called Search Space. It’s aimed to be a curated list of news.

But this is a special issue since, as you know, I’m celebrating 🎉 I recently surpassed 100 posts.

Again, I never imagined Feasible would make it this far when I started 3 years ago. Thanks a lot for supporting this newsletter for that long period of time, really appreciate it.

This special issue is about you: your side projects, your companies, your personal brands or courses.

So let’s go to the contributions you're making to the field, ordered as I received the submissions:

🎎 You can schedule events like Anime Boston optimally. If you want to see an example of how to leverage binary integer programming (BIP) and data science to schedule an enjoyable itinerary for a 3-day anime convention, read this LinkedIn post and look at this GitHub repository by John Woodward.

🚀 Nextmv is a DecisionOps platform that accelerates optimization projects and teams. You can follow them on LinkedIn too, and I personally encourage you to subscribe to their newsletter.

🤖 You can use Explainable AI to understand optimization decisions with detailed constraint analysis and alternative evaluations. How? Visit this blog post by Solvice. They have more to come, including explainability integrated with LLMs.

🏭 Harumi is an AI-powered platform that uses Operations Research to prescribe the optimal scenarios for real economy companies — manufacturers, retailers, and logistics providers. Follow one of the co-founders, Miriam Koga.

💡 In optimization we learn that the solution is implicit in our problem statement. Zohar Strinka took that idea and coined the "Meta-Problem" to capture that we have a choice of which problem to solve, and we should use that power. I also encourage you to subscribe to her newsletter, it’s awesome!

⚡ You know you can accelerate Linear Programming with AI, right? Kamal Basulaiman knows it very well. Yes, you’re right, he’s the collaborator from the latest deep dive here, in Feasible.

🚚 Do you need a fully integrated vehicle routing and 3D loading optimization platform? Clover Optimization is your place, by Leopold Kuttner.

🧬 Learn how a genetic algorithm for lot streaming job shop scheduling works thanks to this GitHub repo, by Francisco Vallejo.

🌐 Open data infrastructure for computational research projects, another GitHub repo that’s worth a try, by Luca Wrabetz. He’s looking for beta testers!

🧩 Timefold offers out-of-the-box solutions for common complex planning problems. The standardization of models makes operations research accessible for companies. Cannot recommend enough their newsletter, Planned.

📦 Do you know what the Bin Packing Problem with Conflicts is about? Well, this research project developed by Rafael Coelho during his time in university explores it. Focuses on applying a metaheuristic to a classic combinatorial optimization problem. Read it here.

Thank you very much for this. I really enjoy giving you visibility here, as it’s the place where OR practitioners get together :-)

Whenever you want to further collaborate with Feasible, remember I’m just one email away from you!

In the meantime…

Let’s keep optimizing,

Borja.

Consider these new posts on Saturdays your dose of OR discoveries without the time investment of finding them yourself.

In summary: a curated list of news.

💭 LLMs and academic research: the capacity of LLMs in doing OR/MS research. An article by Hanzhao Wang, via Johannes Royset.

🪡 This also applies to OR projects… How to select the use cases that actually move the needle? By Deepak Mehta.

💡 Microsoft developed Kamino, a simple, provably effective algorithm with substantial real-world impact (double-digit latency cuts, sizeable resource savings, and headroom for more allocator instances) all achieved with minimal engineering complexity and no disruptive architectural changes. Via Ishai Menache.

🎲 I see RKO (Random-Key Optimizers) gaining momentum. Want another proof? Take a look at Antonio Chaves’ work presented at EURO 2025.

🏅 Google just announced the 87 professors selected for the 2025 Google Research Scholar Program. There are papers about optimization, just in case.

Found something I should see? Send it my way and I’ll credit you next week.

In the meantime…

Let’s keep optimizing,

Borja.

As you know, I'm trying something new.

Instead of letting all the interesting research and tools I discover each week disappear into my bookmark graveyard, I'm sharing the ones that made me pause and think this is clever or I need to remember this.

Consider it your dose of OR discoveries without the time investment of finding them yourself.

Let me know if this hits the mark: your feedback will determine if this becomes a regular thing.

📚 The Polymathic Engineer is a newsletter that talks about algorithms. I found it on LinkedIn as the creator shared how to get better at algorithms and data structures. Via Fernando Franco.

🧗‍♂️ If you want to leave a local maximum, you have to go down, so this technique seems to be effective. Via Benjamin Doerr.

🧭 Moving from strategic to tactical to operational decisions is not that easy, even when the problem resembles from one layer to the other. Sometimes you get infeasibilities or poor solutions. PAMSO comes to the rescue in that part. Via Can Li.

⚙️ MCP seems to be the de-facto standard to communicate with ChatGPT-like apps. What if we can wrap a solver like HiGHS with MCP so we can build models in plain English? It has its own caveats, but these approaches make sense in today’s world. Via Wilfred Springer.

Found something I should see? Send it my way and I’ll credit you next week.

In the meantime…

Let’s keep optimizing,

Borja.

As you know, I'm trying something new.

Instead of letting all the interesting research and tools I discover each week disappear into my bookmark graveyard, I'm sharing the ones that made me pause and think this is clever or I need to remember this.

Consider it your dose of OR discoveries without the time investment of finding them yourself. Read here the first and second edition if you didn’t.

Let me know if this hits the mark: your feedback will determine if this becomes a regular thing.

🤖 GenAI can help OR Engineers at some steps of the optimization workflow, like problem understanding > model development > solver execution > solution interpretation. Where does GenAI fit? An article by Deepak Mehta.

🔌 Artelys, FICO and NVIDIA collaborated to solve energy optimization problems through cuOpt, the GPU-based approach from NVIDIA.

💻 CPU, GPU, TPU, QPU… And what other Processing Units may help in optimization? Exactly: Lasser Processing Units, or LPUs like the ones from LightSolver.

📊 Benchmarking optimization algorithms is challenging, but with the updated chapter on the CP-SAT Primer, you’ll find it easier. Via Dominik Krupke.

🎲 Learning optimization through games is a great idea. And there’s another one in the space: The Cutting Plane Game. I’m a passionate about these games (that’s why OR From 0 to 1 teaches optimization through them). The author is Gonzalo Muñoz, Assistant Professor at Universidad de Chile.

🎮 Talking about games, if you want to explore Graph Theory… Try the Claude Shannon Game, by Mehul Jain.

Found something I should see? Send it my way and I’ll credit you next week.

In the meantime…

Let’s keep optimizing,

Borja.

As you know, I'm trying something new.

Instead of letting all the interesting research and tools I discover each week disappear into my bookmark graveyard, I'm sharing the ones that made me pause and think this is clever or I need to remember this.

Consider it your dose of OR discoveries without the time investment of finding them yourself. Read here the first edition if you didn’t.

Let me know if this hits the mark: your feedback will determine if this becomes a regular thing.

🚀 The XI Iberian Modeling Week will start in one month. It’s a joined initiative between Portugal and Spain designed to introduce undergraduate and master’s students to industrial mathematics through hands-on problem-solving. Each year, students from both countries work in teams on real-world challenges proposed by companies, under the supervision of academic and industry mentors; via Emilio Carrizosa.

🏭 Building a combinatorial optimization supply chain model that changes everything at Danone, enabling rapid scenario analysis across every plant-product-customer combination, by Kyle Cannington.

🏫 DualSchool (paper), or how LLMs may change the class in introduction to optimization courses, either if you’re a teacher or student, by Pascal Van Hentenryck.

🤖 Inventory Optimization AI, a Reinforcement Learning-based simulation environment and agent for intelligent inventory control in supply chain systems, by Adam Chen.

🧩 Formulations for optimization problems: an actively maintained, great website to learn modeling techniques, examples, and algorithms to solve hard problems, by Fabio Furini.

Found something I should see? Send it my way and I’ll credit you next week.

In the meantime…

Let’s keep optimizing,

Borja.

I'm trying something new today.

Instead of letting all the interesting research and tools I discover each week disappear into my bookmark graveyard, I'm sharing the ones that made me pause and think this is clever or I need to remember this.

Consider it your weekly dose of OR discoveries without the time investment of finding them yourself.

Let me know if this hits the mark: your feedback will determine if this becomes a regular thing.

🏎 How to find the shortest path, faster: Dijkstra’s algorithm dominated this problem, but now researchers from the Max Planck Institute for Informatics broke the barrier. If you want to read a free version of the paper, here.

⏳ For algorithms, a little memory outweights a lot of time: Ryan Williams’ 2025 proof shows that any algorithm taking T steps can be simulated with roughly √T memory, breaking a 50-year deadlock and proving that a sliver of space can beat oceans of time.

🎢 Introduction to Minimum Cost Flow Problem: what it is, and how to solve it with Python, by Reinhard Sellmair.

🧬 LLMs meet Evolutionary Algorithms: a nice summary from Eric Bonabeau of how both fields support each other in several ways, via Manuel López-Ibáñez.

🐍 pyOptInterface, a well-written and fast modeling framework, but what can it do? By Richard Oberdieck.

🎆 A parallel implementation in Java of Adaptive Large Neighborhood Search (jPALNS), by Nils Beckman.

🗺 spopt: a spatial optimization framework in Python by PySal team, lead by James Gaboardi, via Kyle Walker.

🚢 DecisionOps, or how to productionize OR models: a talk by ECCO’s Lead Data Scientist, Matthias Als.

Found something I should see? Send it my way and I’ll credit you next week.

In the meantime…

Let’s keep optimizing,

Borja.

Three days for one of the top three conferences in AI and ML.

Three days for Operations Research content too.

I know I told you the other day that I see optimization everywhere. And now I see it on AI and ML conferences too.

Today you’ll read in Feasible 10 papers to take a look at in ICLR. Most are posters, some others are interesting submissions as conference papers. Anyway, I’ll give you:

• 📌 1-liner

• 💡 Why that paper matters

• 🛠️ Proposal

And, of course, a link to the submitted paper.

Ready? Let’s dive in… 🪂

🤖 1. Learning-Guided Rolling Horizon Optimization for Long-Horizon Flexible Job-Shop Scheduling

📌 1-liner:

AI-powered scheduling for long-term, flexible manufacturing operations.

💡 Why it matters:

Integrating learning into Rolling Horizon Optimization (RHO) allows for faster and better solutions to long-horizon scheduling problems by reducing redundant computation across overlapping decision windows.

Flexible Job-Shop Scheduling is a core problem in manufacturing and operations. Improving solve time and quality here has direct ROI in production efficiency and resource usage.

🛠️ Proposal:

The authors propose a neural network to identify parts of the schedule that don’t need to be re-optimized in every horizon, reducing compute while improving outcome quality.

🔗 Paper

🚛 2. Boosting Neural Combinatorial Optimization for Large-Scale Vehicle Routing Problems

📌 1-liner:

Smart, scalable AI for solving massive delivery routing problems.

💡 Why it matters:

VRPs are central to logistics, delivery, and ride-sharing platforms. Better large-scale performance directly translates to cost savings and customer satisfaction.

Carefully designed architectures and decoding strategies make neural solvers more scalable for large vehicle routing problems.

🛠️ Proposal:

The paper improves decoding strategies and uses grouped node embeddings to make neural solvers more scalable to thousands of nodes.

🔗 Paper

🗣️ 3. Decision Information Meets LLMs: The Future of Explainable Operations Research

📌 1-liner:

Talking optimizers: using LLMs to explain complex OR decisions clearly.

Read more

I didn’t write any post with recent news since January 12th.

✅ Good thing: you’ll read the ones that caught my attention.

✖️ Bad thing: you’ll read the ones that caught my attention.

I mean, there’s a lot of stuff to cover lately, and that’s why you also receive now an edition of Feasible on Thursdays.

Today we’ll see AI agents, LLMs, and whether we can automate maths with AI. Be ready because this one will be intense. From more specific to more general use cases.

As we say in Spain… Vamos al turrón 🪂

🤖 Optimization modeling, the AI way

Just in case you missed it, AI Agents are a paradigm in which you can coordinate LLMs for better accuracy in their responses.

But it’s better if you read about them in deep from other, more reliable sources than me.

The first one is Chip Huyen. She explains agents in an easy-to-digest way, I cannot recommend more her Agents post.

Also Google released a white paper explaining agents in a 42-page PDF. For some reason you cannot access the original PDF, but here you can find the one submitted to kaggle.

Now that we know what agents are, their skills, how you can leverage them for your own usage… How can you use them without spending time in writing lines of code?

Read more

However, their ability to solve complex planning and decision-making tasks remains restricted.

Extracting critical information from a text-based optimization problem and building a mathematical model or algorithm to solve it is very challenging.

A traditional solver is the clear choice for complex decision-making tasks. They offer optimality guarantees (where needed), are trusted in Operations Research, and have powered critical decision-making in various fields for decades.

What if the obstacle is structuring the problem into a mathematical model?

Many researchers are using LLMs to extract decision variables, constraints, and objective functions from problem descriptions to automate model formulation. This approach has significant practical limitations and is theoretically promising.

If extracting a MIP model from an LLM fails, is there another way forward? Could we build something that doesn’t require formal mathematical structuring, just operating at a language level?

Today in Feasible we’ll see:

• 💥 Why traditional methods have struggled

• 🧠 A novel approach from DeepMind and its uniqueness

• 🤖 Why it represents a fundamental shift in LLM-driven problem-solving

🎧 At the top, you can listen to a podcast-like version of this post!

Before we start... I recently spoke at the University of Delhi about how OR can drive real-world impact.

Ready? Let’s dive in… 🪂

💥 What was tried and why it failed

There’s a significant research area around solving optimization problems through LLMs, by generating algorithms or mathematical models using solvers.

It’s a hot topic with great potential to change how we -OR practitioners- work.

Several studies have connected AI and optimization by using LLMs to automate problem formalization.

We’ve seen studies on the limits of language models as planning formalizers from Drexel University and a team from Harvard, MIT, and IBM that sought to solve real-world planning with LLMs.

The typical process to automate our work involves (i) extracting decision variables, constraints, and objectives (LLMs analyze natural language descriptions and convert them into a model/algorithm), (ii) coding the model/algorithm, and (iii) solving it with a solver or running the algorithm.

This approach was expected to work as it’s the natural way we usually work. LLMs excel at extracting information from text and solvers are efficient with well-defined problems. Automating this process could reduce human intervention and make optimization tools more accessible.

As Richard Oberdieck highlighted in LLM-ify me, LLMs can’t be a solver since they hallucinate, so this approach is ineffective in practice:

1. 📖 Extracting a model from text is harder than expected. Many real-world problems are loosely defined, making it difficult to extract precise decision variables, constraints, and objective functions. Moreover, LLMs can misinterpret them, leading to incomplete or incorrect problem definitions.

2. 🎯 Solvers need perfectly defined models. If one constraint is missing or incorrectly formulated, they fail to find a valid solution. Some constraints and objectives require fuzzy logic or rewording to incorporate them into the model.

3. 🐛 Coding errors. Generated code often requires manual debugging to run correctly in a solver. LLM-generated code is prone to errors when task complexity increases, such as in optimization problems.

What if there was a different approach? What if it’s not strictly necessary to translate text into a mathematical model? What if LLMs could act as a solver, getting solutions by manipulating them directly in the language space?

Let’s see Mind Evolution.

🧠 Mind Evolution: key contributions and highlights

While DeepMind explored generating heuristics with FunSearch, others proposed solutions to automatically develop algorithms for optimization problems using LLMs.

Like this NeurIPS 2024 paper or this IEEE Transactions on Evolutionary Computation one. (thanks Emilio for the references!)

Mind Evolution’s idea is clear: don’t translate a problem in plain language to a solver/algorithm. It’s unreliable.

Instead, provide solutions to problems manipulating them through language. The novelty lies in not needing defined decision variables, constraints, and objective functions.

It uses a Genetic Algorithm-like style to balance exploration and exploitation while keeping a set of possible solutions rather than static sampling or a single-solution refinement.

They retain the ideas of Genetic Algorithms, such as selection (choosing the best solutions based on task-specific evaluators), crossover (combining high-performing solutions to generate improved candidates), and mutation (introducing variations to avoid stagnation and enhance exploration).

How can they evaluate solutions through a fitness function? According to a Tencent paper, there’s a method to refine and self-improve LLM responses based on feedback.

Mind Evolution evaluates solutions through textual feedback and heuristic scoring rather than mathematical formulations.

🤖 Why Mind Evolution is a shift

OR practitioners are trained to structure problems mathematically.

We might consider other options for optimization problems, especially the challenging ones with fuzzy logic or complex descriptions.

Let’s examine the trade-offs between traditional solvers and heuristic-based refinement approaches like Mind Evolution instead of stating that structured optimization is challenging.

1️⃣ When does a solver make sense?

🔹 Well-defined problems. If a problem can be fully encoded with precise variables and constraints, a solver can find optimal solutions efficiently.

🔹 Optimality is required. Some industries (e.g., finance, manufacturing) need provable guarantees that heuristics cannot provide.

2️⃣ When could Mind Evolution be the preferable choice?

🚀 When problem definition is ambiguous, an iterative language-based approach helps prevent premature formalization errors if decision variables or constraints are hard to extract explicitly.

🚀 When you lack a team of OR experts, Mind Evolution operates directly through language, enabling non-technical users to solve their problems and reducing the time to an automated solution.

Mind Evolution complements traditional solutions like solvers or algorithms, rather than replacing them. It offers a modern way to solve unstructured problems.

🏁 Conclusion

Traditional approaches to optimizing LLM outputs have limitations. Through heuristic sampling, structured solvers, or LLM-assisted problem formalization, they struggle with ambiguity, require exact formulations, and suffer from longer debugging times.

Mind Evolution represents a paradigm shift. By applying evolutionary search principles, it enables LLMs to refine solutions without explicit mathematical formalization. This makes it more adaptable, scalable, and user-friendly than previous approaches.

Mind Evolution embraces the fluid and evolving nature of real-world decision-making, rather than fitting every problem into a rigid framework. This makes it a promising technique for optimizing LLM-driven reasoning.

Let’s keep optimizing,

Borja.

PS: the paper! Here you have it: Evolving Deeper LLM Thinking.

🗞️ Other related posts people like

December was a crazy month: the Advent of OR was super demanding for me (and you too if you participated!) with 24 optimization problems in just 24 days.

Plus 3 open mic live sessions.

Plus a bonus with:

… 21 videos,

… 77 minutes of videos,

… 4 sections covering context, data, model in Google OR-Tools and visualization with Streamlit,

… and a bonus inside the bonus about how to easily deploy the solution in the cloud so you don’t need to go with your laptop to stakeholders’ desks.

(👉 if you want access to that bonus, let me know replying to this email!)

But I didn’t stop writing here. And as much as I love writing, I also love reading. That’s why today you’ll read one of those posts that I needed to split into two 😅

I hope you understand it: there’s a lot to talk about in the OR space lately!

From optimizers for Deep Learning to a pragmatic view of Operations Research…

→ Are you ready? Let’s get started! 🪂

Read more

One of those days in which I share the latest news I’ve read around the Operations Research space!

From cutting-edge applied research for power grids in the US to top conferences around business impact with optimization…

Here we go! 🪂

Read more

That’s true, AI is faster than light, so not at the speed of AI.

In the past few weeks, there’s been a lot of news. For example, the new solver from Google, PDLP.

Let’s start from there. Today’s Feasible post will be about:

• 🧪 Is NVIDIA testing PDLP and GPUs?

• 🕸️ INFORMS and the importance of graphs through Walmart’s vision.

• 🌪️ Google Cloud’s AI infrastructure and its impact on OR companies.

I’ll also share why traditional solvers like Gurobi should adopt PDLP.

Let’s get started! 🪂

🧪 NVIDIA testing PDLP on their GPUs

You’ve read about PDLP in Feasible at least 3 times.

It’s the most important recent development in mathematical optimization solvers 🧩.

What a wonderful era to be alive!

If you haven’t read about PDLP, start here to understand its implications for solving large-scale LPs using GPUs. Then, continue here to see its performance against Gurobi. If you’re more interested, read the original blog post and paper from Google.

I understand you haven’t been isolated from the world lately, have you? 🌍

NVIDIA hasn’t.

They’re testing PDLP.

They integrated PDLP into their cuOpt library and found interesting things. Want to read it? Go here!

Almost a year ago, two University of Chicago researchers tested the potential of PDLP using GPUs through the Julia programming language.

Their main conclusion was awesome:

The numerical experiments show that the prototype GPU implementation cuPDLP.jl can match the performance of commercial solvers like Gurobi and outperform them on large instances. This reveals the potential of using GPU to develop high-performance optimization solvers.

Read the full paper here.

My ✌🏻 two cents. Should traditional solvers like Gurobi embed PDLP in their workflow?

Read more

By far.

And I didn't even intend to post today because I didn’t find enough time during this week.

But I couldn’t miss the opportunity to say that Google Research just released a blog post of its new Linear Programming solver.

It’s not just pure research. They also implemented it inside Google OR-Tools.

I need to carefully read the paper to extract more conclusions, but it seems promising.

Here is the blog post, and thus the link to the paper:

Scaling up linear programming with PDLP

In which situations would you use such a solver?

Do you want me to analyze it and go deeper in another post?

Just let me know!

Have a nice day,

Borja.

PS: this is kind of the middle step between what I told you last week and what I wanted to tell you today (and will be for the next post).

Despite their incredible capabilities, there are still many challenges to overcome.

Optimizers for Deep Learning models still consume a lot -I mean, A LOT- of memory. Researchers around the world investigate more efficient algorithms, with Quantum Computing opening new frontiers. And let's not forget Alan Turing's warning about the limits of computation, given in the 1930s, which still holds true today.

Are we going to break those limits? Will we find super-efficient algorithms that make our lives easier? Will AI soon run on small devices with limited memory?

Let’s dive in! 🪂

Read more

Last weeks have been pretty intense in Operations Research.

The field is evolving rapidly nowadays, I can see a lot of movement inside it. Like everyone now is aware of Operations Research and want to get a piece of the cake.

Usually, academia has dominated the field. It creates innovation, it solves complex problems, and everything stays inside. And thus OR is not widely known even though it impacts business revenues more than any other technology.

But things are changing.

Now, academia creates the innovation in the form of algorithms, and that innovation is transferred to society in the form of companies that build powerful tools. Those companies create the building blocks to actually solve complex business problems in real life with the help of experts in the field.

This is a virtuous cycle since the more innovation, the better algorithms, that creates diverse and better building tools, that contribute to solve problems in very different industries, each with their own details… Which ends up on academia creating more innovation for those specific details.

Today you are going to see such a cycle, but as I love starting from the industry, you’ll see the cycle starting from that point.

In fact, I’m going to start with LLMs and I’m going to finish with LLMs too.

Maybe you wonder why LLMs in this context?

Let’s dive in! 🪂

Read more

4 weekends in a row without any post.

And for a very good reason: I became father for the first time on 30th May, and I’ve been pretty busy learning how to manage my new life.

But I didn’t stop paying attention to the Operations Research community, so here I am with all the things that I did read for the past few weeks:

• A new promising alternative to Adam, the go-to optimizer in Deep Learning.

• Google optimizing shipping routes with 13% reduction in fuel usage and costs.

• A conference in a place where I spent 3 months of my life that united some of the top worldwide researchers around heuristics.

• A hackathon from Shell, and $100 million in funds for power grid optimization lead by Bezos Earth Fund.

• Aaaaand… The long awaited Python version of Timefold, one of the best open-source heuristics solvers.

Let’s go for it! 🪂

Read more

And then I realized that everything I have been thinking and have told you about on occasion (like in the previous post) is truly materializing.

Operations Research is gradually making its way into our lives to the point that it will be inevitable to have systems that, by default, have at least a little bit of this area of knowledge.

What would happen if the mobile phone network went down? Or if a very important Internet node stopped providing service? What if autonomous cars took a long time to understand the objects in an image?

It would be a disaster, honestly. So today on Feasible, I tell you:

• The importance of the minimum cut problem in real life.

• Why Google really nailed it with this.

• The 3 steps they follow to solve the problem.

Come on, one more Sunday… Let’s get to it!

✂️ The Importance of the Minimum Cut Problem in Real Life

The minimum cut problem is a classic problem in network design, graph theory, and combinatorial optimization.

The ultimate goal is to find the smallest subset of edges - or nodes - that, if removed from the graph, would disconnect it.

And you might ask what the point is of solving this problem. In general, it can be used to find critical points of failure or bottlenecks in a network of nodes, but let me tell you about more specific cases:

Read more