ML Foundations and Mathematics
Foundational ML mathematics, probability, optimization, information theory, and core algorithms.
32%
Best tweets about Machine Learning
50 Best Tweets About Machine Learning (2026)
Find the best tweets about machine learning, from models and datasets to training, evaluation, research papers, MLOps, and production systems.
Technical machine learning research, training, evaluation, data, engineering, deployment, and lessons from production.
- Creators
- 44
- Updated
Original Xholic analysis
What 50 top Machine Learning posts reveal
The supplied ML conversation spans learning resources and foundational explainers alongside training, retrieval, production systems, evaluation, and inference efficiency. A smaller set of posts describes automated research and experimentation workflows, while others emphasize reliability, debugging, and deployment constraints. Learning-oriented posts make up all five supplied engagement-score outliers.
- Dominant tone
- Positive
- Median score
- 71.5
- Leading format
- Announcement
- Recent posts
- 38%
62% of posts
All-time engagement
30% of posts
Published in 90 days
Conversation map
The themes creators return to
ML Learning Resources
Courses, books, roadmaps, repositories, and interview-oriented resources for learning ML.
32%
Model Training and Experimentation
Training workflows, fine-tuning, post-training, experiment management, and decentralized or automated training.
28%
LLMs and Agents Engineering and Research and Automation of M
LLM architectures, transformers, language-model construction, RAG, prompting, and agent applications.
26%
ML Systems and Production
Production ML systems, MLOps, deployment, observability, governance, and engineering complexity.
26%
ML Data and Retrieval
Data collection, datasets, preprocessing, embeddings, vector databases, retrieval, and data infrastructure.
14%
Evaluation and Reliability
Evaluation, calibration, validation, benchmarks, reproducibility, and reliability monitoring.
12%
Inference and Efficiency
Inference serving, latency, model compression, hardware acceleration, edge deployment, and efficient runtimes.
12%
Tone and stance
Sentiment Positive leads
Author posture Supportive leads
Performance benchmark
- Median likes
- 115
- Median reposts
- 17
- Median replies
- 7
- Median views
- 8.9K
Posts with media make up 72% of this collection. Their median all-time score is 80.0, compared with 16.6 for text-only posts.
Format mix
- Announcement 30% · score 69.9
- List 24% · score 132.2
- Tutorial 22% · score 72.3
- Opinion 18% · score 8.37
Consensus and debate
Where creators agree, and where they do not
Shared view
Foundations and systems learning resources recur
The supplied posts pair ML foundations—such as probability and prerequisite maps—with structured curricula, systems material, and reading lists.
Shared view
Deployment extends beyond model training
Posts describe production ML work in terms of retrieval and vector data, inference latency and model size, infrastructure, observability, evaluation, and controls such as guardrails and rate limits.
Shared view
Automation targets experimentation workflows
Several posts describe automated experimentation loops that run tasks, measure results, analyze failures, make or propose changes, and retain human approval or oversight for consequential changes.
Open debate
Competing views of the next constraint
One post argues that industry attention and spending are extending to data infrastructure, deployment, inference, governance, and verification. Another presents a robotics view that architecture choices and simulation may help address limited-data constraints.
Open debate
Automation claims alongside reliability concerns
A post reports gains for a recursive-self-improvement ML engineering agent, while other posts argue that reliable AI needs stronger integration of learning with reasoning and values, and that complex models can be harder to understand and debug.
Open debate
Hardware speedups do not remove deployment trade-offs
One post recounts a 200× speedup after moving an SVM implementation from CPU to GPU. Another stresses that production deployments still face latency, memory, and compression-accuracy trade-offs.
What performs
Patterns behind standout posts
Learning-oriented posts account for the supplied score outliers
The five supplied score outliers are learning-oriented posts: an ML-systems curriculum, a book list, a Bayes explainer, an ML knowledge graph, and an MIT course list.
Media posts had the higher supplied median score
Media appeared in 36 of 50 tweets (72%). The supplied median all-time score was 80.03 for media posts and 16.57 for text-only posts.
Lists led the supplied format medians
Lists had the highest supplied format median all-time score, at 132.17, compared with 72.332 for tutorials and 69.888 for announcements.
Statistical standouts
- View standout post 1 Score 1906.7 · 26.67× median
- View standout post 2 Score 1240.3 · 17.35× median
- View standout post 3 Score 1144.1 · 16.01× median
- View standout post 4 Score 685.3 · 9.59× median
- View standout post 5 Score 508.0 · 7.11× median
Creator landscape
Who shapes this conversation
The five most represented creators account for 20% of the selected posts.
-
1. Vaishnavi
@_vmlops
2 posts
-
2. Aurimas Griciūnas
@Aurimas_Gr
2 posts
-
3. BURKOV
@burkov
2 posts
-
4. Towards Data Science
@TDataScience
2 posts
-
5. Tech with Mak
@techNmak
2 posts
-
6. Tivadar Danka
@TivadarDanka
2 posts
Aurimas Griciūnas covers retrieval and compression trade-offs
Aurimas Griciūnas’s posts explain vector-database write and read flows, then frame compression as a trade-off among latency, model size, and evaluation of accuracy effects.
Burkov’s posts promote compact ML learning materials
Posts from Burkov’s account present the Hundred-Page ML and Language Models books as concise resources covering mathematical ML concepts and hands-on language-model implementation.
Tech with Mak connects probability and ML systems
Tech with Mak’s two cited posts combine a Bayes’ theorem explainer with a post about a public ML-systems curriculum covering architecture, data pipelines, production, MLOps, edge AI, and privacy.
How this analysis was made
Themes, sentiment, stance, and post format are classified per tweet. All counts, shares, medians, creator concentration, freshness, and performance comparisons are then calculated directly from the published snapshot.
Xholic's all-time score compares engagement while accounting for reach, post age, and creator consistency. It is used for relative comparisons within this collection.
This report analyzes the exact 50-post snapshot shown below. AI identifies editorial categories and drafts explanations; all statistics are calculated from the snapshot, and every narrative claim is checked against cited posts before publication.
Best Machine Learning tweets
Top Machine Learning tweets from 44 creators
Ranked 01–50
- 01
@techNmak ·
Harvard made its Senior Engineer roadmap available to the public at no cost. Stop paying for $2,000 bootcamps. Prof. Vijay Janapa Reddi just put the entire ML Systems (CS249r) curriculum on GitHub. If you master these 6 pillars, you're ahead of 99% of the field: 🏛️ Architecture 🚿 Data Pipelines 🚢 Production 🛠️ MLOps 🔋 Edge AI 🔒 Privacy This is the "Black Box" of Big Tech infrastructure, open-sourced. Read. Learn. Bookmark.
- 02
@Zachly ·
You only need to read four books to truly get what’s going on in ML and data engineering: - Fundamentals of Data Engineering by Joe Reis - Designing Data Intensive Applications by Martin Kleppmann - AI engineering by Chip Huyen - Designing Machine Learning Systems by Chip Huyen If you read these four technical books and then read these four books on leadership and soft skills, you’ll be well on your way to massive success! - Radical Candor - Atomic Habits - How to Win Friends and Influence People - The Body Keeps Score What books would you recommend?
- 03
@techNmak ·
Most engineers have seen this formula. P(A|B) = P(B|A) × P(A) / P(B) Almost none can explain what it actually does. Here's Bayes' Theorem in plain English, and where it's hiding inside systems you use every day. The core idea in one sentence: Bayes' Theorem updates your belief about something after seeing new evidence. That's it. Four terms: Prior → what you believed before the evidence Likelihood → how probable the evidence is, given your hypothesis Evidence → how common the evidence is overall Posterior → your updated belief after seeing the evidence A concrete example: Say 40% of all emails are spam (your prior). You see a new email containing the word "lottery." 10% of spam emails contain "lottery." Only 1% of legitimate emails do. Plug into Bayes: P(spam | "lottery") = (0.10 × 0.40) / P("lottery") ≈ 87% The word "lottery" updated your belief from 40% → 87%. That's Bayes in action. Prior belief + new evidence = updated belief. Where it lives in AI: 1/ Spam filters The Naive Bayes classifier, the algorithm behind most spam filters - applies this exact calculation word by word across an entire email. Each word shifts the probability up or down. It's called "naive" because it assumes each word is independent of the others, which isn't realistic, but works remarkably well in practice. 2/ Medical diagnosis AI A patient has symptom X. What's the probability of disease Y? Bayes updates the base rate (how common the disease is) with the likelihood of seeing that symptom in patients who have it. Same formula, different domain. 3/ Your LLM's uncertainty Modern language models don't just predict the next token, they assign a probability to every possible token. The sampling process (temperature, top-p) is directly working with those probability distributions. Bayesian reasoning is embedded in every response your model generates. The insight most engineers miss: Bayes doesn't give you certainty. It gives you a rational way to update uncertainty. That's exactly why it's foundational to AI - real-world systems are never certain. They're always working with incomplete, noisy, probabilistic information. Every model that learns from data is, at its core, doing some version of this: Start with a belief. See evidence. Update the belief. That's Bayes. That's machine learning.
- 04
- 05
- 06
@SakanaAILabs ·
The AI Scientist: Towards Fully Automated AI Research, Now Published in Nature Nature: https://t.co/nNfpSV5e5I Blog: https://t.co/i6h8LVQOdl When we first introduced The AI Scientist, we shared an ambitious vision of an agent powered by foundation models capable of executing the entire machine learning research lifecycle. From inventing ideas and writing code to executing experiments and drafting the manuscript, the system demonstrated that end-to-end automation of the scientific process is possible. Soon after, we shared a historic update: the improved AI Scientist-v2 produced the first fully AI-generated paper to pass a rigorous human peer-review process. Today, we are happy to announce that “The AI Scientist: Towards Fully Automated AI Research,” our paper describing all of this work, along with fresh new insights, has been published in @Nature! This Nature publication consolidates these milestones and details the underlying foundation model orchestration. It also introduces our Automated Reviewer, which matches human review judgments and actually exceeds standard inter-human agreement. Crucially, by using this reviewer to grade papers generated by different foundation models, we discovered a clear scaling law of science. As the underlying foundation models improve, the quality of the generated scientific papers increases correspondingly. This implies that as compute costs decrease and model capabilities continue to exponentially increase, future versions of The AI Scientist will be substantially more capable. Building upon our previous open-source releases (https://t.co/H1tBT14Yx8), this open-access Nature publication comprehensively details our system's architecture, outlines several new scaling results, and discusses the promise and challenges of AI-generated science. This substantial milestone is the result of a close and fruitful collaboration between researchers at Sakana AI, the University of British Columbia (UBC) and the Vector Institute, and the University of Oxford. Congrats to the team! @_chris_lu_ @cong_ml @RobertTLange @_yutaroyamada @shengranhu @j_foerst @hardmaru @jeffclune
Watch video
- 07
@the_slai ·
I'm excited to join @SpaceX and @xAI to build X modeling! For the last few years I've strived to build ML systems for massive-scale user engagement and intelligent thinking; having worked on YouTube's Community Discovery team to help creators and audiences connect, then co-founding Lemma Research to help AI tackle complex legal reasoning, I realized the best way to make the most impact was to join forces with the xAI team. I look forward to pushing the boundaries of how AI understands and interacts with real-time information and conversations. My DMs are open if you're exceptional and want to build the future of AI with us.
- 08
@sentient_agency ·
10 BOOKS SERIOUS AI RESEARCHERS ACTUALLY RECOMMEND (NOT THE ONES EVERYONE POSTS) Every AI reading list says the same five names. The people actually building these systems read deeper than that. Here's the shelf they point to when nobody's performing for an audience. 1. Probability Theory: The Logic of Science - E.T. Jaynes The book researchers quietly call life-changing. Jaynes reframes probability not as gambling odds but as the mathematics of reasoning under uncertainty, which is exactly what every modern model is doing. Dense, opinionated, and the closest thing the field has to a sacred text. Almost nobody outside the work has heard of it. 2. Information Theory, Inference, and Learning Algorithms - David MacKay The book that unites information theory and machine learning in one place, written by a Cambridge physicist who made it genuinely fun. Free online, full of puzzles, and on the shelf of nearly every researcher who actually understands why their models compress and predict the way they do. 3. Reinforcement Learning: An Introduction - Sutton and Barto The foundation under everything from AlphaGo to how modern models get fine-tuned with human feedback. Researchers don't recommend it because it's trendy. They recommend it because the ideas in it keep turning out to be the ideas that matter, decades later. Also free. 4. The Book of Why - Judea Pearl A Turing Award winner's argument that today's AI is stuck because it confuses correlation with causation, and a map of what real reasoning would require. The book that names the exact ceiling current systems keep hitting. Researchers cite it constantly. The public reads past it. 5. Vision - David Marr A neuroscientist's framework for how any system, brain or machine, processes information, written before deep learning existed and somehow predicting the questions it would face. The "levels of analysis" idea in here quietly shapes how serious people think about what a model is even doing. 6. Gödel, Escher, Bach - Douglas Hofstadter The cult book about how meaning and selfhood emerge from systems following simple rules. It won a Pulitzer and then got name-dropped to death, but almost nobody finishes it. The ones who do think differently about intelligence forever. The real one, not the summary. 7. Metaphors We Live By - Lakoff and Johnson The argument that human thought runs on metaphor, not cold logic, and that you can't build a mind on first-order logic alone. Researchers working on why language models grasp meaning the strange way they do keep circling back to this one. A genuine left-field pick. 8. The Society of Mind - Marvin Minsky One of AI's founding figures arguing that intelligence isn't one thing, it's a swarm of dumb little processes working together. Written as hundreds of one-page ideas. Out of fashion for years, now looking prophetic in the age of multi-agent systems. Insider catnip. 9. How to Solve It - George Pólya A 1945 book on mathematical problem-solving that quietly shaped how a generation of researchers think about breaking down hard problems, and that keeps surfacing in papers on how to make models reason. The bridge between human heuristics and machine reasoning. 10. The Mathematical Theory of Communication - Claude Shannon The original paper that invented information theory and, with it, the entire conceptual ground that machine learning stands on. Short, brutal, and foundational. Researchers revere Shannon the way physicists revere Newton. Most reading lists skip the source and quote the descendants. The popular books tell you what AI might do. These tell you how the people building it actually think. The difference is the whole point.
- 09
@Aurimas_Gr ·
Fundamentals of a 𝗩𝗲𝗰𝘁𝗼𝗿 𝗗𝗮𝘁𝗮𝗯𝗮𝘀𝗲. With the rise of GenAI, Vector Databases skyrocketed in popularity. The truth - Vector Databases are also useful outside of a Large Language Model context. When it comes to Machine Learning, we often deal with Vector Embeddings. Vector Databases were created to perform specifically well when working with them: ➡️ Storing. ➡️ Updating. ➡️ Retrieving. When we talk about retrieval, we refer to retrieving set of vectors that are most similar to a query in a form of a vector that is embedded in the same Latent space. This retrieval procedure is called Approximate Nearest Neighbour (ANN) search. A query here could be in a form of an object like an image for which we would like to find similar images. Or it could be a question for which we want to retrieve relevant context that could later be transformed into an answer via a LLM. Let’s look into how one would interact with a Vector Database: 𝗪𝗿𝗶𝘁𝗶𝗻𝗴/𝗨𝗽𝗱𝗮𝘁𝗶𝗻𝗴 𝗗𝗮𝘁𝗮. 1. Choose a ML model to be used to generate Vector Embeddings. 2. Embed any type of information: text, images, audio, tabular. Choice of ML model used for embedding will depend on the type of data. 3. Get a Vector representation of your data by running it through the Embedding Model. 4. Store additional metadata together with the Vector Embedding. This data would later be used to pre-filter or post-filter ANN search results. 5. Vector DB indexes Vector Embedding and metadata separately. There are multiple methods that can be used for creating vector indexes, some of them: Random Projection, Product Quantization, Locality-sensitive Hashing. 6. Vector data is stored together with indexes for Vector Embeddings and metadata connected to the Embedded objects. 𝗥𝗲𝗮𝗱𝗶𝗻𝗴 𝗗𝗮𝘁𝗮. 7. A query to be executed against a Vector Database will usually consist of two parts: ➡️ Data that will be used for ANN search. e.g. an image for which you want to find similar ones. ➡️ Metadata query to exclude Vectors that hold specific qualities known beforehand. E.g. given that you are looking for similar images of apartments - exclude apartments in a specific location. 8. You execute Metadata Query against the metadata index. It could be done before or after the ANN search procedure. 9. You embed the data into the Latent space with the same model that was used for writing the data to the Vector DB. 10. ANN search procedure is applied and a set of Vector embeddings are retrieved. Popular similarity measures for ANN search include: Cosine Similarity, Euclidean Distance, Dot Product. How are you using Vector DBs? Let me know in the comment section!
- 10
@shekhu04 ·
Meet Devendra Singh Chaplot (He is teaching machines to see, think and move through the real world) > Born in Rajasthan > All India Rank 25 in IIT JEE 2010 > Same year ranked 5th in International Mathematics Olympiad in the entire world. > https://t.co/TMjo4uhgj1 from IIT Bombay, 2014 > PhD in Machine Learning from Carnegie Mellon University > Built Arnold, an AI that learned to play Doom just by watching the screen > No instructions. Pure visual learning. > Won the Visual Doom AI Competition 2017. > Won CVPR 2019 PointNav Challenge > Won CVPR 2020 ObjectNav Challenge > Won NeurIPS 2022 Rearrangement Habitat Challenge > Three of the most competitive AI benchmarks in the world. All three. > Joined Meta AI as Research Scientist after PhD > Co-founded Mistral AI, one of Europe's most powerful AI companies > In March 2026 joined SpaceX and xAI A boy from Rajasthan who ranked 5th in the world at mathematics is now building superintelligence with Elon Musk. He never chased the spotlight. The spotlight found him.
- 11
@burkov ·
The Hundred-Page Language Models Book by Andriy Burkov is well regarded, and for a specific niche: readers who want to actually build a language model, not just read about one. Why it's good: - Density without fluff. True to the "hundred-page" branding, it moves fast through n-grams → RNNs → Transformers → LLM finetuning/prompting without padding. If you already know general ML but haven't sat down and coded a Transformer, this closes that gap efficiently. - Hands-on code. All examples run in PyTorch on Google Colab, so you're not just reading math — you build three different language model architectures yourself, including a Transformer from scratch. That's the book's real differentiator versus most "intro to LLMs" material. Endorsements from figures like Vint Cerf and Tomáš Mikolov (author of word2vec) call it clear and a solid starting point for language modeling, and reviewers like the CEOs of Weaviate, Qdrant, and LlamaIndex praise its concision and clarity for understanding how LLMs work under the hood. Burkov's prior book (The Hundred-Page Machine Learning Book) has a strong track record and is used as a university textbook, so there's real precedent for his teaching style working. Bottom line: if you want a fast, code-first path to genuinely understanding and building Transformer-based LLMs, and you're comfortable with Python and some math, it's a strong choice.
- 12
@_vmlops ·
A guy landed offers from Google, LinkedIn, Snap, Coupang, and StitchFix during his ML interview run. That kind of insight usually comes with a price tag. He wrote it all down and put it on GitHub for free instead That repo now has 12.4k stars, and it's basically the closest thing to a "cheat sheet" for ML interviews that actually works, because it's based on real questions he was asked, not guesses. Here's what's inside: → a study plan that tells you exactly what to focus on, so you're not wasting weeks on stuff that never comes up → leetcode and SQL practice, including the specific things interviewers keep asking (like window functions and join types) → stats and probability questions taken straight from real interviews → AB testing basics, since almost every company asks about this now → classic ML and deep learning concepts explained simply → actual system design examples, like how to design a recommendation system or a fraud detection pipeline → a reading list of papers from people like Andrew Ng and Yoshua Bengio → an FAQ section answering the questions everyone secretly wonders about, like "do I really need to solve LeetCode Hard" or "how much cloud stuff do they actually ask" It's not trying to teach you everything about machine learning. It's trying to teach you what gets asked, which honestly matters more when you're prepping under a deadline If you're getting ready for an ML or data science interview, this is worth a bookmark
- 13
@SchmidhuberAI ·
Everybody is talking about recursive self-improvement (RSI) and meta learning. Here is my old 2020 talk about this [1]. It has aged well. Example: humans still define the starts & ends of trials of many modern meta learners. My RSI systems since 1994 LEARN to (re)define them [2]! [1] Meta Learning Machines in a Single Lifelong Trial (talk for workshops at ICML 2020 and NeurIPS 2021, based on earlier talks since 1994). Abstract: the most widely used machine learning algorithms were designed by humans and thus are hindered by our cognitive biases and limitations. Can we also construct meta learning algorithms that can learn better learning algorithms so that our self-improving AIs have no limits other than those inherited from computability and physics? This question has been a main driver of my research since I wrote a thesis on it in 1987 [2]. Here I summarize our work on meta reinforcement learning with self-modifying policies in a single lifelong trial (since 1994), and mathematically optimal meta-learning through the self-referential Gödel Machine (since 2003). Many additional publications on meta-learning since 1987 can be found in the RSI overview [2]. [2] J. Schmidhuber (AI Blog, 2020-2025). 1/3 century anniversary of first publication on recursive self-improvement (RSI) and meta learning machines that learn to learn (1987). For its cover I drew a robot that bootstraps itself. 1992-: gradient descent-based neural meta learning. 1994-: meta reinforcement learning with self-modifying policies. 1997: meta RL plus artificial curiosity and intrinsic motivation. 2002-: asymptotically optimal meta learning for curriculum learning. 2003-: mathematically optimal Gödel Machine. 2020-: new stuff!
Watch video
- 14
@pauliusztin_ ·
Microsoft just open-sourced one of the most interesting agent engineering projects I've seen this year... → https://t.co/5NyVSw2ATi Most engineers assume improving an AI agent requires: Better models More data More fine-tuning SkillOpt takes a completely different approach. Instead of optimizing the model... It optimizes the skill. Think of a skill as a standard operation procedure (SOP) for an AI agent. A small document that describes: How to solve a task Which tools to use Which steps to follow How to format outputs What good behavior looks like Traditionally, these skill files are written once and then forgotten. SkillOpt turns them into something trainable. The idea is surprisingly simple: 1. Run the agent on a batch of tasks 2. Measure performance 3. Let a second model analyze failures 4. Propose edits to the skill document 5. Keep only the edits that improve validation performance The loop starts looking surprisingly similar to machine learning. ... except you're training the skill rather than the weights. We've spent years treating skills, prompts, and agent instructions as static artifacts. SkillOpt treats them as something that can improve continuously. The model stays the same. The skill evolves. Learn more here: https://t.co/5NyVSw2ATi
Watch video
- 15
@Hesamation ·
KL divergence is a fundamental concept used in machine learning, from optimization a neural net to RL training of LLMs. but here is what it actually means:
Watch video
- 16
- 17
@aiwithjainam ·
ONE GUY WROTE THE ENTIRE STANFORD AI CURRICULUM INTO FREE NOTES AND PUT IT ON THE OPEN WEB it's called https://t.co/vosmzCUXzV and you just open the tab and the whole field is sitting there. i went in looking for one explainer on attention. that's it. one thing. an hour later i was still scrolling. transformers. mixture-of-experts. RAG. diffusion models. agents. reinforcement learning. flash attention. context engineering. every concept that's been melting your brain on the timeline for two years, written out clean, one at a time. then i found the course notes and actually said out loud "no way." > Stanford CS229, the machine learning class people pay tuition for, fully noted > CS230 deep learning, CS231n vision, CS224n NLP, all of it > the Coursera deep learning and NLP specializations, distilled > plus python, pytorch, numpy, the math, the backprop derivations done by hand the guy who made it is Aman Chadha. he's an actual applied scientist. this is not a content farm. there are no popups. no "enter your email." no $49 course at the bottom. you read it and you leave. he has a primer breaking down DeepSeek-R1. one on Claude 4. one on Kimi K2. the stuff that's barely months old is already in there, explained like a patient TA who actually wants you to get it. and the wildest part to me is the math. a Stanford-grade AI education, the exact notes, free, sitting at a url. and most people will keep paying $2k for a worse version on a course platform. no signup. no paywall. the whole thing is just yours to read. the internet is healing.
- 18
@Suhail ·
/goal for AI model training runs is *so* good - it really feels like the future. Very little babysitting now. Mine: Launch a full training run on 4 nodes. Continuously record things in an experiment document if it exists. Log hyper params, configs, periodic evals, performance insights, analyze training stability, and important changes for future analysis and reproducibility. Fix any major bugs you encounter while you monitor training but do not change the fundamental nature of the experiment without asking. If it crashes, resume again and keep training. Resume from latest reliable checkpoint you have. Reach <num> steps
- 19
- 20
@goyalshaliniuk ·
The AI Ecosystem: Essential Concepts & Model Development 1. AI Tools & Frameworks Includes workflow tools, model training platforms, vector databases, and AI-powered DevOps. 2. Computer Vision Covers image recognition, face detection, medical imaging AI, and 3D vision applications. 3. Natural Language Processing (NLP) Encompasses sentiment analysis, text summarization, retrieval-augmented generation, and speech-to-text. 4. AI Scalability & Deployment Focuses on cloud AI, serverless AI, model monitoring, and chatbot integration. 5. Deep Learning & Neural Networks Explores GANs, reinforcement learning, self-supervised learning, and federated learning. 6. Machine Learning & Model Optimization Includes feature engineering, model evaluation, hyperparameter tuning, and AI bias mitigation. 7. AI Fundamentals Covers data preprocessing, probabilistic AI, decision-making models, and explainability in AI. Explore more in the image below.
Watch video
- 21
- 22
@TechWithKhushi ·
10 GitHub repos that will level up your AI Agent skills (SAVE THIS)🔖 1. Hands-On Large Language Models Complete code notebooks from basics to advanced fine-tuning. 🔗 https://t.co/QT837OIBdw 2. AI Agents for Beginners A free 11-part intro course to build your first agents. 🔗 https://t.co/saJBGtl2eo 3. GenAI Agents Tutorials and code for building generative AI agents. 🔗 https://t.co/0g4pl4U3E0 4. Made with ML Learn to design, build, and deploy real ML apps. 🔗 https://t.co/kq43OoYr0G 5. Prompt Engineering Guide Learn to write powerful and effective prompts. 🔗 https://t.co/plIDcji8pS 6. Hands-On AI Engineering Practical LLM-powered apps and agent examples. 🔗 https://t.co/pmyvVLpIGF 7. Awesome Generative AI Guide Curated hub for genAI research and tools. 🔗 https://t.co/hG93PX13cT 8. Designing Machine Learning Systems Summaries and resources from the popular ML systems book. 🔗 https://t.co/DuzN3uWyZA 9. ML for Beginners (Microsoft) Free beginner-friendly ML curriculum. 🔗 https://t.co/6NMx4TlQxl 10. LLM Course Roadmaps and hands-on notebooks to build LLM apps. 🔗 https://t.co/WNRQ6pOcvP
- 23
@MIT_CSAIL ·
A free throwback MIT course breaking down how machine learning techniques can be applied to healthcare: https://t.co/TrQlckLh8o (v/@MITOCW) Here, MIT prof. & CSAIL principal investigator David Sontag discusses how AI can help sort thru medical data (Lecture 1).
Watch video
- 24
@andy_ai0 ·
1-month playbook to start learning AI In just one month, you’ll already be able to: - understand what AI, machine learning, deep learning actually mean - use beginner AI tools without feeling lost - understand core concepts like training data, models, overfitting - build a few small AI projects - decide which direction to go next Week 1: Build the foundation What to learn: - what AI is - AI vs machine learning vs deep learning vs generative AI - what a model is - what training means - supervised vs unsupervised learning - classification vs regression - training data, labels, and features - overfitting at a basic level - what neural networks are at a high level - what prompts, tokens, and hallucinations are What to do: - write short notes in your own words for the key concepts - take 3-5 real-life examples and identify inputs and outputs - test a few prompts in ChatGPT or Claude and observe where AI is useful and where it struggles Week 2: Start using AI hands-on What to learn: - Python basics - variables, lists, dictionaries, loops, functions - JSON basics - what APIs are - how AI tools receive input and return output - beginner prompt writing - how to give clear instructions - how to ask for structured answers What to do: - set up Python, Jupyter Notebook, or Google Colab - run simple Python exercises - make your first AI prompt experiments more structured - build tiny practice tasks such as: 1. text summarizer 2. bullet-point generator 3. study-note explainer Week 3: Learn the real ML basics underneath What to learn: - how machine learning models learn from data - loss and error at a simple level - train / validation / test split - accuracy, precision, and recall - why accuracy can be misleading - linear models and logistic regression at a beginner level - what embeddings are - semantic similarity - transfer learning - prompting vs retrieval vs fine-tuning What to do: - study one beginner dataset - understand what the input columns are and what the model predicts - compare simple examples of classification and prediction - make a short notebook or notes page explaining: 1. what data goes in 2. what the model tries to predict 3. how performance is measured 4. what can go wrong Week 4: Build small projects and choose your direction What to learn: - how to turn AI knowledge into simple projects - how to define input, process, output, and limitations - how to test outputs - how to improve prompts or workflows - the difference between learning AI for: 1. building apps 2. machine learning 3. deep learning 4. productivity/work use What to do: - choose 1-2 small projects and finish them - test them on real examples - write a short README or notes page for each project - reflect on which direction feels most interesting Everything here is best learned through practice For every concept you learn, try to: - explain it in simple words - test it with an example - build something tiny with it And from there, it becomes much easier to keep going without feeling overwhelmed
- 25
@Aurimas_Gr ·
Understanding and being able to apply 𝗠𝗮𝗰𝗵𝗶𝗻𝗲 𝗟𝗲𝗮𝗿𝗻𝗶𝗻𝗴 𝗺𝗼𝗱𝗲𝗹 𝗖𝗼𝗺𝗽𝗿𝗲𝘀𝘀𝗶𝗼𝗻 will distinguish you as a standout 𝗔𝗜 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿. Here is why 👇 Small Language Models will be the cornerstone of modern Agentic Systems. When you deploy ML models to production you need to take into account several operational metrics that are in general not ML related: 👉 𝗜𝗻𝗳𝗲𝗿𝗲𝗻𝗰𝗲 𝗟𝗮𝘁𝗲𝗻𝗰𝘆: How long does it take for your Model to compute inference result and return it. 👉 𝗠𝗼𝗱𝗲𝗹 𝗦𝗶𝘇𝗲: How much memory does your model occupy when it’s loaded for serving inference results. Both of these are important when considering operational performance and feasibility of your model deployment. 👉 Large models might not fit on a device if you are considering edge deployments. 👉 Latency of retrieving inference results might make business case non feasible. E.g. Recommendation Engines require latencies in milliseconds as ranking has to be applied as the user browses your website or app in real time. Bad news for LLM fans! 👉 … You can influence both latency and size by applying different Model Compression methods, some of them are: ➡️ 𝗣𝗿𝘂𝗻𝗶𝗻𝗴: this method is mostly used in tree-based and Neural Network algorithms. In tree-based ones we prune leaves or branches from decision trees. In Neural Networks we remove nodes and synapses (weights) while trying to retain ML performance metrics. ✅ In both cases the output is a reduction in the number of Model Parameters and model size. ➡️ 𝗞𝗻𝗼𝘄𝗹𝗲𝗱𝗴𝗲 𝗗𝗶𝘀𝘁𝗶𝗹𝗹𝗮𝘁𝗶𝗼𝗻: this type of compression is achieved by: 👉 Training an original large model which is called the Teacher model. 👉 Training a smaller model to mimic the Teacher model by transferring knowledge from it, this model is called the Student model. Knowledge in this context can be extracted from the outputs, internal hidden state (feature representations) or a combination of both. 👉 We then use the “Student” model in production. ➡️ 𝗤𝘂𝗮𝗻𝘁𝗶𝘇𝗮𝘁𝗶𝗼𝗻: a most commonly used method that doesn’t have much to do with Machine Learning. This approach uses fewer bits to represent model parameters. 👉 You can apply quantization techniques both during the training and after the models has been already trained. 👉 In regular Neural Networks what is quantized are Model Weights, Biases and Activation Functions. 👉 Most usual quantization is from float to integer (32 bits to 8 bits. There are talks about 1 bit LLMs nowadays 😅). ➡️ … ❗️ While the above methods do reduce the size of the models, allowing them to be deployed in production scenarios, there is almost always a reduction in accuracy so be careful and evaluate it accordingly. What methods for reducing model size have you used? What were the main challenges? Let me know in the comment section!
- 26
@mark_k ·
Recursive Self Improvement is here! Frontis-MA1 (35B) is a new AI4AI agent trained for recursive self-improvement in machine learning engineering. Full OpenMLE stack released. Model post-trained on four operators: Draft → Improve → Debug → Crossover. Learning and evolution close in one loop with real execution feedback. MLE-Bench Lite (12h/task, single 4090 @ 12GB): Base 39.4% → +Evo 60.6% → +Evo-Max 71.2% Beats GPT-5.5 + Codex. Gains transfer to held-out NatureBench. Weights + code fully open.
- 27
@JesusMartinez ·
Anthropic spent billions training Claude. @DistStateAndMe spent $2-3 million and got comparable results with 70 strangers on the internet. No data center. No corporate backing. Just a Bittensor subnet. 18 months ago, people said this was impossible. Sam Dare is the founder of @covenant_ai and the man behind @tplr_ai, Subnet 3 on Bittensor. On March 10, his team completed the largest decentralized AI training run in history. • 72 billion parameters • ~1.1 trillion tokens • 70+ independent contributors • Commodity internet. No InfiniBand. No whitelist • 67.1 MMLU score. Outperforms Meta's LLaMA-2-70B on multiple benchmarks The model is fully open source under Apache 2.0. Sam is not an ML researcher. He's a blockchain guy. Jake from @opentensor taught him machine learning in a month. With GPT-3 and a dream, he built one of the most significant AI achievements in crypto history. 9 months ago, Templar trained a 1.2 billion parameter model. Now it's 72 billion. He told me Cursor's latest Composer 2 model was trained using technology that references Covenant's published research. A crypto project's work being used by one of the most popular AI coding tools in the world. Jensen Huang mentioned Templar on the All-In Podcast. TAO surged 30%. Grayscale filed for a TAO ETF. DCG built Yuma specifically for the Bittensor ecosystem. But Sam doesn't care about any of that right now. Tether is moving into decentralized training. The Goliath of crypto. And Sam says he doesn't have time to celebrate because he's trying to stay ahead. Next up: an 8 billion parameter domain-specific model trained end to end on decentralized rails. New heterogeneous SparseLoCo algorithm that lets any GPU join. Maybe even a MacBook. His target? One trillion parameters by end of year. He's not slowing down. The man is not playing.
- 28
- 29
@alvarobartt ·
💥 Learn how to build your own tool-calling agent with @huggingface TRL + @Alibaba_Qwen Qwen3.5 on @Azure Machine Learning! - @NousResearch hermes-function-calling-v1, 500 single-turn samples - SFT with TRL on Qwen3.5 2B (released today!) on a single NVIDIA H100 - Everything on Azure, from Container Registry to Machine Learning! Step-by-step in the thread 🧵
- 30
- 31
@TheAITimeline ·
🚨This week's top AI/ML research papers: - Composer 2 Technical Report - LeWorldModel - Claudini - Intern-S1-Pro - Self-Distillation of Hidden Layers for Self-Supervised Representation Learning - Natural-Language Agent Harnesses - Why Does Self-Distillation (Sometimes) Degrade the Reasoning Capability of LLMs? overview for each + authors' explanations read this in thread mode for the best experience
- 32
@shushant_l ·
I'm shocked most people use AI every day without knowing how AI actually learns. Here's the complete AI model training pipeline explained in one simple infographic. --- 📂 AI Model Training ┃ ┣ 📂 AI Training Basics ┃ ┣ 📂 Pattern Recognition ┃ ┣ 📂 Predictions ┃ ┣ 📂 Error Measurement ┃ ┣ 📂 Model Improvement ┃ ┗ 📂 Billions of Iterations ┃ ┣ 📂 Preparation ┃ ┣ 📂 Define The Goal ┃ ┣ 📂 Collect Data ┃ ┣ 📂 Clean The Data ┃ ┣ 📂 Remove Duplicates ┃ ┗ 📂 Filter Low Quality Data ┃ ┣ 📂 Data Processing ┃ ┣ 📂 Tokenization ┃ ┣ 📂 Convert Text To Tokens ┃ ┣ 📂 Numerical Representation ┃ ┣ 📂 Input Formatting ┃ ┗ 📂 Training Ready Data ┃ ┣ 📂 Model Architecture ┃ ┣ 📂 Choose Neural Network ┃ ┣ 📂 Transformer Architecture ┃ ┣ 📂 Initialize Model ┃ ┣ 📂 Random Weights ┃ ┗ 📂 Parameter Setup ┃ ┣ 📂 Pretraining ┃ ┣ 📂 Predict Missing Tokens ┃ ┣ 📂 Learn Language ┃ ┣ 📂 Learn Facts ┃ ┣ 📂 Learn Reasoning ┃ ┗ 📂 Discover Patterns ┃ ┣ 📂 Learning Cycle ┃ ┣ 📂 Calculate Loss ┃ ┣ 📂 Compare Predictions ┃ ┣ 📂 Backpropagation ┃ ┣ 📂 Update Weights ┃ ┗ 📂 Optimization ┃ ┣ 📂 Validation ┃ ┣ 📂 Test On Unseen Data ┃ ┣ 📂 Measure Accuracy ┃ ┣ 📂 Check Reasoning ┃ ┣ 📂 Evaluate Safety ┃ ┗ 📂 Test Generalization ┃ ┣ 📂 Post Training ┃ ┣ 📂 Supervised Fine Tuning ┃ ┣ 📂 Preference Optimization ┃ ┣ 📂 Safety Alignment ┃ ┣ 📂 Improve Helpfulness ┃ ┗ 📂 Improve Reliability ┃ ┣ 📂 Safety ┃ ┣ 📂 Reject Harmful Requests ┃ ┣ 📂 Protect Privacy ┃ ┣ 📂 Reduce Unsafe Outputs ┃ ┣ 📂 Follow Safety Policies ┃ ┗ 📂 Responsible Responses ┃ ┣ 📂 Evaluation ┃ ┣ 📂 Math ┃ ┣ 📂 Coding ┃ ┣ 📂 Science ┃ ┣ 📂 Language ┃ ┣ 📂 Logic ┃ ┗ 📂 Real World Tasks ┃ ┣ 📂 Deployment ┃ ┣ 📂 Optimize Inference ┃ ┣ 📂 Efficient Serving ┃ ┣ 📂 Production Release ┃ ┣ 📂 User Access ┃ ┗ 📂 Scalable Infrastructure ┃ ┣ 📂 Continuous Improvement ┃ ┣ 📂 Collect Feedback ┃ ┣ 📂 Fix Weaknesses ┃ ┣ 📂 Improve Data ┃ ┣ 📂 Enhance Safety ┃ ┗ 📂 Release Updated Models ┃ ┣ 📂 End To End Pipeline ┃ ┣ 📂 Goal ┃ ┣ 📂 Data Collection ┃ ┣ 📂 Data Cleaning ┃ ┣ 📂 Tokenization ┃ ┣ 📂 Model Initialization ┃ ┣ 📂 Pretraining ┃ ┣ 📂 Loss Calculation ┃ ┣ 📂 Backpropagation ┃ ┣ 📂 Optimization ┃ ┣ 📂 Validation ┃ ┣ 📂 Post Training ┃ ┣ 📂 Safety Alignment ┃ ┣ 📂 Evaluation ┃ ┣ 📂 Deployment ┃ ┗ 📂 Continuous Improvement ┃ ┗ 📂 Key Takeaways ┣ 📂 AI Learns Patterns ┣ 📂 Data Quality Matters ┣ 📂 Pretraining Builds Knowledge ┣ 📂 Post Training Improves Safety ┣ 📂 Massive Compute Is Required ┗ 📂 Learning Never Truly Stops
- 33
@toly ·
MetaTimer: Using Large Language Models for Precise, Prompt-Aware Inference Latency Prediction The rapid proliferation of large language models (LLMs) in production systems has exposed a fundamental limitation: inference latency varies dramatically across prompts due to differences in semantic complexity, required reasoning depth, output length, and generation dynamics. Conventional prediction methods—ranging from token-count heuristics and hardware Roofline models to traditional machine-learning regressors—fail to generalize because they cannot capture these prompt-specific nuances. Accurate a priori estimation of processing time is essential for resource scheduling, dynamic batching, cost forecasting, service-level guarantees, and user-experience enhancements. We introduce MetaTimer, the first framework to repurpose a lightweight LLM itself as a high-precision meta-predictor capable of forecasting the exact wall-clock inference duration required by any target LLM for an arbitrary input prompt. A compact 8B-parameter model is fine-tuned on a massive corpus of millions of prompt–execution pairs collected across heterogeneous model families (GPT-4-class, Llama 3.1, Claude, Mistral), quantization levels, decoding strategies, and hardware accelerators. The predictor employs chain-of-thought reasoning to decompose prompt semantics, estimate output token distributions and reasoning trajectories, and integrate model- and hardware-specific performance profiles, yielding fine-grained predictions for Time-to-First-Token (TTFT), Time-Per-Output-Token (TPOT), and total latency. Extensive evaluations on held-out benchmarks spanning reasoning, creative writing, coding, and long-context tasks demonstrate state-of-the-art accuracy: a mean absolute percentage error (MAPE) of 6.3% for end-to-end latency—representing a >40% reduction in mean squared error relative to the strongest Roofline–ML baselines—and strong zero-shot generalization to unseen models and platforms. When integrated into production serving stacks (vLLM, TensorRT-LLM, Triton), MetaTimer delivers up to 31% gains in resource utilization and tail-latency reduction. These results establish that LLMs possess emergent capabilities for computational self-modeling, opening a new paradigm for self-aware, adaptive, and energy-efficient generative AI infrastructure. We publicly release the predictor model, dataset, and serving plugins to accelerate research in meta-performance modeling for frontier AI systems.
- 34
@eric_seufert ·
I'm happy to share some of the research that I've been working on today: As ad platforms become more opaque and automated end-to-end, advertisers are left with few levers of control over campaign performance. I wanted to interrogate an idea: could advertisers treat "black box" platforms as teacher models and use a behavioral distillation process, focusing on multimodal creative and ad context, to predict ROAS performance for ad instances? And if so, what kind of model could best express the interactions between those features? Most machine learning research focused on digital advertising is published by the largest platforms themselves. I wanted to address this asymmetry and explore what advertisers could build independently to improve performance within this increasingly automated environment. That idea ultimately became DeCANT: a Deep Creative Attention-based Network for pre-Testing. The model architecture uses self- and cross-attention to condition the semantic interpretation of a creative on the environment in which it is deployed. The model architecture supports multimodal ad creative and predicts ROAS on a context-conditioned basis. Operationally, a model like DeCANT can fit into a pre-testing regime as an automated filter: creatives are produced through a generative pipeline, and the student model that learns the process for the given advertising channel is invoked on the proposed creative and context. The model produces an expected ROAS, which is compared against the advertiser's testing threshold to determine whether the creative is uploaded to the platform. https://t.co/TgElrkFWwI
- 35
@KanikaBK ·
Just stumbled upon this data: 92% of data science jobs require stats and ML skills. 69% list Machine Learning specifically. CAMBRIDGE just made their 417-page Math for Machine Learning book 100% FREE. No signup is required . Just the full PDF. This is the actual math that sits underneath every ML model you use. I was just going through the content: ↳ Linear algebra and matrix operations ↳ Analytic geometry and vector spaces ↳ Probability and statistics from scratch ↳ Optimization methods including gradient descent ↳ Dimensionality reduction and PCA ↳ Regression, density estimation, classification The engineers who build the next generation of AI systems will not just know the tools. They will understand the math the tools run on.
- 36
- 37
@IamEmily2050 ·
The interview between the two legends, Jeff and Bill, was one hour long, so I used the NotebookLM video overview to capture the key details. In a collaborative discussion, Google's Jeff Dean and Nvidia's Bill Dally examine the rapid evolution of machine learning and its future hardware requirements. They highlight the transition from simple task based models to autonomous agents capable of executing long term, complex workflows. To support these advances, the experts emphasise the need for low latency inference and innovative chip architectures that minimise data movement to conserve energy. The conversation also explores how AI driven design is currently accelerating the creation of more efficient semiconductors at both companies. Finally, they reflect on the profound societal benefits of these technologies, particularly through the potential for personalised healthcare and individualised educational tutors.
Watch video
- 38
@burkov ·
The Hundred-Page Machine Learning Book by Andriy Burkov has a strong reputation in the ML community. Here's the general consensus: - Concise but mathematically substantive — it compresses the math (linear algebra, calculus, probability) rather than skipping it - Covers core ML topics seriously: supervised/unsupervised learning, SVMs, neural networks, ensembles, gradient descent, feature engineering, hyperparameter tuning - Endorsed by Peter Norvig, Aurélien Géron, and Gareth James (who wrote the foreword) - "Read first, buy later" model — you can read it free online - Works as both an introduction and a long-term reference you keep coming back to The book is best for: - The "determined amateur" with a good understanding of high-school maths, or anyone comfortable with calculus, statistics, probability, and vectors/matrices - People who want signal-dense material and don't mind re-reading paragraphs - Practitioners wanting a compact reference across the ML landscape - ML interview prep Burkov also wrote two follow-ups, The Machine Learning Engineering Book, which focuses more on the practical/production side, and the The Hundred-Page Language Models Book, which focus on the the evolution of language models from count-based, to recurrent NNs, to transformers. #LMtrainingData
- 39
@panditdhamdhere ·
If you're building AI applications in Rust, these are five of the strongest libraries to learn by the end of 2026. 🦀 → Burn - Deep learning - native Rust framework with training, inference, GPU acceleration (CUDA, WGPU), autodiff, modular design. Great alternative to PyTorch for Rust developers. → Candle - LLM - inference - Lightweight framework from Hugging Face. Excellent for running Llama, Mistral, Phi, Qwen, Gemma, and other transformer models efficiently in Rust. → tch-rs- PyTorch bindings- Rust bindings for LibTorch. Ideal if you're migrating PyTorch code to Rust or want access to the PyTorch ecosystem. → Ort (ONNX Runtime) - Production inference- High-performance ONNX Runtime bindings. Deploy models exported from TensorFlow, PyTorch, or scikit-learn with excellent speed. SmartCore - Classical machine learning- Pure Rust machine learning library supporting regression, classification, clustering, PCA, KNN, Random Forests, SVMs, and more. Great for non-deep-learning tasks. If your goal is AI agent development. A modern Rust AI stack. LLM Framework → Rig Local Model → Candle Tokenization → tokenizers Embeddings → FastEmbed Vector Database → Qdrant or LanceDB API → Axum Async Runtime → Tokio
- 40
@ForwardFuture ·
“I rewrote my code in 30 minutes — and it ran 200× faster.” @ctnzr VP, Applied Deep Learning Research @NVIDIA on the moment GPUs changed everything: “NVIDIA showed up in our lab and said, ‘You should try CUDA.’ I plugged in a GPU, rewrote my SVM training code, and it ran 200× faster than my CPU version.” “I thought, that’s it. This is dramatically easier and clearly the future of machine learning.” “The vision was simple: accelerate the world’s most important computations by 10× or 100×, and use that to power AI.” “The compute required for intelligence is essentially unbounded.”
Watch video
- 41
@GaryMarcus ·
This was right five years ago, and still is: “Large scale pretrained models are certainly likely to figure prominently in artificial intelligence for the near future, and play an important role in commercial AI for some time to come. The results that have been achieved with them are certainly intriguing and it is worthwhile pursuing them. But it is unwise to assume that these techniques will suffice for AI in general. It may be an effective short-term research strategy to focus on the immediate challenges that seem to be surmountable, but focusing on the surmountable may not get us to what is most necessary: a firm foundation for reliably integrating statistics and machine learning with reasoning, knowledge, common sense and human values.” Gary Marcus and Ernest Davis, 2021, “Has AI found a new Foundation?” at https://t.co/9VqBTg0FtK
- 42
@glenngabe ·
Some interesting nuggets from the latest Search Off The Record Podcast about Google using AI in Search rankings. Here is Google's Nikola Todorovic, Director of Software Engineering at Google Search, about how AI is used to impact rankings and why simpler linear systems are sometimes easier to debug than more complex AI systems: Nikola: "The reason it's not so easy to apply AI everywhere (in Search) is because the models function like a black box. You don't always understand what's happening underneath. It's a complex set of neural networks. The linear models are the easiest ones to understand and debug, because it's not like you can just put your AI or ML system into search and reap the most benefit from your side by side experiments." "Then you will get to something and launch it, but you will have problems with that as well because maybe the systems evolved, the searches evolve, and so on. And then you will need to debug this and replace it (at some point). And this kind of replacement and changes is complicated. So the more you can understand how these things work, which signals you are using, which signals are important for relevance, for quality, for the safety of the results, (the easier it is to debug). So you do need to understand the system and the more complex the AI or the ML systems, the more challenging it is." https://t.co/Pf5stUq4HZ
- 43
@atulit_gaur ·
i think the fusion of physics and ai is the most beautiful one we train neural networks with optimization methods rooted in physics, diffusion models borrow ideas from thermodynamics, physics informed neural networks solve differential equations by embedding physical laws directly into learning, energy based models treat learning as the search for low energy states, statistical mechanics has shaped the mathematics behind entire classes of machine learning algorithms etc the universe spent 13.8 billion years writing its laws and today we are teaching machines to think by borrowing those very laws
- 44
@ttunguz ·
That little black box in the middle is machine learning code. I remember reading Google’s 2015 Hidden Technical Debt in ML paper & thinking how little of a machine learning application was actual machine learning. The vast majority was infrastructure, data management, & operational complexity. With the dawn of AI, it seemed large language models would subsume these boxes. The promise was simplicity : drop in an LLM & watch it handle everything from customer service to code generation. No more complex pipelines or brittle integrations. But in building internal applications, we’ve observed a similar dynamic with AI. Agents need lots of context, like a human : how is the CRM structured, what do we enter into each field - but input is expensive the Hungry, Hungry AI model. Reducing cost means writing deterministic software to replace the reasoning of AI. For example, automating email management means writing tools to create Asana tasks & update the CRM. As the number of tools increases beyond ten or fifteen tools, tool calling no longer works. Time to spin up a classical machine learning model to select tools. Then there’s watching the system with observability, evaluating whether it’s performant, & routing to the right model. In addition, there’s a whole category of software around making sure the AI does what it’s supposed to. Guardrails prevent inappropriate responses. Rate limiting stops costs from spiraling out of control when a system goes haywire. Information retrieval (RAG - retrieval augmented generation) is essential for any production system. In my email app, I use a LanceDB vector database to find all emails from a particular sender & match their tone. There are other techniques for knowledge management around graph RAG & specialized vector databases. More recently, memory has become much more important. The command line interfaces for AI tools save conversation history as markdown files. When I publish charts, I want the Theory Ventures caption at the bottom right, a particular font, colors, & styles. Those are now all saved within .gemini or .claude files in a series of cascading directories. The original simplicity of large language models has been subsumed by enterprise-grade production complexity. This isn’t identical to the previous generation of machine learning systems, but it follows a clear parallel. What appeared to be a simple “AI magic box” turns out to be an iceberg, with most of the engineering work hidden beneath the surface. https://t.co/7uehMMWucf
- 45
- 46
@DylanFeltus ·
curious about building your own model? the actual roadmap: step 1: understand transformers → Karpathy "Let's build GPT" https://t.co/QRC54Uo5es → "Attention Is All You Need" paper https://t.co/JEN33a8Bxw → HF LLM course https://t.co/PbZUWmJbWz → https://t.co/UGTmB0L8Z6 step 2: fine-tune a model (fastest way to "your own") → Unsloth https://t.co/LwOpqmmPz2 — runs on consumer hardware → Axolotl https://t.co/3Go38MIxxT → HF TRL https://t.co/4Wtg4RIQHO — SFT, DPO, GRPO step 3: train from scratch → nanochat https://t.co/ub1AvqQw77 — GPT-2 for ~$48 → LitGPT https://t.co/LajhHXxZnP → datasets: FineWeb, RedPajama, The Pile (all on HF) 📚 more resources: → "Deep Learning" by Goodfellow, Bengio & Courville — the textbook https://t.co/Zrr0jwiALu → "Build a Large Language Model (From Scratch)" by Sebastian Raschka — hands-on, code-first https://t.co/3OeGbGCycv → "Designing Machine Learning Systems" by Chip Huyen — the production/systems side → Karpathy's "Neural Networks: Zero to Hero" playlist https://t.co/iKLe8rtDso → 3Blue1Brown neural networks series https://t.co/knxwlyF9eq
- 47
@Meta_Engineers ·
Our Ranking Engineer Agent (REA) autonomously executes key steps across the end-to-end machine learning lifecycle for ads ranking models. REA reduces the need for manual intervention, managing asynchronous workflows spanning days to weeks through a hibernate-and-wake mechanism, with human oversight at key strategic decision points. Read this post that covers REA’s ML experimentation capabilities: autonomously generating hypotheses, launching training jobs, debugging failures, and iterating on results: https://t.co/W5AcATUFXX
- 48
@_theshash ·
"How can we make the best model possible with the limited data we have? A lot of those folks are converging around liquid neural networks."- @brezshares The robotics industry has a data problem. Everyone knows it. The default answer is collect more data. But Brian flagged a group of people taking the opposite approach: instead of scaling data, make the model architecture work with less. Liquid neural networks update their weights in real time instead of fixing them after training. That means continuous learning in production, the model improves as it operates, not just when you retrain it. The implication: reinforcement learning scales with compute, while data collection scales with people. One of those is easier to throw money at. @castorhat validated the core insight: "It's easy to say let's just go scale the data. It's a lot more nuanced than that. You need to scale data at a pace that matches the financial resources of your model training partners. You can't give them a billion hours of video and make them train on it. They don't have the money to buy the compute." Physical Intelligence and Skild may have $1 billion, but they still only have 30 people on the model team. Each one is a person with a life and a family. You can't just say here's a billion hours, figure it out. @abdulalali added the world model angle: simulations could bridge the data gap by generating unlimited training scenarios through compute rather than physical collection. A hybrid deployment, world models for data augmentation alongside foundation models for generalization, might be the path that actually works. Bayley's historical parallel: "We're mirroring early computer vision. Small models, heavy fine-tuning, product-driven iteration. The foundation-scale backbones come later. Right now you've raised $5 million. The answer is definitely not train on a billion hours of video. It's train something modestly sized with tangible performance benefits so you can talk to your VCs and raise." The teams that solve the model architecture problem might leapfrog the ones trying to solve the data problem. Watch the full Robotics Livestream Ep. 1 on YouTube: https://t.co/PfWgSichbI
- 49
- 50
@TDataScience ·
"Machine Learning systems rarely fail in a single moment. Their performance changes gradually as data distributions shift, calibration drifts, or new patterns emerge in the environment." Gal Arav shares a thorough, accessible introduction to survival analysis for data drift and ML reliability. https://t.co/oXh4lwrRk8
Best Tweets by Topic