What's new

Changes

• After an accepted submission, share your solution with a title and an optional explanation, right from the test result panel.
• A new Community tab on every problem shows shared solutions with their code inline, sorted by top votes or newest.
• Upvote the solutions you find most helpful, and open an author's profile from their solution.
• Edit or delete your own shared solutions at any time.
• Works on free problems, study plan problems (Python, SQL, CUDA and Triton) and research paper problems.

Changes

• Redesigned study plan badges as medal plaques, including new artwork for every sheet.
• Research paper badges are now medal crests.
• Topic mastery badges are now medal gems.
• Badge images load faster across profiles and the dashboard.

Changes

• Grouped LLM Internals articles into Attention and inference, Training and fine-tuning, and Efficiency and scaling.
• Added an introduction to ML Math that covers all seven modules and suggests where to start, from Basic Statistics through Backpropagation.

Changes

• Toni no longer points out issues in code you have already changed, and focuses on your current version instead.
• Code suggestions are checked against the reference solution before they appear, so fewer incorrect snippets reach you.

Changes

• The test result panel now uses the same layout and styling as the test case panel, across Python and SQL problems.
• Replaced the hourglass loading animation with a new TensorTonic dolphin.

Changes

• Try Similar now skips problems you have already solved.
• Suggestions include problems from study plans, not only free problems.
• Pro problems appear with a lock, so you can see what is available before opening them.
• Try Similar now works on study plan problems, including CUDA, SQL and Triton.

Changes

• Add your own code and Markdown cells anywhere in the Qwen fine-tuning notebook, and they are saved with your draft.
• Markdown cells can be previewed in place.
• Run All now works reliably on longer notebooks.
• Clarified the hint for finding the median answer length.

Changes

• Redesigned architecture diagrams for 21 research papers, including Transformer, BERT, GPT-2, Llama, ViT, DeepSeek-V3, Kimi K3 and ResNet.
• All diagrams now share one visual style, making it easier to compare architectures across papers.

Changes

• Tab now accepts the selected autocomplete suggestion and still indents when suggestions are closed.

Changes

• Search free and study plan problems together, with technology filters and dedicated topic pages.
• Faster problem-list loading with fewer unnecessary requests.

Changes

• Refreshed light mode across TensorTonic, with an interactive homepage and clearer pricing comparisons.

Changes

• Expanded ML Design with 29 illustrated lessons and access to 21 design problems, with a yellow Beta label identifying the course as an early release.
• Read each problem description, requirements, and constraints in the course layout before opening its design canvas, while keeping the same left sidebar.
• Moved design problems under /ml-design and redirected previous /system-design links to their new locations.
• Added loading and error states so design problems no longer leave an empty reading area while their content loads.
• Updated Terms and Privacy to use the same navigation and solid header background as the Research page.
• Added Python source locations and expressions to the remaining 28 Research Engineer problems, completing coverage of all 48 Python problems in the track.
• Added detailed traces to the first ten ML Engineer assessment problems, including k-NN, train/test splits, ROC AUC, and Gaussian Naive Bayes.
• Refined Build LLM from Scratch problems.

Changes

• Added real Python tracebacks to the first 20 Research Engineer assessment problems.
• Syntax and runtime errors include the relevant solution file, line number, and expression in Run and Submit results.
• Refined Build Micrograd from Scratch and Build LLM from Scratch problems.

Changes

• Updated the remaining 23 Data Scientist Python problems, completing the 83-problem rollout.
• Failed Python programs now include source locations and expressions while retaining each problem's existing grading rules.
• Refined Build AlphaGo from Scratch and Build Micrograd from Scratch problems.

Changes

• Simplified homepage navigation and made the feature list clickable, with more space for each product preview.
• Refreshed sign-in and account creation with shared artwork, clearer form layouts, and more prominent Google and GitHub buttons.
• Applied the same layout to forgotten-password and password-reset pages, including invalid or expired reset links.
• Reworked the RAG introduction around a simple comparison of answers with and without retrieved context, supported by visuals for retrieval, chunking, and context limits.
• Added detailed syntax and runtime errors to 50 more Data Scientist assessment problems, bringing coverage to 60.
• Nested exceptions retain the relevant calls and source lines from your submitted solution.
• Refined Inference Engineering problems.

Changes

• Added detailed Python traces to all 25 Micrograd, 18 AlphaGo, and 57 Build LLM from Scratch exercises.
• Added source locations and expressions to errors in 17 RAG, ReAct, and MCP project exercises.
• Started the Data Scientist assessment rollout with ten Python problems.
• Fixed the BPE Merge Ranks exercise's Run input setup and corrected inconsistent grading in the muP Invariant Diagnostics exercise.
• Kept AlphaGo test results readable when supported inputs include NaN or infinity.
• Refined PyTorch, NumPy and Pandas sheet problems.

Changes

• Expanded LLM Internals with guides to RAG, attention, FlashAttention, and PagedAttention, including visuals for retrieval, attention weights, and memory use.
• Added gradient accumulation and mixed precision training guides with interactive examples of batch sizes, number formats, and loss scaling.
• Added RLHF, PPO, GRPO, DPO, and chain-of-thought reasoning guides with visual explanations of the underlying methods.
• Added detailed Python traces to all 25 NumPy, 30 PyTorch, 25 Pandas, and 30 Inference exercises.
• Completed the Computer Vision trace rollout with the remaining 15 problems.
• Fixed negative-axis handling in Per-Channel Quantization and clarified decode ordering in Chunked Prefill Scheduling.
• Fixed the default null-bias input validation in Convolution Operation so its public cases can run.
• Aligned the BPE, ResNet Block, and GAN Training explanations with the implementations each exercise asks for.
• Refined Cracking ML, Cracking DL, Cracking NLP, Cracking RL and Cracking CV problems.

Changes

• Added detailed error traces to 35 Natural Language Processing and 30 Reinforcement Learning problems.
• Started the Computer Vision rollout with 20 problems.
• Run and Submit errors show the relevant solution lines, including nested and chained exceptions.
• Refined Cracking DL, Cracking NLP, Cracking RL and Cracking CV problems.

Changes

• Completed the Linear Algebra trace rollout and expanded detailed errors across Probability and Statistics, Optimization, Calculus, Machine Learning, and Deep Learning.
• Fixed unreadable result output for NaN and infinite values in affected Linear Algebra and Statistics exercises.
• Mean, Median and Mode, Sample Variance, and Linear Regression Closed Form now reject invalid non-finite answers that could previously pass.
• Linear Regression Closed Form failures now retain the correct test identity and visibility in the results panel.
• Refined Linear Algebra, Probability and Statistics, Optimization, Calculus, Cracking ML and Cracking DL problems.

Changes

• Syntax and runtime errors in Implement Dot Product now identify the solution file, line number, and triggering expression.
• Nested function failures retain their Python call trace in both Run and Submit results.

Changes

• Start with bits, bytes, and number formats, then learn how quantization works and how it combines with LoRA.
• Explore interactive visuals for model memory, NF4, and the QLoRA training step, with a Hugging Face setup to connect the concepts to code.

Changes

• Apply to the Creator Program and receive your personal referral link after approval.
• Earn 25% on eligible Pro and Plus purchases and renewals from new accounts created through your link.
• Track referred customers, commissions, and payout history in your creator portal.

Changes

• Expanded the rollout across the 200 free Python problems so errors include the relevant solution file, line number, and source expression.
• Corrected the Anchor Box Generation starter docstring to describe corner coordinates [x1, y1, x2, y2], matching the problem and grader.

Changes

• Browse study plans with new topic illustrations on the Problems and Study Plans pages.
• Interview-prep cards now feature dedicated artwork and updated company-logo layouts.
• Refreshed card typography and colors improve readability in light and dark themes.

Changes

• See the relevant lines of your solution in Python syntax and runtime error traces.
• Tracebacks focus on your submitted code, with internal runner frames removed.
• Error output preserves line breaks and indentation to make debugging easier.

Changes

• Select a GPU for CUDA and Triton problems and see kernel timings across multiple input sizes after a successful submission.
• Compare your kernel against a reference implementation on the same GPU, and keep results from different GPUs side by side for the same code.
• Run kernels on NVIDIA H100, H200, RTX 5090, and RTX PRO 6000, and T4.

Changes

• Ask Toni for hints, explanations, and debugging help directly from study-plan problems.
• Use Toni while practicing interview problems without leaving the editor.

Changes

• Added a course introduction that connects requirements, data, modeling, evaluation, and production through a reusable design framework.
• Introduced system diagrams and a model-lifecycle visual showing how data collection, training, deployment, and monitoring fit together.

Changes

• Ask Toni, our AI assistant for every problem for progressive hints, debugging guidance, and detailed explanations while keeping the implementation yours.
• Research-problem starter code now includes data types and clearer return formats.

Changes

• Reload the Problems page faster with the initial problem list ready when the page opens.
• See each user's leaderboard rank directly on their profile.
• Free-problem starter code now includes data types and clearer return formats.

Changes

• Open all your saved problem notes from the account menu or directly from the notes editor.
• Check your remaining CUDA, Triton, and GPU project usage from the new GPU usage indicator.

Changes

• Learn how LoRA fine-tunes language models through an interactive LLM Internals guide.
• Pause and resume timed interview prep sessions without losing your remaining solve time.

Changes

• Accepted code and SQL interview prep problems now count toward progress.
• Solved interview prep problems now appear in recent activity.

Changes

• Share interview experiences, ask questions, vote, and reply.
• Continue the LLM Internals series with speculative decoding.

Changes

• Thirteen computer vision visualizations were refreshed.
• Solutions for free problems in the Pandas, NumPy, PyTorch, and SQL Sheets are now available to signed-in learners.

Changes

• Fine-tune Qwen3-4B in a guided GPU notebook.
• Build an inference server and explore Mixture of Experts routing.
• Added a Mixture of Experts guide covering feed-forward layers, expert selection, and the path a token takes through an MoE layer.

Changes

• Follow a guided roadmap from foundations through complete projects.
• Track your TensorTonic 125 progress and continue directly into practice.

Changes

• Build a RAG system and a ReAct agent step by step.
• Create an MCP server through guided notebook exercises.

Changes

• Practice language modeling with NumPy, PyTorch, CUDA, and Triton.
• Move from core components to training and systems exercises.

Changes

• Submit product ideas directly from TensorTonic.
• Browse requests and see what other learners want next.

Changes

• Understand common use cases for conventional machine learning.
• Follow best practices from data preparation through model evaluation.

Changes

• Learn policy networks, value networks, and Monte Carlo tree search.
• Work through the system with interactive visual explanations.

Changes

• Build a scalar autograd engine and neural network from first principles.
• Use TensorTonic in a complete light theme across the platform.

Changes

• Study the Kimi K3 architecture through ten interactive visualizations.
• Inspect its model design and training concepts in one guided page.

Changes

• Add a TensorTonic progress badge to your GitHub profile README.
• Keep a generated index of solved problems in your solutions repository.

Changes

• Walk through GPT pretraining with interactive visualizations.
• Study attention, batching, quantization, KV cache, and MoE serving.

Changes

• Implement transformer components through a structured capstone.
• Run and evaluate the completed model inside TensorTonic.

Changes

• Explore GPT-2, Llama, Gemma 3, Word2Vec, and DenseNet visually.
• Inspect DeepSeek-V3, GLM-4.5, GPT-OSS, and Arcee Trinity.

Changes

• Practice LLM internals and research frontier mathematics.
• Test training, decoding, post-training, and alignment knowledge.

Changes

• Find problem sets and study plans through a redesigned Problems page.
• See the full learning experience through a refreshed landing page.