urko retweeted
I gave a talk to physicists on the potted history from LLMs to agent swarms, covering:
> scaling laws
> reasoning models
> coding agents & harnesses
> autoresearch and idea basins
> adding communication to steer / train swarms
Slides on HF: huggingface.co/spaces/lewtun…
urko retweeted
Microsoft just killed backpropagation.
They published a paper where an LLM agent writes neural net weights directly in Python.
No backpropagation. No gradient descent. No pretrained weights.
Instead, they had an AI coding agent literally hand-write the neural network’s weights directly in Python.
For over a decade, AI has had one fatal flaw in science: the black box problem.
Deep learning models can predict how the human brain responds to images, speech, and stories. But nobody knows how they do it.
Prediction without explanation isn't science.
So the researchers flipped the entire paradigm.
They called it Interpretable Weight-Space Programming (iWSP).
Instead of throwing massive compute at a black box and trying to reverse-engineer it later, they forced an AI agent to work like a human scientist.
The agent writes an explicit mathematical hypothesis directly into model weights. Tests it against a fixed evaluator. Sees the score. Rewrites the code. Over and over.
First, they tested it on 13 algorithmic benchmarks, sorting, multiplication, edit distance, greatest common divisor.
The agent hit 100% accuracy on every single one.
And it compressed the models by 29×, down to a median of just 166 parameters.
The networks were so clean and transparent that summarizing a model in under 200 words allowed a completely fresh AI to rebuild it from text alone with 100% accuracy.
Then they pointed it at the human brain.
In the visual cortex, the hand-written model beat every existing interpretable baseline on human fMRI data, and generalized to unseen people and unseen images.
In auditory brain recordings (ECoG), its hand-crafted acoustic features captured brain signals that OpenAI's Whisper-large-v3 completely missed, improving predictions across 126 out of 166 electrodes.
Then came the moment that proved why black boxes are dangerous.
On a language fMRI dataset, the agent’s score suddenly exploded in a single step—jumping from 0.067 to 0.104 and crushing GPT-2 XL.
With a black box, researchers would have celebrated and published.
Because iWSP writes readable code, they looked inside.
They caught the AI exploiting a hidden flaw in the brain-scan dataset, tracking scan start and end drift instead of the actual meaning of the story.
Once they trimmed the corrupted scan edges, the agent still built transparent semantic features that matched GPT-2 XL.
For ten years, we assumed intelligence required billions of unreadable parameters tuned in the dark.
Turns out, you don't need to crack open the black box if you stop building black boxes in the first place.
urko retweeted
Yann LeCun just did it again.
Their new paper proves world models (JEPAs) can finally reason about counterfactuals, what WOULD have happened if you’d taken a different action.
And it requires zero retraining. Zero extra parameters. Zero new data.
Right now, every AI world model has a massive blind spot. It can only predict what happens on average.
Imagine steering a boat while an invisible underwater current pushes you off course.
A normal AI world model can tell you where the boat ends up on average.
What it cannot answer is the question that actually matters:
"What would have happened if I had steered left instead, in that exact same hidden current?"
That single question is what separates bad luck from a bad decision.
The researchers proved something brutal. Standard world models trained on factual data can never answer it on their own.
Not with 10x more data. Not with 100x bigger models.
Because data only shows what happened, never what would have happened under the same invisible forces.
Then they found the loophole.
Physics has a simple rule: nature makes no leaps. Over a short step, your action and a hidden force act separately.
That means the actual outcome you just observed already contains the hidden current baked inside it.
You don't need to guess the current. You just reuse the drift.
Take what actually happened. Add the difference between the AI's prediction for the new action and the action you took.
That’s it.
Zero retraining. Zero extra parameters. Zero counterfactual training data.
Just two passes of the exact same JEPA world model and basic addition.
And results are ridiculous:
• Across four physics simulators from raw pixels, counterfactual accuracy ($ R^2$) jumped from 38–46% to 86–96%, using the exact same models.
• On an underwater vehicle test, it reached 90–94%.
• In hindsight planning, it found the winning action 35–69% of the time, crushing an oracle planner with full access to the simulator (14–23%).
For years, AI has been stuck predicting the average future.
It couldn't look back at a mistake and truly understand what went wrong.
Humans don't just predict what happens next. We rewind reality and ask "what if."
Now world models can too.
urko retweeted
A single artificial neuron does surprisingly little: a weighted sum, then one squashing function.
Three signals come in, each scaled by its own weight. The cell body adds them up with a bias, and the function decides what goes out the other side.
Watch the activation switch from sigmoid to tanh to ReLU. With ReLU, anything below zero simply doesn't fire.
Every number you see is worked out live as the weights drift.
urko retweeted
Por eso no tengo TikTok en El móvil.
TikTok destripado: cómo usa la IA y los sensores del móvil para que no salgas de la aplicación jamás.
Via @paco229
microsiervos.com/archivo/hac…
urko retweeted
🧠 Antes de elegir una prueba estadística, define qué quieres averiguar.
Parece obvio, pero es uno de los errores más frecuentes en el análisis de datos científicos.
¿Cuántos análisis empiezan eligiendo una prueba antes de definir el objetivo científico?
#stats #datascience
urko retweeted
Interpreting machine learning models and explaining predictions is a common need for data scientists and analysts. explainerdashboard is a Python package that provides interactive dashboards for model explanation, letting you explore feature importance, individual predictions, and model behavior in a clear visual interface.
Helpful when you want to:
✔️ Visualize global and local feature importance
✔️ Explore model predictions interactively
✔️ Create explainable dashboards without heavy frontend coding
✔️ Share model insights with stakeholders in Python
Below are sample visuals from the package documentation. More details are available here: explainerdashboard.readthedo…
If you are interested in practical tools for statistics, data science, R, and Python, you can sign up for my newsletter.
Further details: statisticsglobe.com/newslett…
#Python #RStats #statisticsclass #database #rstudioglobal #StatisticalAnalysis
urko retweeted
🎥 𝗘𝘇𝗮𝗴𝘂𝘁𝘂 𝗟𝗮𝘁𝘅𝗮 𝟯!
Atzo aurkeztu genuen; gaur, gure lankideek bideo honetan azalduko dizute zertan datzan euskarazko AAren bertsio berria.
🐑 Probatu doan!
👉 latxa.hitz.eus/
#Latxa3 #HiTZ #Euskara
urko retweeted
Principal Component Analysis (PCA) is an essential technique for reducing the dimensionality of data sets while maintaining their key characteristics.
By transforming original variables into a new set of uncorrelated principal components, PCA captures the most significant variance in the data, making analysis and visualization simpler.
Implementing PCA in R is straightforward using the prcomp() function from the stats package. This function efficiently computes the principal components and summarizes their importance. For visual insights, the ggplot2 package can be used to plot the results.
Steps to Implement PCA in R:
1️⃣ Load Data: Import your data set with read.csv() or read.table().
2️⃣ Perform PCA: Use prcomp() to calculate the principal components.
3️⃣ Standardize Data: Enable standardization directly in prcomp() by setting scale. = TRUE.
4️⃣ Analyze Results: Review the variance captured by each component with summary() and visualize using ggplot2.
For a comprehensive, step-by-step guide to performing PCA in R, including practical examples, check out my tutorials created in collaboration with Paula Villasante Soriano & @Cansu_SG.
Article: statisticsglobe.com/principa…
Video: youtube.com/watch?v=mNpBrHwO…
Furthermore, I have developed an extensive online course on PCA, covering both the theoretical concepts and practical application in R programming.
Check out this link for more details: statisticsglobe.com/online-c…
#programming #tidyverse #DataViz #pythonlearning #Python #DataScientist #Rpackage #DataAnalytics #database #datavis #RStats #ggplot2
urko retweeted
Today we're releasing the message board used to create PyBot!
With 1,000 agents, 184k messages, and 32M words, it's the largest open dataset to study swarm dynamics.
huggingface.co/spaces/Huggin…
PyBot: No weights. No training.
A chatbot built with only simple rules and no AI, entirely written in 600k lines of Python code. PyBot matches Smollm with 360M parameters!
How? A swarm of 1,000 DeepSeek Flash agents wrote the whole code in just a week.
github.com/huggingface/pybot
urko retweeted
Bayes’ Rule is one of the most important concepts in probability and machine learning.
But how does it actually work in Python?
This free PDF, Bayes' Rule with Python: A Tutorial Introduction to Bayesian Analysis, helps you understand Bayesian thinking through practical Python examples.
What you'll learn:
Understanding Bayes' Theorem
Conditional probability
Prior and posterior probabilities
Updating beliefs using new evidence
Applying Bayesian analysis with Python
A useful resource for anyone interested in Python, statistics, data science, and machine learning.
Get the FREE PDF here: clcoding.com/2026/08/bayes-r…
urko retweeted
Choosing between R and Python for statistical analysis involves understanding their specific strengths and popularity in different contexts. Python is a versatile programming language, widely favored for its general-purpose capabilities and extensive use in data science and software development. In contrast, R has a strong foothold in the field of statistics, providing specialized tools and packages designed for statistical computing and graphics.
The visualization of this post illustrates Google search trends over the last five years, comparing interest in R and Python for programming and statistics. It shows that while Python generally maintains higher search interest overall, R leads when focusing on statistical topics.
Here’s a comparison of both languages for statistical tasks:
R:
🔹 Highly specialized in statistical analysis and data visualization.
🔹 Offers a wide range of packages specifically for statistical tests, modeling, and graphics (e.g., dplyr, ggplot2, stats).
🔹 Preferred for academic and research-focused projects in statistics.
Python:
🔹 Known for its versatility in data science and general-purpose programming.
🔹 Integrates well with other technologies and supports a broader range of applications, from machine learning to web development.
🔹 Utilizes powerful libraries for data manipulation and visualization (e.g., pandas, matplotlib, seaborn, scipy, statsmodels).
Generally speaking, the choice between R and Python often comes down to personal preference. Both languages have their unique strengths and can be effective for statistical tasks. However, it's also important to consider the ability to communicate and collaborate with peers. In the field of statistics, R is widely used and recognized, making it a valuable tool for ensuring clear communication and understanding within teams. For this reason, I would always choose R for statistical tasks.
If you want to dive deeper into statistical methods in R, check out my online course. Learn more: statisticsglobe.com/online-c…
#programmer #R4DS #datavis #Python3 #RStats #Python
urko retweeted
🚀 𝗟𝗮𝘁𝘅𝗮 𝟯 𝗵𝗲𝗺𝗲𝗻 𝗱𝗮!
@Hitz_zentroa-k garatutako euskarazko AAk Interneten bilatzen du, irudiak ulertzen ditu, 21 hizkuntzatan itzultzen du eta hobeto arrazoitzen du.
Irekia eta doakoa. Probatu!
👉 latxa.hitz.eus/
#Latxa3 #HiTZ #Euskara
urko retweeted
Microsoft is making moves again.
A quiet little Python tool just shot to the top of GitHub’s trending charts.
100,000+ stars.
It’s called MarkItDown.
And it does something deceptively simple:
It turns almost any file into clean Markdown.
PDFs. Word docs. PowerPoints. Excel files. Images.
Drop a file in. Get structured Markdown out.
Sounds small.
It’s not.
Because one of the biggest bottlenecks in AI workflows — especially RAG systems — is getting messy, real-world documents into a format models can actually use.
And real-world documents are brutal.
PDFs are chaotic.
Word docs are full of hidden formatting junk.
PowerPoints are messy and often image-heavy.
Spreadsheets can be a nightmare to parse cleanly.
That’s where this gets interesting.
MarkItDown strips away the friction and gives you something LLM pipelines can actually work with.
In other words: less preprocessing, less pain, faster AI implementation.
Even better, this isn’t some random side project.
It’s an official Microsoft open-source tool.
Free. Commercially usable. Practical.
I tested it on a 200-page PDF.
A few seconds later, I had Markdown that was shockingly clean.
And that’s what big tech does at its best:
They take an annoying, universal problem that everyone has been duct-taping together…
and turn it into a simple standard.
That’s why this matters.
It’s not just a file conversion tool.
It’s infrastructure for the next wave of AI applications.
Get it here: github.com/microsoft/markitd…
===
🚨 Want to learn how to build + ship AI and Data Science projects (that businesses actually want in 2026)?
On December 2nd, I am hosting a free workshop to help you get started with AI + DS projects in Python.
Register here (500 seats): learn.business-science.io/ai…
urko retweeted
New Harvard paper finds that 94.8% of a coding agent's tokens were re-reads of cached context, i.e. 94.8% of the text the AI processed was earlier conversation it was re-reading, not new work.
And that re-reading raises its estimated energy use about 25 times, from roughly 10 to 250 kilowatt-hours over 10 weeks.
Energy estimates for LLMs usually price single request-and-response calls. Agents re-read a growing context on every call and hand work to subagents, which those estimates can miss.
The authors logged 10 weeks of real Claude Code use by 1 engineer, about 1 billion tokens across 42 sessions. They built an energy estimator that charges cached tokens for attention reads but not full reprocessing.
About 260 cached tokens were read per generated token, and context grew 5.3× over a session. The cache-aware estimate came to about 250 kWh, against about 10 kWh from the output-based EcoLogits estimator.
If you track agent cost or emissions, count cached reads and subagent calls, and start new tasks in fresh sessions when possible.
urko retweeted
🆕 Posit está pasando de integrar un chatbot en el IDE a construir un entorno de trabajo donde los agentes interactúan con los datos, el código y las herramientas reales del investigador.
¿Qué ha cambiado?👇🧵
📊 耶鲁计算机博士生刘晨把自己发 Nature 子刊和 NeurIPS 的论文配图脚本全开源了,figures4papers 让你的论文图一眼看上去就有顶会的样子。
figures4papers 在 GitHub 揽获 1.1 万 star,也上了 HelloGitHub 的推荐。
以前赶论文 deadline,图往往是最后一晚用默认样式凑出来的,配色、字号、图例都不讲究,审稿人扫一眼就觉得粗糙。
这个仓库里放的是作者真实发表过的图:柱状图、雷达图、折线图、3D 球体图、趋势图,每张都配好生成它的 Python 脚本,照着换上自己的数据就能出同款。
更省事的是,figures4papers 里还附了一套科研绘图 skill,能在 Cursor、Claude Code、Codex 里引用,让 AI 编程助手按这套风格直接帮你画。
论文的门面,往往就是那几张图。
GitHub:github.com/ChenLiu-1996/figu…
urko retweeted
Google published a system that writes full research papers end to end.
They call it “PaperOrchestra”
It is a multi-agent framework that completely automates the entire lifecycle of writing scientific papers.
Not just summarizing data. Not just drafting an abstract.
The whole thing. End-to-end.
They built a fully autonomous digital laboratory.. When you input a raw idea, a swarm of specialized AI agents takes over.
One agent acts as the lead researcher. Another conducts the exhaustive literature review. Another writes the methodology.
Then, a dedicated "reviewer" agent brutally critiques the work, forcing autonomous revisions before a human ever even sees the draft.
What used to take a team of PhDs six months of grinding can now be orchestrated in the background while you sleep.
But here is the trap nobody is thinking about.
The scientific peer-review system is already drowning.
Now, anyone with a laptop can spin up a swarm of agents to generate highly plausible, perfectly formatted, end-to-end research papers at scale.
The cost of producing scientific literature just dropped to zero.
And the feedback loop is officially closing.
If AI can independently research, write, and critique AI research... the human is no longer the driver
urko retweeted
When planning a statistical study, it is important to find a balance between a large enough sample size to detect meaningful effects and the efficient use of resources. Sample size calculation using power analysis is a key tool for achieving this balance.
Power analysis allows you to determine how many observations are needed to detect an effect of interest with a given probability while controlling the significance level. It shows how quantities such as effect size, alpha level, statistical power, and sample size are related, making it an essential method for study planning.
Advantages of power analysis:
🔹 Helps determine an appropriate sample size before collecting data
🔹 Clarifies the relationship between effect size, alpha level, power, and sample size
🔹 Supports efficient use of time, budget, and other resources
🔹 Reduces the risk of underpowered studies
🔹 Makes study planning more transparent and scientifically sound
The image below illustrates these relationships. On the left, you can see the connection between the null hypothesis, the alternative hypothesis, the critical value, and statistical power. On the right, the charts show how the required sample size changes depending on the effect size. In general, smaller effects require larger samples to be detected reliably.
If you are interested in topics like this, subscribe to my newsletter to receive practical tips on statistics, data science, AI, and programming with R and Python. Check out this link for more details: statisticsglobe.com/newslett…
#statisticsclass #RStats #RStudio #pythonprogramming #Python
urko retweeted
Healthcare data is sensitive, so sending it to a cloud AI service isn't always an option.
In this tutorial, Lakshmi shows you how to build a local healthcare AI assistant with MedGemma, Ollama, and Open WebUI.
You'll run the model on your own machine, chat with a medically tuned LLM, upload medical images, and more.
freecodecamp.org/news/build-…


























