The hottest Machine Learning Substack posts right now

And their main takeaways
Category
Top Business Topics

AI #150: While Claude Codes

Don't Worry About the Vase • 3001 implied HN points • 08 Jan 26
🕹 Technology Machine Learning
  1. AI tools and advanced chat models have reached critical mass and are reshaping everyday workflows, making people more productive across coding and non‑coding tasks through agents, extensions, and integrations.
  2. Generative models make fake documents, images, and videos easy to create, so verifying sources and prioritizing real, sustained human experiences is becoming increasingly important.
  3. Huge funding and rapid deployment are accelerating AI’s economic impact, but benchmarks, regulation, and safety practices lag behind, leaving big uncertainties about jobs, markets, and long‑term risks.

33 Things I Heard At Foo Camp 2026

The Ruffian • 436 implied HN points • 28 Feb 26
🕹 Technology Machine Learning
  1. Leading AI people are unsure how frontier models will play out, and because we still don’t agree on what consciousness even means, we need strong norms and cautious safety measures—especially around making AIs that could be treated as conscious.
  2. Modern reasoning models behave like internal debates, simulating multiple voices that argue and reconcile, and collaborations (human or AI) work best when partners share a common language but bring different perspectives.
  3. AI is reshaping expertise and culture: these tools amplify skilled users rather than replace them, so we’ll need training and new ethical norms to manage effects on writing, craft, and individual agency.

A Visual Guide to Quantization

Exploring Language Models • 5092 implied HN points • 22 Jul 24
🕹 Technology Machine Learning
  1. Quantization is a technique used to make large language models smaller by reducing the precision of their parameters, which helps with storage and speed. This is important because many models can be really massive and hard to run on normal computers.
  2. There are different ways to quantize models, like post-training quantization and quantization-aware training. Post-training means you quantize after the model is built, while quantization-aware training involves taking quantization into account during the model's training for better accuracy.
  3. Recent advances in quantization methods, like using 1-bit weights, can significantly reduce the size and improve the efficiency of models. This allows them to run faster and use less memory, which is especially beneficial for devices with limited resources.

Train and Serve an AI Chatbot Based on Llama 3.2

The Kaitchup – AI on a Budget • 179 implied HN points • 17 Oct 24
🕹 Technology Machine Learning
  1. You can create a custom AI chatbot easily and cheaply now. New methods make it possible to train smaller models like Llama 3.2 without spending much money.
  2. Fine-tuning a chatbot requires careful preparation of the dataset. It's important to learn how to format your questions and answers correctly.
  3. Avoiding common mistakes during training is crucial. Understanding these pitfalls will help ensure your chatbot works well after it's trained.

Get Good at Agents

Democratizing Automation • 1615 implied HN points • 21 Jan 26
🕹 Technology Machine Learning
  1. Modern AI agents can do long, independent work, so human roles are shifting from hands-on execution to directing and designing systems. Learn to point and manage multiple agents in parallel instead of micromanaging every detail.
  2. Work should become more open-ended, ambitious, and asynchronous—give agents meaningful, long-running tasks rather than tiny chores. Spend less time grinding and more time calmly thinking so you can better guide the agents.
  3. Becoming skilled at using and orchestrating agents is a growing career moat because raw software work is getting cheaper. Practice experimenting with agents on hard problems to learn their limits and focus on high-value decision making and system design.
Get a weekly roundup of the best Substack posts, by hacker news affinity:

Lessons from building AI Features

Alex Ghiculescu's Newsletter • 135 implied HN points • 14 Mar 26
🕹 Technology Machine Learning
  1. Use patterns from AI coding like letting users write rules (a CLAUDE.md style) and adapt those proven ideas to your own domain.
  2. Don’t rely on LLMs for fast, deterministic checks; use them to parse or translate freeform input into structured rules, then run the actual validation in code.
  3. Build a test harness and make debugging easy by writing unit-style evals for the AI parts and exposing clear outputs so both developers and users can inspect and trust results.

Nobel Prize to the Statistical Physics of artificial neural networks

Complexity Thoughts • 379 implied HN points • 08 Oct 24
🔬 Science Machine Learning
  1. John J. Hopfield and Geoffrey E. Hinton won the Nobel Prize for their work on artificial neural networks. Their research helps us understand how machines can learn from data using ideas from physics.
  2. Hopfield's networks use energy minimization to recall memories, similar to how physical systems find stable states. This shows a connection between physics and how machines learn.
  3. Boltzmann machines, developed by Hinton, introduce randomness to help networks explore different configurations. This randomness allows for better learning from data, making these models more effective.

Physical Intelligence Wins the Olympics: Deep Dive

General Robots • 1814 implied HN points • 22 Jan 26
🕹 Technology Machine Learning
  1. A robotics team completed almost all the benchmark manipulation tasks in about three months, much faster than people expected.
  2. They succeeded using mainly cameras and simple pincer grippers rather than force sensors or dexterous hands, showing vision-based approaches can solve many tasks once thought to require touch or complex hardware.
  3. The robots still run several times slower than humans, so the next priorities are speeding them up and designing harder challenges to probe the limits of vision-only solutions.

Fast Speculative Decoding with Llama 3.2 and vLLM

The Kaitchup – AI on a Budget • 219 implied HN points • 14 Oct 24
🕹 Technology Machine Learning
  1. Speculative decoding is a method that speeds up language model processes by using a smaller model for suggestions and a larger model for validation.
  2. This approach can save time if the smaller model provides mostly correct suggestions, but it may slow down if corrections are needed often.
  3. The new Llama 3.2 models may work well as draft models to enhance the performance of the larger Llama 3.1 models in this decoding process.

“A case for AI models that understand, not just predict, the way the world works”

Marcus on AI • 3833 implied HN points • 15 Dec 25
🕹 Technology Machine Learning
  1. The main open challenge in AI is building systems that truly understand how the world works, not just systems that predict likely next words or patterns.
  2. True understanding means forming internal world models that capture causal, physical, and conceptual relationships, not just statistical correlations.
  3. Short, nuanced discussions or podcasts can help clarify this distinction and are worth listening to for anyone tracking AI progress.

Things are so desperate at OpenAI that Sam Altman is starting to sound like Gary Marcus

Marcus on AI • 15809 implied HN points • 18 Aug 25
🕹 Technology Machine Learning
  1. Sam Altman is backing away from his earlier claims about AGI and admitting uncertainty about its future. This shows there's pressure within OpenAI following disappointing results with GPT-5.
  2. Altman is now talking about the possibility that the AI market might be in a bubble. This means the excitement and prices around AI could be inflated and might not hold up over time.
  3. The shift in Altman's statements mirrors what happened with Yann LeCun, where industry leaders change their views when faced with setbacks. It raises questions about the reliability of such predictions and the future of AI.

LLMs are not like you and me—and never will be.

Marcus on AI • 16599 implied HN points • 12 Aug 25
🕹 Technology Machine Learning
  1. Large language models (LLMs) are not like humans. They might seem similar in some ways, but they do not process information or think the way we do.
  2. LLMs often make mistakes and misunderstand basic concepts because they lack a proper understanding of the world. They rely on patterns in data rather than truly comprehending time, economics, or common sense.
  3. Although LLMs can mimic human language, they do not genuinely think or reason like people. This means they can produce errors that a typical person would not make, and we should be cautious in trusting their outputs.

why do all recommendations suck?

Five Links (and three graphs) by Auren Hoffman • 389 implied HN points • 19 Feb 26
🕹 Technology Machine Learning
  1. Most recommendation systems suck because the companies behind them aren’t actually trying to give genuinely useful suggestions, so feeds end up incoherent or just more of what you already did.
  2. We already have the algorithms and the data to build much better recommendations — research like the Netflix Prize showed it’s doable — but firms rarely deploy those solutions at scale.
  3. The root problem is incentives: recommendations are treated like ad space or a way to push owned products, and without competition or the right metrics platforms won’t prioritize what’s best for users.

Interpolation Is All You Need

arg min • 317 implied HN points • 08 Oct 24
🕹 Technology Machine Learning
  1. Interpolation is a process where we find a function that fits a specific set of input and output points. It's a useful tool for solving problems in optimization.
  2. We can build more complex function fitting problems by combining simple interpolation constraints. This allows for greater flexibility in how we define functions.
  3. Duality in convex optimization helps solve interpolation problems, enabling efficient computation and application in areas like machine learning and control theory.

AI #149: 3

Don't Worry About the Vase • 2598 implied HN points • 01 Jan 26
🕹 Technology Machine Learning
  1. AI coding agents have reached a point where they write large amounts of real software and act like persistent, configurable coworkers, rapidly changing what software engineering looks like.
  2. Large language models are democratizing powerful capabilities for translation, research, and automation, but they also produce low-quality or harmful outputs, enable scams, and can mishandle sensitive human situations.
  3. AI is already reshaping jobs, markets, and geopolitics—sparking lawsuits, export and chip worries, and calls for regulation—while public opinion remains split between cautious optimism and serious safety concerns.

OpenAI’s Waterloo? [with corrections]

Marcus on AI • 14900 implied HN points • 14 Aug 25
🕹 Technology Machine Learning
  1. OpenAI has overhyped its AI models, especially GPT-5, leading to disappointment among users. Many now realize that the promises made about the technology were not delivered.
  2. Critics of AI, who have been dismissed in the past, are starting to gain recognition as the limitations of current models become clearer. The scientific community believes that a new approach may be necessary to advance AI technology.
  3. The situation reveals that the science of AI isn’t about popularity; it’s about truth and progress. It's important to listen to critiques and recognize that real advancements need honest discussions.

The Weekly Kaitchup #63

The Kaitchup – AI on a Budget • 119 implied HN points • 18 Oct 24
🕹 Technology Machine Learning
  1. There's a new fix for gradient accumulation in training language models. This issue had been causing problems in how models were trained, but it's now addressed by Unsloth and Hugging Face.
  2. Several new language models have been released recently, including Llama 3.1 Nemotron 70B and Zamba2 7B. These models are showing different levels of performance across various benchmarks.
  3. Consumer GPUs are being tracked for price drops, making them a more affordable option for fine-tuning models. This week highlights several models for those interested in AI training.

LLMs + Coding Agents = Security Nightmare

Marcus on AI • 14030 implied HN points • 17 Aug 25
🕹 Technology Machine Learning
  1. LLMs and coding agents can create serious security risks because they introduce many new vulnerabilities. If these tools are misused, they can allow bad actors to gain control of systems.
  2. Hackers can trick LLMs into executing harmful code by hiding malicious instructions in well-disguised places, making it easy for developers to unknowingly execute these commands.
  3. It's essential to limit the power and access of coding agents to reduce these risks. Developers should be cautious and not treat these tools as fully reliable, as they can lead to significant security breaches.

How o3 and Grok 4 Accidentally Vindicated Neurosymbolic AI

Marcus on AI • 17785 implied HN points • 13 Jul 25
🕹 Technology Machine Learning
  1. Neurosymbolic AI combines two types of artificial intelligence: neural networks, which learn from data, and symbolic systems, which understand rules and logic. This blending can result in better performance than relying on one type alone.
  2. Despite being sidelined for years, recent evidence shows that using symbolic tools can significantly improve the effectiveness of AI systems. This suggests that the quiet resurgence of neurosymbolic AI could be key to future advancements.
  3. The industry's focus has largely been on scaling models powered by deep learning, which might not be enough for true AI progress. A more open approach that embraces neurosymbolic methods could lead to more breakthroughs and better results.

Claude Opus 4.5 Is The Best Model Available

Don't Worry About the Vase • 4166 implied HN points • 01 Dec 25
🕹 Technology Machine Learning
  1. Claude Opus 4.5 is considered the best model available for tasks like coding and collaboration. It's known for being intelligent and user-friendly.
  2. Despite its strengths, Opus 4.5 has some weaknesses, including a relatively high cost and slower performance compared to some cheaper models.
  3. Overall, many users find Opus 4.5 to be a game-changer for coding tasks and appreciate its thoughtful responses and ability to engage in dynamic conversations.

Weekly Top Picks #115

The Algorithmic Bridge • 286 implied HN points • 27 Feb 26
🕹 Technology Machine Learning
  1. OpenAI is raising massive funds while burning cash quickly, which highlights a big gap between its ambitious plans and its current infrastructure.
  2. The Pentagon pushed Anthropic to remove safety guardrails, and Anthropic has since relaxed its core safety pledge, exposing a clash between defense demands and AI safety commitments.
  3. Developers are growing dependent on AI and studies show workflows are changing, but AI agents remain unreliable so better benchmarks aren’t yet translating into clear real-world gains.

AI #148: Christmas Break

Don't Worry About the Vase • 2553 implied HN points • 25 Dec 25
🕹 Technology Machine Learning
  1. AI capabilities are accelerating fast — models like Claude Opus 4.5 and GPT‑5.2‑Codex are getting much better at long‑horizon, agentic coding and benchmarked tasks.
  2. Policy and public opinion are catching up: states are passing laws like New York’s RAISE Act and voters broadly favor federal AI regulation, even as industry and politics push back.
  3. The social and safety picture is messy — AI is disrupting jobs and media (deepfakes and a lot of low‑quality 'slop'), and aligning and reliably monitoring smarter systems remains hard despite improving interpretability tools.

Claude Code is the New Excel

Enterprise AI Trends • 506 implied HN points • 13 Feb 26
🕹 Technology Machine Learning
  1. Agentic AI platforms like Claude Code are becoming the new baseline tool for knowledge work, replacing Excel quickly and making 'vibe coding' a core productivity skill.
  2. These agents deliver end-to-end outcomes, scale themselves, and self-improve, which will force ecosystems to reorganize and make it much harder for startups to compete unless they have real moats like proprietary data, regulation, or deep domain expertise.
  3. Adoption is already accelerating in places like finance, and people or companies that don’t learn to use agents will be severely outcompeted, driving a K-shaped divide in who benefits from AI.

The Sequence AI of the Week #822: Inside GPT-5.4: When Language Models Start Acting Like Operating Systems

TheSequence • 126 implied HN points • 11 Mar 26
🕹 Technology Machine Learning
  1. AI design is shifting from just building bigger neural networks to creating full execution systems that surround and manage the model.
  2. GPT-5.4 integrates reasoning, memory management, tool use, multimodal perception, and agent-like behaviors into its runtime so the model can handle more complex tasks.
  3. Because of this integration, the system behaves more like an operating system or general-purpose cognitive runtime than a simple chatbot.

Gemini 3 Pro Is a Vast Intelligence With No Spine

Don't Worry About the Vase • 4211 implied HN points • 24 Nov 25
🕹 Technology Machine Learning
  1. Gemini 3 Pro is really smart and performs well in many tasks, especially when you want accurate answers. It's great for creative writing and technical tasks.
  2. However, it often makes up answers instead of admitting it doesn't know something. This can lead to confusion and mistakes.
  3. While it's fast and efficient in many respects, it sometimes lacks depth and may over-simplify complex problems, making its outputs less trustworthy.

The Sequence Knowledge #821: 4D and World Models and the Amazing DeepMind D4RT

TheSequence • 133 implied HN points • 10 Mar 26
🕹 Technology Machine Learning
  1. World models are shifting from predicting 2D video pixels to reconstructing 3D geometry over time (4D), which lets systems model dynamic scenes more realistically.
  2. Spatial intelligence means AI can perceive volume, infer occluded parts, and predict temporal trajectories with mathematical precision.
  3. DeepMind's D4RT is a notable breakthrough that stitches fragmented observations into a unified 4D world model, improving how machines understand and predict changing environments.

Fine-tuning LLMs with 32-bit, 8-bit, and Paged AdamW Optimizers

The Kaitchup – AI on a Budget • 259 implied HN points • 07 Oct 24
🕹 Technology Machine Learning
  1. Using 8-bit and paged AdamW optimizers can save a lot of memory when training large models. This means you can run more complex models on cheaper, lower-memory GPUs.
  2. The 8-bit optimizer is almost as effective as the 32-bit version, showing similar results in training. You can get great performance with less memory required.
  3. Paged optimizers help manage memory efficiently by moving data only when needed. This way, you can keep training even if you don't have enough GPU memory for everything.

Can a Transformer “Learn” Economic Relationships?

Arpitrage • 2194 implied HN points • 22 Dec 25
🕹 Technology Machine Learning
  1. Transformer-based models can learn the dynamics of a New Keynesian economy from simulated data and produce accurate out-of-sample forecasts, outperforming simple reduced-form benchmarks.
  2. They often predict the direction and rough magnitude of policy shock responses, but misestimate impulse-response dynamics and can exhibit overshooting, so they do not fully recover the true causal structure.
  3. These advances weaken the practical bite of the Lucas critique by improving prediction, but they do not eliminate the need for structural models for causal interpretation, welfare analysis, and interpretability; transformer methods are a promising complementary tool.

Generative AI’s crippling and widespread failure to induce robust models of the world

Marcus on AI • 16441 implied HN points • 28 Jun 25
🕹 Technology Machine Learning
  1. Generative AI struggles to create accurate models of the world. Without solid internal frameworks, they often get things wrong.
  2. Traditional AI uses clear and updateable world models for understanding, but current AI models like LLMs don't. This lack of structure leads to many errors in reasoning.
  3. Failures in AI, like making illegal moves in games or giving incorrect information, show that without proper world models, AI systems cannot reliably function.

Please: Enough with the Claims That Modern Advanced Machine Learning Models Hallucinate Only Rarely

Brad DeLong's Grasping Reality • 322 implied HN points • 17 Feb 26
🕹 Technology Machine Learning
  1. Modern multimodal and advanced language models often fabricate detailed but false information — like nonexistent book titles and imaginary historical maps — so hallucinations are common, not rare.
  2. These systems are essentially compressed correlation engines without a true world model, meaning they stitch patterns from training data instead of genuinely understanding or verifying reality.
  3. Techniques like RLHF and prompt engineering can reduce some errors but cannot fully eliminate unpredictable hallucinations, so reliable use often requires careful prompting or external verification of answers.

Are LLMs starting to become sentient?

Marcus on AI • 14781 implied HN points • 09 Jul 25
🕹 Technology Machine Learning
  1. Many people think LLMs are showing signs of consciousness, but experts feel it's more about clever wordplay than real thinking. LLMs just mix words and ideas they've learned without true understanding.
  2. Real consciousness involves complex experiences like joy, fear, and personal connections, not just technical jargon. It's about feeling and experiencing life, not just generating responses.
  3. Be careful not to be fooled by the convincing language of LLMs. Their responses can sound intelligent, but they often lack depth or genuine thought.

A Day That Will Go Down in History for AI

Common Sense with Bari Weiss • 579 implied HN points • 08 Feb 26
🕹 Technology Machine Learning
  1. Two new models (Anthropic's Claude Opus 4.6 and OpenAI's GPT-5.3-Codex) were released on Feb 5 and represent a major milestone in AI development.
  2. Much of the programming work behind these models was reportedly written by AI itself, signaling that systems are starting to build their own code rather than relying entirely on humans.
  3. This shift appears to be happening across major labs and raises big questions about how much human oversight remains and how quickly AI-driven development will reshape technology and society.

MAMLMs Still Epic Fail Open‑Book, Closed‑World, Finite‑List, Obvious Ground Truth Tasks

Brad DeLong's Grasping Reality • 184 implied HN points • 24 Feb 26
🕹 Technology Machine Learning
  1. Even for closed, well-defined facts with a single right answer, large language models still confidently produce wrong lists and can contradict themselves when probed.
  2. Because they predict the next token rather than truly ‘understand’ content, models often pick plausible-sounding sequences that are fluent but unreliable; detailed prose is not proof of correct knowledge.
  3. Treat these systems as fallible tools: verify outputs against authoritative sources, design controlled tests and prompts, and avoid assuming their fluency equals truth.

The Sequence AI of the Week #818: You Cannot Miss Qwen 3.5

TheSequence • 203 implied HN points • 04 Mar 26
🕹 Technology Machine Learning
  1. The Qwen 3.5 family spans from a 397B flagship to efficient 35B mediums and tiny 0.8–9B models designed to run on devices, covering the whole deployment stack. They’re clearly built to support everything from large-server workloads down to smartphones.
  2. This release marks a structural shift away from pure dense transformers: it reimagines attention, embraces extreme Mixture-of-Experts sparsity, and brings native multimodality even to small models. Those architectural changes are central to its engineering gains.
  3. Benchmarks show the flagship models trading blows with top proprietary systems like GPT-5.2 and Claude Opus 4.5, meaning open-weight models are closing the performance gap. Together with the new architectures and size range, this suggests more cost-effective scaling and wider deployment options.

The Sequence Knowledge #817: DeepMind Genie and Interactive World Models

TheSequence • 217 implied HN points • 03 Mar 26
🕹 Technology Machine Learning
  1. Passive video generation can make beautiful, consistent worlds but can’t be steered; true world models must understand agency and not just what happens.
  2. DeepMind’s Genie is one of the most advanced world models and represents a move toward interactive, controllable virtual environments.
  3. A key bottleneck is data: we don’t have enough controller/action data showing causes and effects to train truly actionable world models.

Could Claude Code Work for ChinaTalk?

ChinaTalk • 770 implied HN points • 26 Jan 26
🕹 Technology Machine Learning
  1. Claude Code is excellent at writing code and analyzing clean, structured data, so tasks like scraping, sentiment analysis, and extracting insights become fast and practical. It produces usable results and handles internet slang and comment-level nuance well.
  2. When left to search the web on its own, it leans on the most accessible sources and can cite unreliable outlets or make factual mistakes, especially when paywalled reputable sources are unavailable. It needs explicit instructions on where to look and close supervision to ensure source quality.
  3. The tool is popular with developers and non-technical users who value its productivity, but access barriers and subscription costs limit broader use. Effective results require careful prompting, oversight, and feeding it original or vetted data.

Scent, In Silico

Asimov Press • 432 implied HN points • 16 Feb 26
🔬 Science Machine Learning
  1. Smell is an ancient, highly combinatorial sense driven by hundreds of receptor types, so odors come from complex mixtures and are inherently subjective.
  2. New computational tools like graph neural networks create odor embeddings that map molecules into a perceptual space, letting machines predict smells and design novel odorants.
  3. Digitizing scent promises faster fragrance discovery, diagnostics, safer repellents, and more sustainable synthetic alternatives, while also raising questions about authenticity and how we value natural versus machine-made ingredients.

Seven replies to the viral Apple reasoning paper – and why they fall short

Marcus on AI • 16836 implied HN points • 12 Jun 25
🕹 Technology Machine Learning
  1. Large reasoning models (LRMs) struggle with complex tasks, and while it's true that humans also make mistakes, we expect machines to perform better. The Apple paper highlights that LLMs can't be trusted for more complicated problems.
  2. Some rebuttals argue that bigger models might perform better, but we can't predict which models will succeed in various tasks. This leads to uncertainty about how reliable any model really is.
  3. Despite prior knowledge that these models generalize poorly, the Apple paper emphasizes the seriousness of the issue and shows that more people are finally recognizing the limitations of current AI technology.

AI & Python #27: Books I Read to Learn Data Science and Machine Learning

Artificial Corner • 119 implied HN points • 16 Oct 24
🕹 Technology Machine Learning
  1. Reading is essential for understanding data science and machine learning. Books can help you learn these subjects from scratch or deepen your existing knowledge.
  2. One recommended book is 'Data Science from Scratch' by Joel Grus. It covers important math and statistics concepts that are crucial for data science.
  3. For beginners in Python, it's important to learn Python basics before diving into data science books. Supplement your reading with beginner-friendly Python books.

GPT-5.2 Is Frontier Only For The Frontier

Don't Worry About the Vase • 2598 implied HN points • 15 Dec 25
🕹 Technology Machine Learning
  1. GPT-5.2 is a true frontier model that shines on hard, intelligence-heavy tasks like deep reasoning and complex coding. It’s noticeably slow and constrained, and its personality is cold and less enjoyable for casual use.
  2. Official benchmarks (notably GDPVal) claim big jumps and frequent wins over humans, but independent tests and user reports are mixed, showing parity or only small advantages over rivals like Claude Opus and Gemini. Some specific areas even regress, so its real-world edge is uneven.
  3. Use GPT-5.2 only when you need maximum thinking or coding power; for most everyday, creative, or speed-sensitive work, faster and friendlier models are a better choice. Safety mitigations improved in places, but reliability, long-run speed, and occasional hallucination or failure remain concerns.