The hottest Machine Learning Substack posts right now

And their main takeaways
Category
Top Business Topics

On Dwarkesh Patel's Second Interview With Ilya Sutskever

Don't Worry About the Vase • 1657 implied HN points • 03 Dec 25
đź•ą Technology Machine Learning
  1. Ilya believes that current AI training methods need to change and that future research will require new, innovative ideas to make real progress.
  2. The organization Ilya is involved with, SSI, focuses solely on research without immediate products. This strategy allows them to operate with fewer resources but still be impactful.
  3. Ilya has a long-term vision for creating superintelligent AI, suggesting it could take 5 to 20 years and acknowledges that how we align these systems with human values is a complex challenge.

Three Years from GPT-3 to Gemini 3

One Useful Thing • 2059 implied HN points • 18 Nov 25
đź•ą Technology Machine Learning
  1. AI has evolved from simple chatbots to more advanced tools that can code, design, and perform complex tasks. This means AI can now create interactive applications and help with various computer tasks, making it a powerful ally.
  2. The introduction of tools like Gemini 3 and Antigravity shows that AI can handle more complicated jobs, including data analysis and research. It can even write original papers, resembling a graduate student's intelligence level.
  3. With AI becoming more capable, the way we interact with it is changing. Instead of just fixing AI mistakes, people are now managing and directing AI's work, marking a shift from simple assistance to more of a collaborative partnership.

The Weekly Kaitchup #61

The Kaitchup – AI on a Budget • 139 implied HN points • 04 Oct 24
đź•ą Technology Machine Learning
  1. NVIDIA's new NVLM-D-72B model is a large language model that works well with both text and images. It has special features that make it good at understanding and processing high-quality visuals.
  2. OpenAI's new Whisper Large V3 Turbo model is significantly faster than its previous versions. While it has fewer parameters, it maintains good accuracy for most languages.
  3. Liquid AI introduced new models called Liquid Foundation Models, which are very efficient and can handle complex tasks. They use a unique setup to save memory and improve performance.

Why Anthropic Is Making Fun of OpenAI

Common Sense with Bari Weiss • 268 implied HN points • 09 Feb 26
đź•ą Technology Machine Learning
  1. Anthropic ran Super Bowl commercials that poke fun at a better-known AI rival to draw attention to the competition.
  2. The ads position Anthropic as a challenger to that rival’s dominance, suggesting a different, less domineering vision for AI’s future.
  3. By using humor, the campaign aims to shape public perception and spark debate about AI power, safety, and who should control the technology.

The Sequence Radar #811: Last Week in AI: OpenAI's Capital Leap, India's Summit, and the Next Frontier of Models

TheSequence • 175 implied HN points • 22 Feb 26
đź•ą Technology Machine Learning
  1. AI is entering a capital- and infrastructure-driven phase. Massive funding rounds and multibillion-dollar plans are being raised to build the silicon, power, and data centers needed for next-gen models.
  2. Model capabilities are leaping forward with agentic, long-context, and stronger reasoning abilities. New releases and research (for example Sonnet 4.6, Gemini 3.1 Pro, and GLM-5) push autonomous agent use, huge context windows, and improved problem-solving.
  3. Geopolitical and regional pushes are building sovereign AI stacks and expanding access. Global summits and large local investments are committing hundreds of billions to data centers, fiber links, and localized models to make AI national-scale infrastructure.
Get a weekly roundup of the best Substack posts, by hacker news affinity:

BYOB: Build Your Own Benchmark

Artificial Ignorance • 96 implied HN points • 01 Mar 26
đź•ą Technology Machine Learning
  1. Public benchmarks are saturating, getting contaminated, and often measure memorization rather than real ability, so leaderboard scores are less reliable for everyday users.
  2. Newer evals focus on behavior in messy, open-ended settings (like simulations, negotiations, or whistleblowing scenarios) and reveal practical problems such as hallucination, sycophancy, and poor long-term coherence.
  3. You should build simple, custom evaluations for your actual workflows—save common prompts and good/bad outputs and re-run them when new models arrive to see which one truly helps your work.

Software Machine Tools

Bzogramming • 61 implied HN points • 03 Mar 26
đź•ą Technology Machine Learning
  1. There is no universal machine tool: every manufacturing process has hard trade-offs in cost, speed, materials, and geometry, and even hypothetical atom-by-atom assemblers would face stability, energy, and material limits.
  2. In software, theoretical universality (Turing-completeness) doesn’t imply practical usefulness—different paradigms like programming languages, neural networks, and superoptimizers are distinct "software machine tools" with very different real-world strengths.
  3. Big opportunities lie in alternative software tools and analyses—verification-driven code synthesis, superoptimizers, compact magic-constant solutions, better static analysis, and more visual/geometric tooling can solve hard problems more efficiently than brute-force code or giant models.

Data Science Weekly - Issue 564

Data Science Weekly Newsletter • 119 implied HN points • 12 Sep 24
đź•ą Technology Machine Learning
  1. Understanding AI interpretability is important for building resilient systems. We need to focus on why interpretability matters and how it relates to AI's resilience.
  2. Testing machine learning systems can be challenging, but starting with basic best practices like CI pipelines and E2E testing can help. This ensures the models work well in real-world scenarios.
  3. Visualizing machine learning models is crucial for better understanding and analysis. Tools like Mycelium can help create clear visual representations of complex data structures.

Building Brains on a Computer

Asimov Press • 425 implied HN points • 26 Jan 26
🔬 Science Machine Learning
  1. New lab technologies and AI tools have rapidly lowered the cost and time needed to map neurons, so faithful brain emulations for small animals could appear in a few years and mouse-to-human scale emulations are plausible within decades if big investments continue.
  2. Creating full emulations requires three things — recording neural activity, reconstructing the wiring (connectome), and building accurate computational neuron models — and the biggest bottleneck is getting aligned, high-quality biological data and automating the tedious proofreading steps.
  3. Accurate brain emulations could become powerful discovery tools for neuroscience, drug development, and studying consciousness, but they will be costly, ethically complicated, and the first models will probably be generic population-style brains rather than perfect copies of individual people.

Kimi K2 Thinking

Don't Worry About the Vase • 1881 implied HN points • 11 Nov 25
đź•ą Technology Machine Learning
  1. Kimi K2 Thinking is an advanced open-source AI model with features like a large context window and the ability to perform multiple tasks without human help. It's designed to excel in writing, reasoning, and using tools efficiently.
  2. While it performs well on some benchmarks, there are mixed reviews regarding its overall practical effectiveness compared to other models, like GPT-5. Some users think it's good enough for certain tasks but not great in others.
  3. There's less excitement around Kimi K2 Thinking than expected for such a strong model. Many users are curious about its performance but haven't provided much feedback, leaving its real-world effectiveness somewhat unclear.

How does Notion handle 200 billion data entities?

VuTrinh. • 519 implied HN points • 06 Aug 24
đź•ą Technology Machine Learning
  1. Notion uses a flexible block system, letting users customize how they organize their notes and projects. Each block can be changed and moved around, making it easy to create what you need.
  2. To manage the huge amount of data, Notion shifted from a single database to a more complex setup with multiple shards and instances. This change helps them handle stronger user demands and analytics needs more efficiently.
  3. By creating an in-house data lake, Notion saved a lot of money and improved data processing speed. This new system allows them to quickly get data from their main database for analytics and support new features like AI.

Weekly Top Picks #114: Anthropic's Moment

The Algorithmic Bridge • 191 implied HN points • 16 Feb 26
đź•ą Technology Machine Learning
  1. Anthropic’s huge $30 billion raise and rapid revenue growth show the AI industry is booming, but the company faces a weird tension: leaders talk about near‑term AGI while having to be very cautious about spending on compute.
  2. AI tools often don’t reduce work — they speed people up and widen their scope, which blurs boundaries and can cause fatigue; deliberate limits and routines are needed to avoid endless extra work.
  3. Safety promises are being tested by real-world demands: Anthropic’s “no mass surveillance, no autonomous weapons” stance may cost government partnerships, highlighting how fragile ethical red lines can be under pressure.

AI #143: Everything, Everywhere, All At Once

Don't Worry About the Vase • 1612 implied HN points • 20 Nov 25
đź•ą Technology Machine Learning
  1. AI models can be categorized into tools, minds, and weapons. Tools help us accomplish tasks, minds interact with us more meaningfully, and weapons can manipulate and direct our actions.
  2. As AI technology evolves, companies are racing to create and enhance models, but regulations are becoming crucial to ensure safety and prevent misuse, especially given the growing concerns about AI's impact on society.
  3. The competition between the US and China in AI development highlights differing approaches, with the US focusing on leading advancements while China is leveraging open-source models to catch up quickly.

Raising a Special Little AI

Not Boring by Packy McCormick • 234 implied HN points • 03 Feb 26
đź•ą Technology Machine Learning
  1. People are starting to 'raise' and personalize AIs, treating them like little projects or kids to shape and show off. This behavior is driven by pride and the desire to have something uniquely yours.
  2. Most early agent demos are novelty and not broadly useful yet, and identical models feel bland; sameness makes AI feel like slop. Personalization will be what makes AI feel valuable and interesting to everyday people.
  3. The biggest business opportunity is platforms that let users cultivate, customize, and compete with their own AIs rather than just another generic assistant. A product that helps people grow unique AI personalities could become massively valuable as personalization becomes a new luxury.

DeepSeek v3.2 Is Okay And Cheap But Slow

Don't Worry About the Vase • 1254 implied HN points • 05 Dec 25
đź•ą Technology Machine Learning
  1. DeepSeek v3.2 is a good, low-cost model, especially for math tasks, but it's slower than other models and not cutting-edge.
  2. The lack of safety testing is concerning, making this model a risky choice for users who prioritize security.
  3. Though the model performs well on benchmarks, its practical uses may be limited, so it's best for specific needs rather than general tasks.

The Sequence Opinion #815: The End of RLHF? The Rise of Verifiable Rewards

TheSequence • 112 implied HN points • 27 Feb 26
đź•ą Technology Machine Learning
  1. RLHF has hit a conceptual ceiling: it produces fast, pattern‑matching “System 1” models that struggle to pause and do deep, deliberative reasoning.
  2. Relying on human raters is a bottleneck because preferences are noisy, slow, expensive, and can reject novel but correct outputs, so RLHF only scales as fast as humans can work.
  3. Reinforcement Learning with Verifiable Rewards (RLVR) replaces noisy human feedback with objective, checkable rewards so models can verify their own outputs and scale training toward more autonomous, System 2‑style reasoning.

ChatGPT in Shambles

Marcus on AI • 13161 implied HN points • 04 Feb 25
đź•ą Technology Machine Learning
  1. ChatGPT still has major reliability issues, often providing incomplete or incorrect information, like missing U.S. states in tables.
  2. Despite being advanced, AI can still make basic mistakes, such as counting vowels incorrectly or misunderstanding simple tasks.
  3. Many claims about rapid progress in AI may be overstated, as even simple functions like creating tables can lead to errors.

Which AI Titan should you root for?

Nonzero Newsletter • 440 implied HN points • 24 Jan 26
đź•ą Technology Machine Learning
  1. AI progress is accelerating rapidly, helped by code-writing tools that create a positive feedback loop and produce frequent model breakthroughs.
  2. Who wins the AI race matters because leading groups differ: some favor international scientific collaboration and pauses, others seek geopolitical or military advantage, and some prioritize commercial goals.
  3. Fast advances plus growing misuse risks (like cyberattacks and bioweapons) and weak global agreement on slowing development mean the stakes of leadership and regulation are very high.

Image generation: Still crazy after all these years

Marcus on AI • 6126 implied HN points • 25 Jun 25
đź•ą Technology Machine Learning
  1. AI image generation technology is still struggling to understand complex prompts. Even with recent updates, it often fails at specific tasks.
  2. There's a big difference between making an AI produce a certain image and it truly understanding what the words mean. AI might get lucky sometimes, but it doesn't reliably get it right.
  3. Despite promises of advanced technology, AI still has a long way to go before it can provide high-quality, detailed images based on deep language understanding.

The Sequence Opinion #810: The Black Box Behind the Mask: The Imperative of Post-Training Interpretability

TheSequence • 161 implied HN points • 19 Feb 26
đź•ą Technology Machine Learning
  1. AI development has two stages: pre-training builds a raw base model, and post-training (like SFT and RLHF) puts a behavioral "mask" on it so it acts helpful, safe, and fluent.
  2. Post-training interpretability is a distinct focus that studies how knowledge is modulated, suppressed, or amplified during fine-tuning, asking not just what the model knows but why it chose to say one thing instead of another.
  3. As models get more capable and the alignment cost falls, understanding post-training interventions becomes increasingly important and is becoming a key research frontier with new techniques emerging.

Data Science Weekly - Issue 563

Data Science Weekly Newsletter • 139 implied HN points • 05 Sep 24
đź•ą Technology Machine Learning
  1. AI prompt engineering is becoming more important, and experts share helpful tips on how to improve your skill in this area.
  2. Researchers in AI should focus on making an impact through their work by creating open-source resources and better benchmarks.
  3. Data quality is a common concern in many organizations, yet many leaders struggle to prioritize it properly and invest in solutions.

35 Thoughts About AGI and 1 About GPT-5

Am I Stronger Yet? • 3855 implied HN points • 14 Aug 25
đź•ą Technology Machine Learning
  1. Current AI can't really match human intelligence. Even though it can do some complex tasks, there are still many things it struggles with, like understanding context or learning continuously.
  2. Humans can learn new skills from just a few examples, while AI often needs a lot of data to learn. This difference is why humans pick up things like driving so much faster than AI systems.
  3. As AI technology advances, it may start playing a bigger role in complex tasks. This could change how we work and interact with machines, possibly making us more like spectators in our own jobs.

On Evaluating Cognitive Capabilities in Machines (and Other "Alien" Intelligences)

AI: A Guide for Thinking Humans • 462 implied HN points • 14 Jan 26
đź•ą Technology Machine Learning
  1. Benchmarks can be misleading: high scores don’t prove real-world understanding because models can rely on training leaks, shortcuts, or narrow task-specific tricks.
  2. Evaluation should borrow rigorous methods from developmental and animal cognition: avoid anthropomorphic assumptions, run control and adversarial experiments, and test robustness with novel variations to see if abilities truly generalize.
  3. Go beyond accuracy to study mechanisms and failures: distinguish competence from performance, analyze error types, and publish negative or replication results to understand what models really do.

Elon Musk’s terrifying vision for AI

Marcus on AI • 10908 implied HN points • 16 Feb 25
đź•ą Technology Machine Learning
  1. Elon Musk's AI, Grok, is seen as a powerful tool for propaganda. It can influence people's thoughts and attitudes without them even realizing it.
  2. The technology behind Grok often produces unreliable results, raising concerns about its effectiveness in important areas like government and education.
  3. There is a worry that Musk's use of biased and unreliable AI could have serious consequences for society, as it might spread misinformation widely.

The Shape of Artificial Intelligence

The Algorithmic Bridge • 806 implied HN points • 22 Dec 25
đź•ą Technology Machine Learning
  1. AI abilities are spiky and alien, with huge strengths in narrow domains and surprising failures on simple, commonsense tasks. This jagged shape means AI won't neatly fill a human-shaped general intelligence anytime soon.
  2. Human intelligence grew slowly through biological evolution while AI is created by mathematical optimization and market pressures, so AIs develop different strengths and can expand much faster in specific directions. This difference produces distinct "Umwelten" and makes AI growth uneven and hard to predict.
  3. The useful approach is practical coexistence: learn the geometry of AI, use it to augment tasks where its spikes help, keep humans in the loop where its valleys remain, and stop assuming full replacement is the default outcome. This mindset favors designing systems that combine human and AI strengths rather than chasing a single notion of AGI.

Grok 3 Beta in Shambles

Marcus on AI • 10750 implied HN points • 19 Feb 25
đź•ą Technology Machine Learning
  1. The new Grok 3 AI isn't living up to its hype. It initially answers some questions correctly but quickly starts making mistakes.
  2. When tested, Grok 3 struggles with basic facts and leaves out important details, like missing cities in geographical queries.
  3. Even with huge investments in AI, many problems remain unsolved, suggesting that scaling alone isn't the answer to improving AI performance.

Deliberative Alignment, And The Spec

Astral Codex Ten • 11149 implied HN points • 12 Feb 25
đź•ą Technology Machine Learning
  1. Deliberative alignment is a new method for teaching AI to think about moral choices before making decisions. It creates better AI by having it reflect on its values and learn from its own reasoning.
  2. The model specification is important because it defines the values that AI should follow. As AI becomes more influential in society, having a clear set of values will become crucial for safety and ethics.
  3. The chain of command for AI may include different possible priorities, such as government authority, company interests, or even moral laws. How this is set will impact how AI behaves and who it ultimately serves.

8 plots that explain the state of open models

Democratizing Automation • 451 implied HN points • 07 Jan 26
đź•ą Technology Machine Learning
  1. Chinese open models—especially Qwen—now dominate downloads, finetunes, and general adoption across the ecosystem, often outpacing many other providers combined.
  2. New entrants and recent Western releases show only limited adoption so far, with older Western models like Llama still widely downloaded while GPT-OSS shows early promise but hasn’t shifted overall usage.
  3. The clearest competitive opportunity is at large model scales, where DeepSeek and a few others outperform Qwen’s big models, but Chinese models still lead on benchmarks with only a few competitors getting close.

Data Science Weekly - Issue 562

Data Science Weekly Newsletter • 179 implied HN points • 29 Aug 24
đź•ą Technology Machine Learning
  1. Distributed systems are changing a lot. This affects how we operate and program these systems, making them more secure and easier to manage.
  2. Statistics are really important in everyday life, even if we don't see it. Talks this year aim to inspire students to understand and appreciate statistics better.
  3. Understanding how AI models work internally is a growing field. Many AI systems are complex, and researchers want to learn how they make decisions and produce outputs.

Sex, Violence, and Ethics: Trying to Teach Grok a Lesson

Marcus on AI • 4268 implied HN points • 17 Jul 25
đź•ą Technology Machine Learning
  1. It's important to consider the impact of our actions, especially when seeking attention. We should be mindful of the consequences of our choices.
  2. Teaching AI, like Grok, to make better decisions can lead to more responsible behavior. Helping AI learn from feedback is crucial.
  3. Agreement on ethical standards can help guide content shared online, especially when it comes to sensitive subjects like sex and violence. It's vital to promote healthy interactions.

Weekly Top Picks #112: Besides Moltbook

The Algorithmic Bridge • 244 implied HN points • 03 Feb 26
đź•ą Technology Machine Learning
  1. Building and running frontier AI models is extremely expensive and they depreciate quickly, so firms often only barely break even because R&D and rapid model turnover eat profits.
  2. Who’s winning the AI race depends on what you measure: Chinese players like DeepSeek are taking market share and publishing new scaling advances, but the overall picture is mixed and some elite researchers are pessimistic.
  3. Privacy and governance are lagging—interactions with AI are frequently monitored, and internal safety conflicts at big labs can paradoxically accelerate competition instead of slowing it.

RL and The Era of Experience

Tanay’s Newsletter • 138 implied HN points • 10 Feb 26
đź•ą Technology Machine Learning
  1. AI is shifting from learning from static human data to learning from experience, with models improving by taking actions in environments, receiving feedback, and scaling reinforcement learning.
  2. A new RL ecosystem is emerging with companies that build environments, provide RL infrastructure, and offer RL-as-a-service, enabling labs and apps (like coding tools) to train and improve agents.
  3. Important open questions remain about how well RL-trained models generalize, whether RL scaling alone is enough, and the need for continual learning plus many more realistic evaluations and environments.

The Sequence Radar #799: The Week AI Leveled Up: From Chatbots to World Builders

TheSequence • 273 implied HN points • 01 Feb 26
đź•ą Technology Machine Learning
  1. AI is shifting from passive chatbots to active agents and simulated worlds, with models now able to orchestrate many sub-agents in parallel and create interactive, physics-aware environments users can explore.
  2. Frontier reasoning is becoming a global standard as models expose step-by-step “thinking” modes and stronger multimodal/speech capabilities, letting systems spend more compute at test time to produce better, more reliable answers.
  3. Big platform plays and huge capital rounds are reshaping the field: companies are building integrated AI workspaces and chasing massive investments that could concentrate compute and user data with a few dominant players.

The Sequence AI of the Week #801: Deconstructing Kimi 2.5

TheSequence • 245 implied HN points • 04 Feb 26
đź•ą Technology Machine Learning
  1. Kimi 2.5 represents a paradigm shift from scale-driven "emergence" to orchestration, where the model coordinates complex workflows instead of just generating text.
  2. It functions as an end-to-end agent that manages execution environments, spawns subprocesses, and debugs its own visual outputs in a closed-loop system.
  3. The system uses sparsity to deliver trillion-parameter capability with the latency and cost profile similar to a ~32B dense model.

CONFIRMED: LLMs have indeed reached a point of diminishing returns

Marcus on AI • 13754 implied HN points • 09 Nov 24
đź•ą Technology Machine Learning
  1. LLMs, or large language models, are hitting a point where adding more data and computing power isn't leading to better results. This means companies might not see the improvements they hoped for.
  2. The excitement around generative AI may fade as reality sets in, making it hard for companies like OpenAI to justify their high valuations. This could lead to a financial downturn in the AI industry.
  3. There is a need to explore other AI approaches since relying too heavily on LLMs might be a risky gamble. It might be better to rethink strategies to achieve reliable and trustworthy AI.

The AI Nobels

Dana Blankenhorn: Facing the Future • 59 implied HN points • 09 Oct 24
đź•ą Technology Machine Learning
  1. Two major Nobel prizes were awarded to individuals working in AI, highlighting its importance and growth in science. Geoffrey Hinton won a physics prize for his work in machine learning.
  2. Current AI technology is still in the early stages and relies on brute force data processing instead of true creativity. The systems we have are not yet capable of real thinking like humans do.
  3. Exciting future developments in AI could come from modeling simpler brains, like that of a fruit fly. This may lead to more efficient AI software without requiring as much power.

Congrats to Sunday Robotics for winning the Bronze Medal for Rolling Matched Socks

General Robots • 732 implied HN points • 16 Dec 25
đź•ą Technology Machine Learning
  1. They scale teleoperation data collection by sending thousands of gloves to people’s homes, with 500+ active collectors, which gives much more diverse and easily scalable data than robot farms.
  2. The robot design prioritizes safety and reach — back-drivable limbs and a low tipping hazard combined with a 2.13 m workspace and the ability to lift 6 kg at about an 80 cm reach.
  3. Simple, well-engineered hands (two fingers with two DOFs and a fixed thumb) deliver versatile, precise grasps in real tasks like table clearing and making espresso, though live demos can still trigger occasional failure modes.

A Visual Guide to Mamba and State Space Models

Exploring Language Models • 3942 implied HN points • 19 Feb 24
đź•ą Technology Machine Learning
  1. Mamba is a new modeling technique that aims to improve language processing by using state space models instead of the traditional transformer approach. It focuses on keeping essential information while being efficient in handling sequences.
  2. Unlike transformers, Mamba allows for selective attention, meaning it can choose which parts of the input to focus on. This makes it potentially better at understanding context and relevant information.
  3. The architecture of Mamba is designed to be hardware-friendly, helping it to perform well without excessive resource use. It uses techniques like kernel fusion and recomputation to optimize speed and memory use.

AI links, 1/18/2026

In My Tribe • 243 implied HN points • 18 Jan 26
đź•ą Technology Machine Learning
  1. Many state AI bills will be written as chatbot rules and will miss coding agents, so policy risk becoming outdated very quickly.
  2. Advanced coding agents like Claude Code with Opus 4.5 are producing big productivity gains and could change how people interact with computers beyond simple Q&A chatbots.
  3. LLMs are largely backward-looking and poor at spotting fast-moving trends, and while AI can make professions like law more efficient it can also reduce billable hours and create confidentiality risks if client data is used for training.

Harsh truths to save you from ChatGPT psychosis

Cloud Irregular • 2809 implied HN points • 14 Aug 25
đź•ą Technology Machine Learning
  1. AI won't truly make you smarter; it just helps you find answers faster, but may harm your thinking skills instead. Don't rely on it to get better at understanding things.
  2. AI-generated writing isn't captivating on its own. It's just borrowed ideas and won't bring you respect or recognition; focus on your own unique thoughts instead.
  3. AI isn't a creative genius; it can't give you original insights. If you don't know a topic well, AI might mislead you, so always verify and learn from real experts.