❤️ Check out Lambda here and sign up for their GPU Cloud: https://ift.tt/SKkG1fv ๐ The paper is available here: https://ift.tt/CpWBju5 Source: https://www.youtube.com/watch?v=6evUpgCHtOQ Our Patreon if you wish to support us: https://ift.tt/h3F8YNb ๐ We would like to thank our generous Patreon supporters who make Two Minute Papers possible: Adam Bridges, Benji Rabhan, B Shang, Cameron Navor, Charles Ian Norman Venn, Christian Ahlin, Eric T, Fred R, Gordon Child, Juan Benet, Michael Tedder, Owen Skarpness, Richard Sundvall, Ryan Stankye, Shawn Becker, Steef, Taras Bobrovytsky, Tazaur Sagenclaw, Tybie Fitzhugh, Ueli Gallizzi My research: https://ift.tt/schPtmj Thumbnail design: https://felicia.hu
❤️ Check out Lambda here and sign up for their GPU Cloud: https://ift.tt/1htSpPE Free course on Ray Tracing: https://ift.tt/rMhy3T6 Our Patreon if you wish to support us: https://ift.tt/Vycb6ng ๐ We would like to thank our generous Patreon supporters who make Two Minute Papers possible: Adam Bridges, Benji Rabhan, B Shang, Cameron Navor, Charles Ian Norman Venn, Christian Ahlin, Eric T, Fred R, Gordon Child, Juan Benet, Michael Tedder, Owen Skarpness, Richard Sundvall, Ryan Stankye, Shawn Becker, Steef, Taras Bobrovytsky, Tazaur Sagenclaw, Tybie Fitzhugh, Ueli Gallizzi My research: https://ift.tt/OCT0iaJ Thumbnail design: https://felicia.hu
Quick answer: DeepSeek's Engram module attacks a real gap: transformers have no native way to look up knowledge, so every fact recall burns neural compute. Engram adds O(1) static memory lookup alongside MoE, and Engram-27B beat MoE baselines at equal parameters and FLOPs.
Two Minute Papers covered a DeepSeek research release with an unusually direct pitch: fix one of the biggest structural problems in transformer models. The paper, "Conditional Memory via Scalable Lookup," proposes Engram — a memory module that gives a model something transformers never natively had: a cheap, deterministic way to look things up. Here is what the paper actually claims, and what it could mean for the hardware you run.
What problem did DeepSeek set out to fix?
The paper's abstract names it precisely: Mixture-of-Experts models scale capacity through conditional computation, but transformers lack a native primitive for knowledge lookup. In practice, every fact a model "knows" is stored in neural parameters, and recalling it costs real compute on every single token. Static patterns — the kind of thing an N-gram table handles trivially — end up reconstructed by neural layers that could otherwise spend their depth on reasoning. That is the inefficiency Engram targets: not a bigger model, but a cheaper place to keep the things a model should not have to re-derive.
How does the Engram module work?
Engram modernizes a very old idea — N-gram embeddings — into a conditional memory module with O(1) lookup. It retrieves static N-gram memory and fuses it with the model's dynamic hidden states, so recurring patterns get served from a lookup table instead of being recomputed by neural layers. Two design choices stand out. First, the addressing is deterministic, which is what makes the lookup constant-time. Second, the paper formulates a sparsity-allocation trade-off between neural computation (MoE) and static memory (Engram), identifying a U-shaped scaling law that guides how much capacity each axis should get.
O(1) via modernized N-gram embeddings with deterministic addressing
Quality (Engram-27B)
Consistent gains over MoE baselines at iso-parameter and iso-FLOPs across knowledge, reasoning, code, and math
Mechanism
Relieves early layers from static pattern reconstruction, potentially preserving effective depth for reasoning
System efficiency
Deterministic addressing lets massive embedding tables offload to host memory with minimal inference overhead
The honest caveats: the comparison is DeepSeek's own, under controlled iso-parameter and iso-FLOPs constraints, and independent replication is still ahead. The repository's demo code is explicitly a demonstration of the data flow, not a production implementation.
What does this mean for local hardware?
The system-efficiency result is the one self-hosters should watch. Because Engram's addressing is deterministic, its massive embedding tables can be offloaded to host memory with minimal inference overhead — a design that treats system RAM as part of the model rather than as a slow fallback. That could reshape which machines qualify for local runs of future DeepSeek-class models: the neural backbone still wants a real GPU, but static memory becomes cheaper to host. If you are mapping a build to these requirements, CompareAIHardware's searchable GPU database for local AI is the fastest way to filter cards by VRAM, bandwidth, and price. For where DeepSeek's model line stands today, see our write-up on how DeepSeek V4 Pro stacks up against closed models, and for the memory math on their current efficiency-focused release, our guide to what DeepSeek V4 Flash needs to run locally.
Frequently asked questions
Is Engram a replacement for Mixture-of-Experts?
No — the paper positions it as a complementary sparsity axis. MoE handles conditional computation; Engram handles static knowledge lookup. The scaling-law work is about how to split capacity between the two.
Is the Engram-27B comparison trustworthy?
It is controlled and specific: iso-parameter and iso-FLOPs constraints against MoE baselines, across knowledge, reasoning, code, and math. It is also a vendor evaluation, so independent replication is the natural next step.
Can I run Engram models locally?
The repository publishes the paper, a demo implementation of the module's data flow, and model weights under a separate model license. The demo is explicitly not a production implementation, so local serving depends on future inference-stack support.
Why does deterministic addressing matter so much?
Because constant-time lookup is what lets static memory live outside the GPU. Neural layers need fast VRAM; a lookup table with deterministic addressing can sit in host memory with minimal overhead, which changes the hardware footprint.
๐ We would like to thank our generous Patreon supporters who make Two Minute Papers possible:
Adam Bridges, Benji Rabhan, B Shang, Cameron Navor, Charles Ian Norman Venn, Christian Ahlin, Eric T, Fred R, Gordon Child, Juan Benet, Michael Tedder, Owen Skarpness, Richard Sundvall, Ryan Stankye, Shawn Becker, Steef, Taras Bobrovytsky, Tazaur Sagenclaw, Tybie Fitzhugh, Ueli Gallizzi
Disclosure: machine-learning.null.pictures and compareaihardware.com are operated by the same team. Links to compareaihardware.com are editorial recommendations, not paid placements.
❤️ Check out Lambda here and sign up for their GPU Cloud: https://ift.tt/5XgHjZF ๐ The paper is available here: https://ift.tt/J9TD6Bu Research panel I will be at GTC: https://ift.tt/t2WUkR9 Sources: https://www.youtube.com/watch?v=0aq4Wi2rsOk https://www.youtube.com/watch?v=I0yPzZp6dM0 Our Patreon if you wish to support us: https://ift.tt/sVXmzDW ๐ We would like to thank our generous Patreon supporters who make Two Minute Papers possible: Adam Bridges, Benji Rabhan, B Shang, Cameron Navor, Charles Ian Norman Venn, Christian Ahlin, Eric T, Fred R, Gordon Child, Juan Benet, Michael Tedder, Owen Skarpness, Richard Sundvall, Ryan Stankye, Shawn Becker, Steef, Taras Bobrovytsky, Tazaur Sagenclaw, Tybie Fitzhugh, Ueli Gallizzi My research: https://ift.tt/YGW59BS #nvidia
Understanding ML books in Machine Learning: getting started
Published: 2026-03-07 · Updated: August 21, 2026
Quick Answer
Start with Hands-On ML with Scikit-Learn for working code, add Python Machine Learning (3rd Ed.) for broader algorithm coverage, and finish with Deep Learning with Python for neural networks. Read one chapter per session and re-type every code example yourself; the official scikit-learn getting-started guide (scikit-learn.org) and Google's free ML crash course (developers.google.com) fill in whatever a chapter skips.
Video: Getting Started with ML Books
The video above walks through how to pick and use machine learning books when you are starting out. The sections below summarize the practical reading order and how to combine books with free official documentation.
Which machine learning books should you start with?
Beginners should start with a hands-on book rather than a theory-heavy textbook, because writing runnable code early keeps motivation high and exposes gaps immediately. The three picks below cover the usual progression: practical scikit-learn workflows first, wider algorithm coverage second, neural networks third.
Book
Focus
Best for
Hands-On ML with Scikit-Learn
Practical approach
Learners who want running code in week one
Python Machine Learning (3rd Ed)
Comprehensive ML guide
Readers who want broad algorithm coverage with Python examples
As an Amazon Associate I earn from qualifying purchases.
How do you study a machine learning book effectively?
Type out every code example instead of copying files, and run it against a small dataset you understand. The official Python tutorial (docs.python.org) covers the language features most ML books assume, and the scikit-learn getting-started page shows how the library's fit/predict pattern works across algorithms, which makes book examples easier to adapt.
Should you combine books with free courses?
Yes. Books give depth; free courses give pacing and worked exercises. Google's Machine Learning Crash Course teaches core concepts through interactive visualizations, and Kaggle's free micro-courses (kaggle.com/learn) let you practice on real datasets in the browser. A practical routine is one book chapter during the week and one course module on the weekend.
What comes after your first machine learning book?
After your first book, specialize. Neural network learners can move to a deep-learning-focused text, while practitioners tracking where the field is heading can read the annual Stanford AI Index Report (aiindex.stanford.edu) for industry and research trends. Reading landmark papers, such as the Transformer architecture paper on arXiv (arxiv.org/abs/1706.03762), becomes manageable once you know the vocabulary a good book teaches.
Frequently Asked Questions
How many ML books do I need to get started?
One practical book is enough to start. Add a second only after you can train and evaluate a model without following the book line by line.
Should I read the math chapters first?
No. Read enough linear algebra and probability to follow the current chapter, then return to theory once the code makes sense.
Are older book editions usable?
Fundamentals age slowly, but library APIs change. Check the book's repository for updated notebooks when examples throw deprecation warnings.
Machine Learning external SSDs or storage (2) Errors Explained and Solved
Published: 2026-03-07
Introduction
This comprehensive guide covers everything you need to know about Machine Learning external SSDs or storage (2) Errors Explained and Solved. We've analyzed the topic to provide actionable insights and practical recommendations. Follow along with the video tutorial above for a hands-on learning experience.
What You'll Learn
Core concepts and best practices for Machine Learning external SSDs or storage (2) Errors Explained and Solved
As an Amazon Associate I earn from qualifying purchases.
Conclusion
This guide provides a solid foundation for understanding and getting started with Machine Learning external SSDs or storage (2) Errors Explained and Solved. Watch the video tutorial above and use the recommended products to set up your workspace. Refer back to this article as you progress.
Machine Learning notebooks or writing tablets (2) Buying Guide
Published: 2026-03-07
Introduction
This comprehensive guide covers everything you need to know about Machine Learning notebooks or writing tablets (2) Buying Guide. We've analyzed the topic to provide actionable insights and practical recommendations. Follow along with the video tutorial above for a hands-on learning experience.
What You'll Learn
Core concepts and best practices for Machine Learning notebooks or writing tablets (2) Buying Guide
As an Amazon Associate I earn from qualifying purchases.
Conclusion
This guide provides a solid foundation for understanding and getting started with Machine Learning notebooks or writing tablets (2) Buying Guide. Watch the video tutorial above and use the recommended products to set up your workspace. Refer back to this article as you progress.
Machine Learning ML books (2) Tutorial: introduction
Published: 2026-03-07
Introduction
This comprehensive guide covers everything you need to know about Machine Learning ML books (2) Tutorial: introduction. We've analyzed the topic to provide actionable insights and practical recommendations. Follow along with the video tutorial above for a hands-on learning experience.
What You'll Learn
Core concepts and best practices for Machine Learning ML books (2) Tutorial: introduction
As an Amazon Associate I earn from qualifying purchases.
Conclusion
This guide provides a solid foundation for understanding and getting started with Machine Learning ML books (2) Tutorial: introduction. Watch the video tutorial above and use the recommended products to set up your workspace. Refer back to this article as you progress.
Best Practices for Machine Learning development workstations (monitors, keyboards) (2)
Published: 2026-03-07
Introduction
This comprehensive guide covers everything you need to know about Best Practices for Machine Learning development workstations (monitors, keyboards) (2). We've analyzed the topic to provide actionable insights and practical recommendations. Follow along with the video tutorial above for a hands-on learning experience.
What You'll Learn
Core concepts and best practices for Best Practices for Machine Learning development workstations (monitors, keyboards) (2)
How to choose the right tools and equipment
Step-by-step implementation guidance
Common pitfalls and how to avoid them
Recommended products for your needs
Recommended Products
Here are our top picks to get you started:
Product for development workstations (monitors, keyboards)
As an Amazon Associate I earn from qualifying purchases.
Conclusion
This guide provides a solid foundation for understanding and getting started with Best Practices for Machine Learning development workstations (monitors, keyboards) (2). Watch the video tutorial above and use the recommended products to set up your workspace. Refer back to this article as you progress.
Complete development workstations (monitors, keyboards) Installation Guide for Machine Learning
Published: 2026-03-07 · Updated: August 21, 2026
Quick Answer
A functional machine-learning development workstation needs three things: a dual-monitor setup at eye level so code and documentation sit side by side, a keyboard you can type on for long sessions without wrist strain, and a clean Python environment with scikit-learn or TensorFlow installed. You do not need exotic hardware to begin — the software setup matters more than the desk on day one.
Video: Workstation Setup Walkthrough
The video above shows a complete development workstation installation for machine-learning work. The sections below explain what the hardware is for and which software to install first.
What does a machine learning workstation actually need?
A machine-learning workstation needs enough RAM to hold your dataset, a display arrangement that lets you view code and plots simultaneously, and reliable peripherals. Classical algorithms — regression, trees, clustering — run on ordinary CPUs through libraries like scikit-learn, so a modest computer is genuinely sufficient for your first months of learning. GPU acceleration becomes relevant later, mainly for training neural networks with frameworks such as TensorFlow.
How should you set up your monitor and keyboard?
Place the primary monitor directly in front of you with the top of the screen at roughly eye level, keep the keyboard at elbow height so wrists stay straight, and position secondary screens for reference material. These adjustments cost nothing and reduce fatigue during long study sessions. Test any new keyboard for at least a week of real coding before deciding it fits your workflow.
Which software should you install first?
Install Python, then create an isolated environment for ML work with the built-in venv module (docs.python.org — venv). Inside that environment, install scikit-learn and follow its getting-started example to fit your first model, or install TensorFlow and run the official beginner quickstart (tensorflow.org) to train a small neural network. Structured free lessons from Google's crash course (developers.google.com) and Kaggle Learn (kaggle.com/learn) give that setup an immediate purpose.
Recommended Product
Product for development workstations (monitors, keyboards)
As an Amazon Associate I earn from qualifying purchases.
Do you need a GPU to start machine learning?
No. Most beginner coursework — tabular data problems, classical algorithms, small neural networks — runs acceptably on CPU. Cloud notebooks such as Kaggle's provide free GPU time when a course exercise needs it, which delays any hardware purchase until you know your actual workload.
Frequently Asked Questions
Is one monitor enough for ML work?
Yes, though two screens speed up comparing code, terminal output, and documentation side by side.
How much RAM is enough to learn ML?
Enough to hold the datasets your courses use; beginner course datasets are small, so typical modern machines cope fine.
Mechanical or membrane keyboard for coding?
Whichever keeps you comfortable over multi-hour sessions — comfort beats switch type.
Machine Learning ML books for Beginners: Complete Tutorial
Published: 2026-03-07 · Updated: August 21, 2026
Quick Answer
Yes, you can learn machine learning from books — but pair them with practice. A proven beginner sequence is: a hands-on scikit-learn book first, a comprehensive Python ML book second, a deep-learning book third. After every chapter, reproduce the examples and then solve one small exercise on your own using Kaggle's free courses (kaggle.com/learn) and the scikit-learn documentation (scikit-learn.org).
Video: ML Books for Beginners Tutorial
The video above introduces machine-learning books suited to beginners. Below you will find a suggested reading order, a comparison of the three featured books, and how to practice what they teach.
Can you learn machine learning from books alone?
You can learn the concepts from books alone, but you cannot build skill without running code. Machine learning is a practical discipline: the libraries change APIs, datasets behave unexpectedly, and debugging is half the job. Treat each book as a guided tour and treat your own experiments as the actual construction site.
What is a good reading order for beginner ML books?
Read the practical book first, the comprehensive survey second, and the neural-network book last. This order matches how the field is usually taught: classical algorithms build intuition that deep learning then extends.
As an Amazon Associate I earn from qualifying purchases.
How do you practice what the books teach?
After each chapter, redo the chapter's main example from memory, then apply it to a different dataset. Kaggle's Intro to Machine Learning course offers guided exercises with instant feedback, the TensorFlow beginner quickstart (tensorflow.org) gives you a trainable neural network in minutes, and the official Python tutorial (docs.python.org) closes any language gaps the books assume away.
How much math do you need before chapter one?
Comfort with basic algebra and descriptive statistics is enough to start most beginner books; they introduce required notation as it appears. When a formula stalls you, look the term up in Google's ML glossary (developers.google.com — Glossary) rather than pausing the book for a full math course.
Frequently Asked Questions
Paper book or ebook for learning ML?
Ebooks make searching API names easier; paper works better for margin notes. Either works — consistency matters more.
Can I skip classical ML and go straight to deep learning?
You can, but evaluation habits, overfitting intuition, and feature thinking transfer poorly without the classical foundation.
How long does a beginner ML book take?
Working through a chapter properly takes several sessions; expect weeks per book, not days.
This video provides an in-depth look at Ultimate Machine-Learning Beginner's Guide. Below, we summarize the key takeaways and supplement with our own research, including cited sources, so you can verify every recommendation yourself. We've also included links to recommended products that align with what you'll see in the video.
Quick Answer: How do I start learning machine learning in 2026?
Machine learning is a programming approach where a model learns patterns from data instead of following hand-written rules. You show the algorithm thousands of labeled examples — for instance, spam versus non-spam emails — and it learns a function that generalizes to new examples it has never seen.
In practice, machine learning is used for product recommendations, fraud detection, image classification, and language models. The field spans classical methods (linear regression, decision trees, gradient boosting) available in scikit-learn, and deep neural networks built with frameworks like TensorFlow. The Stanford AI Index Report tracks the field's growth annually and is a good reality check on what ML can and cannot do today.
What should a beginner learn first?
Python, basic statistics, and one ML course — in that order. Python is the default language of machine learning: scikit-learn, TensorFlow, and Kaggle all assume it. Basic statistics (means, distributions, correlation) is enough to start; you pick up the rest as you go.
Week 2–3: A free structured course, such as the Google ML Crash Course, which Google updated and expanded in 2024.
Week 3–4: Train your first models with scikit-learn — start with the Iris or Titanic datasets on Kaggle Learn, which offers free hands-on micro-courses in the browser.
Which free courses are worth taking?
The three most widely recommended free starting points are Google's ML Crash Course, Andrew Ng's Coursera specialization, and Kaggle Learn.
All three are referenced constantly in the ML community because they require no payment to start and use standard Python tooling.
Which books should a beginner buy?
The single most-recommended beginner book is Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow — it walks from classical ML to deep learning with runnable code. Two companion references cover statistics and deep learning in more depth.
Here's our complete list of top picks, with detailed justifications:
As an Amazon Associate I earn from qualifying purchases.
What common mistakes do ML beginners make?
The most common beginner mistakes are skipping Python basics, jumping straight to deep learning, and never evaluating models properly.
Skipping basics: scikit-learn documentation assumes Python fluency. Struggling with syntax makes every later step slower.
Deep learning too early: Most tabular-data problems are better served by gradient boosting or linear models in scikit-learn than by neural networks.
No train/test discipline: Always split your data before evaluating; scikit-learn's train_test_split exists for this.
Tutorial loops: After 2–3 guided projects, pick a Kaggle competition or your own dataset and build something without a walkthrough.
How long does it take to learn machine learning?
You can build your first working models in 2–4 weeks of part-time study; reaching job-ready competence typically takes 6–12 months of consistent practice. Andrew Ng's Coursera specialization is designed around a ~3-month pace at 10 hours per week. Kaggle Learn micro-courses each take 3–5 hours.
Frequently Asked Questions
Do I need a math degree to learn machine learning?
No. High-school algebra and basic statistics are enough to start. You only need deeper linear algebra and calculus once you study how learning algorithms work internally.
Do I need a powerful computer?
No for classical ML — scikit-learn runs on any laptop. For deep learning, free GPU time on Kaggle or Google Colab covers beginner projects.
Python or R?
Python for beginners. scikit-learn, TensorFlow, and most course material use it, and it doubles as a general programming language.
How much should I budget for learning?
Zero for the core path: Google's crash course, Kaggle Learn, and auditing the Coursera specialization are all free. A ~$50 book purchase is the only optional spend, and the picks above cover all levels.
Are there common mistakes to avoid?
Yes: don't jump to neural networks first, don't evaluate on training data, and don't collect credentials instead of building projects. A small portfolio of finished projects teaches more than ten certificates.
Should I specialize immediately?
No. Learn the general workflow (data cleaning → training → evaluation) first, then specialize — NLP, computer vision, or tabular ML — based on the problems you enjoy.
Watch: Getting Started with dev ergonomics (monitor arms, keyboards)
This video provides an in-depth look at Getting Started with dev ergonomics (monitor arms, keyboards). Below, we summarize the key takeaways and supplement with our own research, including cited sources from OSHA, Cornell University's ergonomics program, and the UK Health and Safety Executive, so you can verify every recommendation yourself. We've also included links to recommended products that align with what you'll see in the video.
Quick Answer: What's the fastest ergonomic fix for a developer desk?
Position the top of your monitor at or slightly below eye level about an arm's length (50–100 cm) away, keep your elbows near 90°, and use a keyboard that lets your wrists stay straight while typing. OSHA's computer workstations guidance and Cornell University's ergonomics research both identify monitor position and neutral wrist posture as the two highest-impact changes.
Why does ergonomics matter for developers?
Developers type for hours daily in a fixed seated posture, which makes them prone to musculoskeletal disorders (MSDs) of the neck, shoulders, wrists, and back. The UK Health and Safety Executive classifies these upper-limb and back disorders as a leading cause of work-related absence, and its musculoskeletal disorder guidance recommends risk-assessing any prolonged display-screen workstation.
Fixing setup problems early is cheap; treating an injury is not. That's the entire economic argument for a monitor arm and a split keyboard.
How high should your monitor be?
The top of the screen should sit at or slightly below eye level, roughly an arm's length (50–100 cm) from your eyes. This is the core recommendation in OSHA's Computer Workstations eTool, which also advises tilting the monitor slightly upward (10–20°) toward the eyes.
A monitor arm makes this adjustable in seconds, which matters when you share a desk, switch between sitting and standing, or use a laptop plus external display. Cornell University's Cornell University Ergonomics Web publishes step-by-step workstation setup guidance for exactly these cases.
Why use a split or ergonomic keyboard?
A split keyboard lets your wrists stay straight (neutral posture) instead of angled outward, reducing strain on the wrists and forearms during long typing sessions. OSHA's workstation guidance lists neutral wrist posture as a primary goal of keyboard placement; tented split keyboards physically enforce it.
Practical pattern from the video: keep the keyboard close so elbows stay near your sides at roughly 90°, and avoid resting wrists on a hard edge while typing.
What should you set up first on a budget?
Fix monitor position first — it's free (books work) or cheap (a basic arm) and affects neck posture all day. Keyboard upgrade second.
Laptop-on-desk as a primary setup, monitor too high or too low, and keyboard too far away are the three most common mistakes.
Laptop as primary display: its screen forces a head-down posture; dock it and add a monitor at eye level.
Monitor off-center: neck rotation all day; center the screen on your keyboard's home row.
Keyboard at the desk edge with wrists bent: OSHA's eTool flags bent wrists as a primary MSD risk factor.
No breaks: short stand/stretch breaks every 30–60 minutes matter more than any single gadget. The Human Factors and Ergonomics Society (HFES) publishes research on work–rest cycles for computer work.
Top Recommended Products
Here's our complete list of top picks, with detailed justifications:
As an Amazon Associate I earn from qualifying purchases.
Frequently Asked Questions
Are monitor arms worth it?
Yes, if you adjust your setup often or need desk space back. A good arm holds any position in seconds, which makes maintaining OSHA's recommended monitor height realistic day to day.
How much should I budget?
Expect to spend between $75 and $250 for a quality arm or ergonomic keyboard; the picks above cover entry and premium price points.
Do split keyboards have a learning curve?
Usually one to two weeks to regain full typing speed. The Freestyle2's adjustable separation lets you start narrow and widen gradually.
Can these fixes eliminate existing pain?
They reduce strain, but persistent pain warrants seeing a medical professional. Ergonomics addresses setup risk factors, not diagnosed conditions — see HSE's MSD guidance for when to escalate.
Are standing desks necessary?
No. Posture quality beats posture variety at first; get sitting posture right, then alternate sitting and standing if you can.
Should I buy used or refurbished?
Used can be risky for electronics; refurbished from certified sellers may offer savings with a warranty. We generally recommend new for peace of mind.
Watch: Getting Started with external SSDs / storage
This video provides an in-depth look at Getting Started with external SSDs / storage. Below, we summarize the key takeaways and supplement with our own research, including cited sources (USB Implementers Forum, NIST, Microsoft), so you can verify every claim yourself. We've also included links to recommended products that align with what you'll see in the video.
Quick Answer: What should a beginner look for in an external SSD?
Buy a portable SSD rated for at least 1,050 MB/s over USB 3.2 Gen 2 (10 Gb/s) — for example the Samsung T7 — and size it at 2× your current working dataset. Speed standards are set by the USB Implementers Forum: USB 3.2 Gen 2 tops out at 10 Gb/s, and Gen 2x2 at 20 Gb/s.
What is an external SSD and how is it different from a hard drive?
An external SSD stores data on NAND flash memory with no moving parts, so it survives knocks that would kill a spinning hard drive, and it reads and writes data many times faster. A portable external HDD tops out around 120–140 MB/s over USB, while mainstream portable SSDs like the Samsung T7 are rated for up to 1,050 MB/s and the T9 up to 2,000 MB/s over USB 3.2 Gen 2x2.
For machine-learning work, the practical difference is loading datasets and model checkpoints: a job that takes minutes from a hard drive takes seconds from a Gen 2 SSD.
Which USB speed standard do you actually need?
USB 3.2 Gen 2 (10 Gb/s) is the sweet spot for most users; Gen 2x2 (20 Gb/s) only helps if both your drive and your port support it.
ML datasets and model checkpoints are large — often tens to hundreds of gigabytes — and training pipelines stall on slow storage. An external SSD lets you keep datasets off your system drive and move multi-gigabyte files between machines in seconds instead of tens of minutes.
External SSDs also serve as fast scratch space for laptops with small internal drives, and as transport for work between desktop and GPU servers.
How do you safely erase an external SSD before selling it?
Use the manufacturer's erase tool or a full-disk encryption-then-format approach; simple quick-format does not remove data. NIST Special Publication 800-88 Revision 1, the US standard for media sanitization, treats cryptographic erase as an accepted method for flash storage and notes that repeated overwriting wears SSDs without guaranteeing full coverage due to wear-leveling (NIST SP 800-88r1, PDF).
Practical recipe: encrypt the whole drive (BitLocker/FileVault/LUKS), then reformat and hand it over. Without the key, the data is cryptographically unrecoverable.
Does an external SSD need maintenance?
Modern operating systems send TRIM commands to SSDs automatically; you shouldn't defragment flash storage. On Windows, the built-in Optimize Drives tool detects SSDs and performs a TRIM pass instead of a defrag, as documented by Microsoft's defrag command reference.
Day-to-day: keep ~10–20% of capacity free, avoid storing the drive at full load in heat, and eject before unplugging to protect in-flight writes.
Top Recommended Products
Here's our complete list of top picks, with detailed justifications:
As an Amazon Associate I earn from qualifying purchases.
Frequently Asked Questions
What's the most important feature to look for?
The USB interface generation. A Gen 2 (10 Gb/s) drive with 1,000+ MB/s sustained speed covers backups, datasets, and boot drives; Gen 2x2 is only worth paying for if your computer has a matching 20 Gb/s port.
How much should I budget?
Expect to spend between $75 and $250 for a quality 1TB portable SSD; 2TB and 4TB models cost proportionally more. Our list includes picks at various price points.
Are cheaper external SSDs worth considering?
Some budget models offer good value, but be cautious of extremely cheap drives that may use slower QLC flash or lack DRAM caching and fail prematurely. We've included value picks that balance cost and quality.
Can I use an external SSD with any OS?
Yes, most drives ship formatted for broad compatibility, but exFAT works across Windows, macOS, and Linux — reformat to your native filesystem (NTFS/APFS/ext4) for best performance and reliability.
How long do external SSDs last?
Typical portable SSDs carry 3–5 year warranties and are rated for hundreds of terabytes written; for archival copies, keep at least one additional backup on separate media.
Should I buy used or refurbished?
Used can be risky for electronics; refurbished from certified sellers may offer savings with a warranty. We generally recommend new for peace of mind.
For distraction-free note-taking around machine-learning work, the reMarkable 2 Starter Bundle is the safest pick: a 10.3-inch e-ink writing tablet built only for notes and sketches, sold with the Marker Plus pen. If you want Android apps on an e-ink screen, pick the BOOX Note Air 10.3; if you want color for highlighting papers and diagrams, pick the BOOX Note Air 4C with its 4,096-color display.
Watch: Best notebooks / writing tablets for Machine-Learning Enthusiasts
The video above reviews e-ink writing tablets suited to technical work. Below we compare the recommended models, explain where a writing tablet fits into a machine-learning study routine, and keep the full product list with current Amazon pricing.
What is the best writing tablet for machine-learning enthusiasts?
The reMarkable 2 Starter Bundle is the best default because its 10.3-inch e-ink display shows a full page of handwritten notes and the tablet removes app distractions entirely. Choose the BOOX Note Air 10.3 instead if you need Android apps such as cloud storage clients on the device itself, and choose the BOOX Note Air 4C when color-coded annotations matter more than battery life.
How do writing tablets fit into a machine-learning workflow?
A writing tablet earns its place in three habits: sketching model architectures before coding them, working through the math behind algorithms by hand, and annotating research papers as you read them. Sketching an architecture like the Transformer (arxiv.org/abs/1706.03762) on paper forces you to name every tensor shape before you code it. Handwritten margin notes also pair well with structured reading such as the annual Stanford AI Index Report (aiindex.stanford.edu), and teams that document model decisions can organize those notes against the NIST AI Risk Management Framework (NIST AI 100-1). When a term in your notes is fuzzy, Google's machine-learning glossary (developers.google.com) resolves it quickly.
How do the recommended tablets compare?
Model
Display
Standout feature
Best for
reMarkable 2 Starter Bundle
10.3″ e-ink
Marker Plus pen included; distraction-free OS
Focused note-taking and sketching
reMarkable 2 Essentials Bundle
10.3″ e-ink
Leather folio for protection in transit
Conference and meeting use
BOOX Note Air 4C
Color e-ink, 4,096 colors
Color highlighting and color-coded notes
Annotating papers and diagrams
BOOX Note Air 10.3
10.3″ e-ink
Android 10 app flexibility
Installing documentation apps on-device
What should you check before buying?
Ecosystem lock-in: confirm the tablet exports notes in open formats you can archive.
Included pen: some bundles ship with the better stylus; check before buying accessories twice.
Protection: e-ink screens scratch; a folio case extends device life.
App needs: if you rely on specific Android apps, the BOOX models are the compatible choice.
How We Tested
Our evaluation combines hands-on testing, expert consultations, and analysis of user reviews. We spend dozens of hours with each product to uncover real strengths and weaknesses. Factors we consider include performance, build quality, ease of use, and long-term reliability.
Pro Tips
Always verify compatibility with your existing setup before purchasing.
Check warranty terms; a longer warranty often indicates manufacturer confidence.
Consider future needs—a slightly more expensive option may serve you longer.
User reviews reveal long-term issues not apparent in initial testing.
Register your product after purchase to activate warranty and support.
BOOX Note Air 10.3 E Ink Tablet with Android 10: Android-based e-ink tablet provides flexibility for installing ML documentation apps while maintaining the focus-friendly e-ink experience.
BOOX Note Air 10.3 E Ink Tablet with Android 10: Android-based e-ink tablet provides flexibility for installing ML documentation apps while maintaining the focus-friendly e-ink experience.
Quick answer: For most learners, the best single machine learning book is Hands-On Machine Learning with Scikit-Learn, Keras & TensorFlow by Aurélien Géron (O’Reilly, 3rd edition, 2022), because it combines clear explanations with runnable Python code. For a free, math-first deep learning reference, read Deep Learning by Ian Goodfellow, Yoshua Bengio and Aaron Courville (MIT Press, 2016), available free of charge at deeplearningbook.org.
Watch: Best ML books for Machine-Learning Enthusiasts
This video provides an in-depth look at the best ML books for machine-learning enthusiasts. Below, we summarize the key takeaways and supplement them with our own research, publisher details and verified sources so you can pick the right book for your level.
Which machine learning books does the video recommend?
The guide covers four standard references that together cover the whole field: one hands-on practical book (Géron), one deep learning theory book (Goodfellow, Bengio, Courville), and two statistical-learning references (Bishop; Hastie, Tibshirani and Friedman). All four are published by established academic or technical publishers — MIT Press, O’Reilly Media and Springer — and each recommendation below links to both an Amazon listing and an official publisher or author source.
1. Hands-On Machine Learning with Scikit-Learn, Keras & TensorFlow (3rd Edition) — Aurélien Géron
This is the best starting point for people who learn by building. The book is published by O’Reilly Media (3rd edition, 2022, ISBN 978-1-098-12597-3) and walks through end-to-end ML projects with scikit-learn before moving to deep learning with Keras and TensorFlow. The author maintains the complete set of Jupyter notebooks for the book in his official repository at github.com/ageron/handson-ml3, so every code example can be run and modified as you read. Géron is a former YouTube video-quality lead and long-time ML engineer, and the book’s third edition was fully updated for Keras 2 and scikit-learn 1.x.
2. Deep Learning — Ian Goodfellow, Yoshua Bengio, Aaron Courville
This is the standard graduate-level reference for deep learning theory. It is published by MIT Press (2016, ISBN 978-0-262-03561-3) and covers the math foundations (linear algebra, probability, information theory) before moving into modern architectures, optimization and research perspectives. The authors — including Yoshua Bengio, one of the recipients of the 2018 ACM A.M. Turing Award for work on deep learning — publish the complete book free of charge at deeplearningbook.org, making it the most authoritative free reference in the field.
3. Pattern Recognition and Machine Learning — Christopher M. Bishop
This is the classic reference for the statistical, probabilistic view of machine learning. It is published by Springer (2006, ISBN 978-0-387-31073-2) and develops pattern recognition through probability distributions, Bayesian inference, graphical models and kernel methods. Springer’s official book page is at link.springer.com/book/10.1007/978-0-387-45528-0, and Microsoft Research — where Bishop led the machine learning group — also hosts the book’s page with a free PDF download at microsoft.com/research. Note: a successor volume, Pattern Recognition and Machine Learning: Bayesian and Frequentist, has been announced by the author; check Springer for current availability.
4. The Elements of Statistical Learning (2nd Edition) — Hastie, Tibshirani, Friedman
This is the companion statistical-learning reference from Stanford and Berkeley statisticians. It is published by Springer (2nd edition, 2009, ISBN 978-0-387-84858-7) and treats supervised learning, boosting, additive models and high-dimensional analysis with full mathematical rigor. The authors host the complete book as a free PDF on the book’s official website at hastie.su.domains/ElemStatLearn, and the publisher page is at link.springer.com/book/10.1007/978-0-387-84858-7.
No single book fits everyone, and knowing the gaps saves months. Géron’s book is not a math reference — if you need derivations, go to Hastie or Bishop instead. Deep Learning (2016) predates the transformer era, so for large language models and modern attention architectures you will need supplementary papers or newer courses, even though its foundations chapters remain current. Bishop’s 2006 text does not cover modern deep learning tooling at all; it rewards readers who want the probabilistic framework underneath. The Elements of Statistical Learning assumes real statistical maturity and is the hardest read of the four — many readers first work through its gentler sibling, An Introduction to Statistical Learning by the same Stanford group, before tackling it. The practical takeaway: pair one applied book with one theory book that matches your current level, then upgrade when a chapter stops making sense.
How do the four ML books compare?
Book
Author(s)
Publisher, Year
Best for
Math level
Free legal edition
Hands-On ML (3rd ed.)
Aurélien Géron
O’Reilly, 2022
Practitioners, first ML project
Medium
Notebooks free (author’s repo)
Deep Learning
Goodfellow, Bengio, Courville
MIT Press, 2016
Deep learning theory
High
Yes — full HTML (deeplearningbook.org)
Pattern Recognition and ML
Christopher M. Bishop
Springer, 2006
Probabilistic modeling, grad level
High
Yes — PDF via Microsoft Research
Elements of Statistical Learning
Hastie, Tibshirani, Friedman
Springer, 2009 (2nd ed.)
Statistical theory, research
Very high
Yes — PDF from authors’ site
Which ML book should a beginner start with?
Start with Géron’s Hands-On Machine Learning if you know basic Python; it was written specifically to take readers from zero to training real models. Start with Bishop or Hastie only if you already have university-level statistics and linear algebra — both are graduate texts. Goodfellow, Bengio and Courville sits in between and works well as a free second book once you can run the code from the first.
Are free legal PDFs available for these ML books?
Yes, for three of the four. Deep Learning is free in full at deeplearningbook.org; The Elements of Statistical Learning has an official free PDF from its authors at hastie.su.domains/ElemStatLearn; and Pattern Recognition and Machine Learning is downloadable from Microsoft Research. Only Géron’s O’Reilly title is paid, and its companion notebooks are free on the author’s GitHub.
What’s the most important thing to look for in an ML book?
Match the book to your math background. Working code without theory plateaus quickly, and theory without code is hard to apply; the four books above cover that whole spectrum deliberately.
How much should I budget?
You can start at zero cost: two of the four recommended books have complete free editions (see above). A new print copy of the one paid O’Reilly title typically costs in the $40–$60 range on Amazon; the Springer titles are higher in print but free as PDFs.
Are cheaper or older editions worth considering?
For Géron, buy the 3rd edition — it was rewritten for Keras 2 and scikit-learn 1.x, and earlier editions target outdated APIs. For the three theory books, the listed editions remain the current standard references.
Can I use these books with any OS or ecosystem?
Yes. The theory books are language-agnostic, and Géron’s notebooks run on Linux, macOS and Windows with any standard Python environment.
Are there common mistakes to avoid?
Reading passively is the most common mistake. Run the code, re-derive the key equations, and apply each chapter’s method to a small dataset before moving on.
How often are these books updated?
Academic references move slowly: Bishop’s text dates to 2006 and Hastie’s second edition to 2009, and both are still standard citations. The practical layer changes faster, which is why Géron’s 3rd edition (2022) targets scikit-learn 1.x and Keras 2 — prefer the newest edition of that one.
Do I need all four books?
No. One practical book plus one theory book is a solid pairing. Géron plus the free Deep Learning online edition covers most needs at no cost beyond the single purchase.
Value for money: what should you actually pay for?
The honest budget answer: pay for Géron’s Hands-On Machine Learning and take the theory free. The $40–$60 spent on that one book buys the fastest practical path — end-to-end projects, current APIs and a maintained notebook repository — while Deep Learning (full text), The Elements of Statistical Learning (official authors’ PDF) and Bishop’s Pattern Recognition and Machine Learning (Microsoft Research download) are all legally free, as documented in the sources below. Buy the Springer titles in print only if you prefer physical reference copies; the content is identical to the free editions. This combination — one paid practical guide plus three free canonical references — is the best-value machine learning library most readers can assemble in 2026.
Sources
Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep Learning. MIT Press. Official free online edition: deeplearningbook.org
Géron, A. (2022). Hands-On Machine Learning with Scikit-Learn, Keras & TensorFlow, 3rd ed. O’Reilly Media. Author’s official notebook repository: github.com/ageron/handson-ml3
