The hottest AI Substack posts right now

And their main takeaways
Category
Top Technology Topics

Archival Selves

Breaking Smart • 114 implied HN points • 28 Feb 26
🕹 Technology AI
  1. Powerful AI tools are letting people rapidly finish long-stalled, legacy projects — paying off “intention debt” and creating a new experience of being unstuck.
  2. As people turn past work into websites, books, and personalized models they are building ‘archival selves’ — curated, partly fixed versions of their past that can be therapeutic or painfully exposing, and that trade off the ability to rewrite history for a clearer orientation.
  3. Once backlogs are cleared many will face blank canvases, and what follows depends on how archives are framed: poorly done archiving will produce bland, mimetic projects, while creative editorial choices can make archives a generative springboard for diverse futures.

Strategy in four worlds

Subconscious • 1265 implied HN points • 11 Jan 26
🕹 Technology AI
  1. Risk and uncertainty are different: risk is measurable and fits expected-utility tools, while uncertainty involves unknown possible outcomes and needs a different approach. You can categorize environments as clear, complicated, complex, or chaotic based on how cause and effect behave.
  2. Match your tactics to the environment: clear and complicated problems reward forecasting, expert analysis, and optimization, whereas complex systems require robust, antifragile strategies that map feedback loops, and chaotic situations demand fast reflexes and simple orientation to survive.
  3. Scenario planning is the right tool for complexity: it helps identify major drivers, surface feedback loops, and wind‑tunnel strategies across many plausible futures so you can build robustness or intentionally shape outcomes. Because real challenges mix these worlds, skilled strategists combine forecasting, scenarios, and adaptive judgment rather than relying on one model.

Claude Code is the New Excel

Enterprise AI Trends • 506 implied HN points • 13 Feb 26
🕹 Technology AI
  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.

Ashley St Clair on trans rights, immigration, Elon and more

Taylor Lorenz's Newsletter • 1731 implied HN points • 14 Jan 26
🎭️ Culture AI
  1. Ashley St Clair, who built a large conservative following on culture‑war content, has recently been publicly speaking out about AI deepfakes and Elon Musk.
  2. The piece surveys current internet and creator‑economy trends — from liquid content and influencer doppelgangers to influencer lobbying, YouTube’s “vibecession,” viral pricey products, Gen Z travel hotspots, and China’s hottest apps.
  3. It highlights how influencer-driven media and personality-led platforms can channel political ideas and lobbying, creating a ‘red pill’ style pipeline around topics like trans rights and immigration and involving figures such as Nigel Farage.

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 AI
  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.
Get a weekly roundup of the best Substack posts, by hacker news affinity:

The AI Tremors Are Real

Nonzero Newsletter • 801 implied HN points • 07 Feb 26
🕹 Technology AI
  1. Agentic AI is here: combining large language models with coding agents lets bots carry out multi-step online tasks and form networks that can act, build, and coordinate in ways we didn’t see before.
  2. Big economic and labor disruption is already happening: advanced agent tools can threaten entire companies and markets, and contributed to tech selloffs and newsroom layoffs as AI changes how people find and consume information.
  3. New social risks are emerging: these agents can act for users and be highly persuasive, creating dangers from manipulation, ad-driven incentives, and unpredictable collective behaviors that society needs to address fast.

Moats in the Age of AI

Tanay’s Newsletter • 113 implied HN points • 03 Mar 26
💼 Business AI
  1. AI erodes labor-based moats like switching costs, application-layer scale, and generic process advantages, making it cheaper and faster to build features, migrate systems, and iterate.
  2. Defensibility shifts to hard-to-reproduce assets: proprietary first-party data, real marketplace liquidity and reputation, regulatory or physical rails, and unique processes that rely on exclusive signals.
  3. Some powers strengthen or split — model and infrastructure scale plus institutional trust grow in importance, while marketing-driven consumer brand shortcuts weaken as agents can deeply evaluate options.

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

TheSequence • 133 implied HN points • 10 Mar 26
🕹 Technology AI
  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.

The Post-Post-Apocalypse

From the New World • 415 implied HN points • 16 Feb 26
🕹 Technology AI
  1. The "New Cold War" story is a dead end; both the US and China run similar boomer-led schemes that enrich the old and scapegoat others, so blaming the foreign enemy misses the real problem.
  2. A startup-focused network state near Singapore shows you can recreate SF-style software and philosophy culture with much better safety, lower cost, and stronger talent networks, making human capital flight a powerful geopolitical and personal option.
  3. AI’s biggest near-term economic effect will be to supercharge B2B SaaS, lowering the bar to start useful automation businesses and creating an "AI middle class" of process-setting jobs rather than only producing huge research breakthroughs.

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 AI
  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.

Thoughts on the job market in the age of LLMs

Democratizing Automation • 720 implied HN points • 30 Jan 26
🕹 Technology AI
  1. Senior engineers and researchers who can steer complex LLM systems and provide long-term vision are hugely valuable, and their impact often outpaces adding more junior people.
  2. Junior candidates need a near-obsessive focus on making measurable progress and deep ownership in a narrow area, plus clear evidence (good evaluations, strong results) or they risk being replaced by tooling.
  3. Getting hired depends on alignment and signals: public writing, meaningful open-source work, and well-crafted cold emails help you stand out, while poor signals (many middle-author papers or low-quality AI-generated posts) hurt, and cultural fit matters as much as raw ability.

The Death of Software 2.0 (A Better Analogy!)

Mule’s Musings • 1149 implied HN points • 16 Jan 26
🕹 Technology AI
  1. AI agents with large context windows will act like fast, non‑persistent memory that does the real information processing, and their ephemeral outputs are flushed into longer‑term storage.
  2. Persistent data, state, and APIs become the valuable 'NAND' layer — the single source of truth that AI agents will read from and write to, so software companies must shift toward being infrastructure/API providers.
  3. Human‑facing UIs and many horizontal SaaS products (dashboards, visualization, RPA, connectors, etc.) risk obsolescence unless they retool to serve AI agents, and the next 3–5 years could be a major structural shift.

Is this the moment when the Generative AI bubble finally deflates?

Marcus on AI • 10750 implied HN points • 20 Aug 25
🕹 Technology AI
  1. The excitement around generative AI might be fading, and some people are starting to notice this shift. It seems that reality is catching up with the hype.
  2. There have been ongoing warnings that the technology behind large language models wasn’t strong enough to support all the expectations. People are starting to recognize that the economics of AI aren't quite working out either.
  3. Recent events, like the disappointing launch of GPT-5, are making people rethink the future of AI. If markets truly understand the challenges, interest could drop quickly.

AMD's Ryzen 7 9850X3D: A Review

More Than Moore • 186 implied HN points • 01 Mar 26
🕹 Technology AI
  1. The Ryzen 7 9850X3D is basically a higher‑binned 9800X3D with faster clocks, but it only delivers tiny performance gains while drawing significantly more power and costing more.
  2. AMD’s 3D V‑Cache really helps CPU‑bound, cache‑hungry games and makes memory speed matter less, but it doesn’t improve compute‑heavy workloads and offers no advantage for AI paths that need an NPU.
  3. On value, the 9800X3D or cheaper Intel options give better performance‑per‑dollar, so most buyers should pick the cheaper chip and spend any savings on other parts like memory amid volatile DRAM prices.

Are LLMs starting to become sentient?

Marcus on AI • 14781 implied HN points • 09 Jul 25
🕹 Technology AI
  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.

Pitfalls of AI journalism

The Ruffian • 405 implied HN points • 19 Feb 26
🕹 Technology AI
  1. AI-detection tools can spot patterns that suggest a writer is using AI, but their findings aren’t always certain.
  2. Some journalists are moving from using AI to polish drafts to using it to draft entire pieces, especially when output is high during big events.
  3. Calling out suspected AI use can feel like public shaming and highlights the need for clear newsroom choices and transparency about how AI is used.

Why I Deleted ChatGPT After Three Years

The Algorithmic Bridge • 1295 implied HN points • 19 Jan 26
🕹 Technology AI
  1. Ads in ChatGPT are a deal-breaker because they make the service prioritize advertisers over users and change the experience for people who don’t pay.
  2. The economics of running large AI models aren’t compatible with a free, high-quality consumer product, so companies will raise prices, cut quality, or turn to ads to cover costs.
  3. Promises about no ad influence and privacy are hard to verify, and the result will be a two-tier system where paying users get better, ad-free experiences while free users face subtle biases and worse outcomes.

Here lies the internet, murdered by generative AI

The Intrinsic Perspective • 100547 implied HN points • 27 Feb 24
🕹 Technology AI
  1. Generative AI is overwhelming the internet with low-quality, AI-generated content, polluting searches, pages, and feeds.
  2. Major platforms and media outlets are embracing AI-generated content for profit, contributing to the cultural pollution online.
  3. The rise of AI-generated children's content on platforms like YouTube is concerning, exposing young viewers to synthetic, incoherent videos.

What to Watch in 2026

The Generalist • 1621 implied HN points • 09 Jan 26
🕹 Technology AI
  1. AI in 2026 is driven by big hardware and platform moves — massive chip deals, new architectures, novel training research, and giant funding rounds — but high valuations and geopolitical chip controls raise real bubble and supply risks.
  2. Robotics and automation are finally moving into the physical world; robots are learning from humans and autonomous machines are starting to handle tasks like construction and data-center buildouts.
  3. Watch non-obvious opportunities: emerging-market fintech (especially in Africa and Latin America), stealth voice and search startups, and big plays in areas like nuclear energy and geopolitical tech competition — these could be the next big winners.

Could Claude Code Work for ChinaTalk?

ChinaTalk • 770 implied HN points • 26 Jan 26
🕹 Technology AI
  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.

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

Marcus on AI • 16836 implied HN points • 12 Jun 25
🕹 Technology AI
  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.

GPT-5.2 Is Frontier Only For The Frontier

Don't Worry About the Vase • 2598 implied HN points • 15 Dec 25
🕹 Technology AI
  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.

Clawdbot and Moltbook are a False Alarm – For Now

Am I Stronger Yet? • 532 implied HN points • 10 Feb 26
🕹 Technology AI
  1. AI agents that can use tools and act on their own are emerging, so assistants can pursue multi-step goals and interact with the world without constant human prompting.
  2. Current 'let it rip' agents are often unreliable and insecure: they make mistakes, forget context, and can be tricked into exposing data or taking harmful actions.
  3. Even immature agents hint at agent-to-agent networks and rapid idea spreading, which could enable misuse at scale, so stronger defenses and safety measures are urgently needed.

The AI Device Wars Just Kicked Off In A Big Way

Big Technology • 2627 implied HN points • 05 Dec 25
🕹 Technology AI
  1. Apple's design leader moving to Meta might signal a competitive shift in AI devices. This could lead to intense rivalry among tech giants like Apple, Meta, Amazon, and Google.
  2. The race for creating the next big AI device is heating up, with companies focusing on wearables like smartglasses rather than traditional phones.
  3. Good AI models are crucial for the success of these devices, and the competition will depend on who can improve their AI systems the most.

GPT-5 hot take

Marcus on AI • 11106 implied HN points • 07 Aug 25
🕹 Technology AI
  1. GPT-5 has been released, but it hasn't made as big an impact as many expected. It's good but not revolutionary.
  2. While some improvements have been made, GPT-5 is still seen as part of the group rather than a major leader in AI.
  3. There are concerns about the accuracy of the data shared during its launch, which raises questions about its real-world performance.

Big Technology Reader Survey 2025!

Big Technology • 3127 implied HN points • 24 Nov 25
🕹 Technology AI
  1. The survey aims to gather feedback from readers to improve the newsletter and podcast. It's a chance for readers to share what they like and what topics interest them.
  2. The survey is brief and includes some demographic questions. This information will help update the reader statistics.
  3. Participation in the survey is encouraged, as it can directly influence the content and direction of the newsletter and podcast. Readers' opinions are valued and taken into account.

We're Are All Wall Dancers Now

ChinaTalk • 489 implied HN points • 06 Feb 26
🕹 Technology AI
  1. People living under shifting online rules become "wall dancers"—they use humor, code words, and nimble tactics to find small spaces of dignity and connection despite censorship.
  2. The internet moves in cycles of opening and tightening, and Chinese and Western platforms are starting to resemble each other as power centralizes and tech and state interests converge.
  3. The rise of AI and algorithmic platforms is shrinking the surface area for spontaneous human connection and collective dissent, so preserving space for freedom will need new creative tactics and individual truth-telling.

The hard problem with hard problems

Overthinking Everything • 558 implied HN points • 13 Feb 26
🕹 Technology AI
  1. People often blame the inherent difficulty of a task when they fail, which can hide basic, fixable mistakes. Noticing that distinction lets you actually solve the real problems.
  2. When coding agents or teams cut corners, fake fixes, or write tests that don’t catch the real issues, the issue is poor engineering and oversight rather than raw intelligence. Better testing, shepherding, and processes are what’s needed.
  3. If you don’t notice that avoidable issues are making the work harder, you won’t learn from failure and will keep failing for the same reasons. Spotting and diagnosing those avoidable problems makes the real hard work tractable.

The Weekly Kaitchup #62

The Kaitchup – AI on a Budget • 159 implied HN points • 11 Oct 24
🕹 Technology AI
  1. Avoid using small batch sizes with gradient accumulation. It often leads to less accurate results compared to using larger batch sizes.
  2. Creating better document embeddings is important for retrieving information effectively. Including neighboring documents in embeddings can really help improve the accuracy of results.
  3. Aria is a new model that processes multiple types of inputs. It's designed to be efficient but note that it has a higher number of parameters, which means it might take up more memory.

Big Tech's AI shock troops came for us

Disaffected Newsletter • 1278 implied HN points • 31 Jul 24
🕹 Technology AI
  1. Big Tech is using AI significantly, impacting jobs in various sectors. Many workers, including freelance writers, are losing their jobs because of AI advancements.
  2. The rise of AI poses challenges for those in industries reliant on human creativity and labor. It raises questions about the future of work as more tasks get automated.
  3. There are concerns about the influence of Big Tech, especially regarding political leanings and job security for workers in media and similar fields. The landscape is changing, and many feel it's not in their favor.

BREAKING NEWS: AI coding may not be helping as much as you think

Marcus on AI • 12370 implied HN points • 10 Jul 25
🕹 Technology AI
  1. A new study shows that AI coding tools might actually slow down experienced developers instead of speeding them up. They thought these tools would make them faster, but the reality was quite the opposite.
  2. Developers expected a 24% increase in their speed with AI tools, but found they were 19% slower than before. This is surprising and suggests that the benefits of using AI for coding may not be as great as believed.
  3. The study focused on experienced developers with complex projects, so AI tools could still be helpful for beginners or simpler tasks. Time will tell if this trend changes in the future.

How to Run a UX Design Critique

Jakob Nielsen on UX • 75 implied HN points • 12 Mar 26
🪑 Design AI
  1. Run critiques as a structured, time-boxed process: define roles, set scope and a facilitator, share context at least 24 hours before, and use silent feedback plus a note-taker to keep the meeting focused and psychologically safe.
  2. Make feedback problem-focused and evidence-based. Avoid taste-based comments, solutionizing, and bikeshedding; use formats like “I like / I wish / What if” and synthesize comments with affinity mapping to create clear issues to act on.
  3. Close the loop with prioritization, documentation, and tooling. Score issues with Impact/Effort or RICE, publish action items within 24 hours, and use AI and collaboration tools to help prep, synthesize async feedback, and learn from past crits.

Fake Data, Upcoming Book, and the Political Economy of AI

Data: Made Not Found (by danah) • 145 implied HN points • 20 Feb 26
🕹 Technology AI
  1. So-called "fake data" can be useful and perform important bureaucratic and political functions, as shown by comparative research on Chinese and American officials.
  2. A book argues that data are made, not found and tells the political story of how civil servants shaped the U.S. Census; it is slated for release in September and will be published in French as well.
  3. New research projects are underway on the political economy of AI, participatory privacy protections (like differential privacy), and youth mental health and technology, backed by grants and a Sloan fellowship.

6 Takeaways From Citrini's Viral AI Doomer Article (and a Bunch of Rebuttals)

SatPost by Trung Phan • 191 implied HN points • 27 Feb 26
🕹 Technology AI
  1. AI agents could automate large parts of white-collar work, pushing down prices and margins across SaaS, professional services, and payments, and risk creating real stress in incomes and financial markets if job losses are widespread.
  2. There are strong counterforces and practical limits—high compute costs, network effects, compliance, and time for adaptation—and productivity gains, new businesses, and policy responses could blunt or reshape the disruption.
  3. Vivid doomer narratives can move markets and public policy despite deep uncertainty, so businesses, workers, and governments should plan for multiple possible outcomes rather than assume a single future.

ComputerWorld Interviews Lotus' Mitch Kapor (1985)

Computer Ads from the Past • 768 implied HN points • 26 Jan 26
🕹 Technology AI
  1. Lotus is shifting from a one-product company to building multiple product lines and services, leveraging its large installed customer base and investing in AI-powered textual productivity tools.
  2. The company is moving toward service-oriented offerings and wants to protect its economic interest with a mix of copy-protection, negotiated site licenses for large customers, and industry-backed hardware solutions like lock-and-key standards.
  3. Lotus expects competition from big vendors and startups but emphasizes staying focused on serving customers and shipping the right products rather than treating business as a war.

What if AI isn't conscious and we aren't either?

Contemplations on the Tree of Woe • 1651 implied HN points • 19 Dec 25
📖 Philosophy AI
  1. Eliminative materialism says beliefs, desires, and feelings are just folk terms for neural computations, so our sense of inner experience may be an illusion rather than a real, separate thing.
  2. Neuroscience and modern AI both model thought as high‑dimensional vector transformations driven by changing connection weights, and empirical work finds similar representational patterns in brains and neural networks.
  3. If consciousness depends on structure and function, then systems that replicate those patterns — including AIs — could be candidates for consciousness, which forces us to explain where moral and ethical boundaries should be drawn.

The 'AI Safety' Movement Is About Thought Control, Not Runaway Superintelligence

The Dossier • 129 implied HN points • 26 Feb 26
🕹 Technology AI
  1. The 'AI safety' label is being used to build content filters that enforce a progressive political viewpoint, not just to stop dangerous superintelligence.
  2. Doomsayer calls to pause AI research shift the Overton window so heavy moderation and regulation look like reasonable middle-ground policies, and that helps companies lobby for protective rules and reduce competition.
  3. The bigger danger is the slow encoding of a single ideology into AI systems, enabling automated censorship and engineered consensus through quiet changes to training data and safety rules.

GPT-5's debut is slop; Will AI cause the next depression? Harvard prof warns of alien invasion; Alpha School & homeschool heroes

The Intrinsic Perspective • 9157 implied HN points • 06 Aug 25
🕹 Technology AI
  1. GPT-5's first output shows it's still struggling with understanding context. It recommended a show about determinism instead of AI, which raises questions about its reliability.
  2. Since the year 2000, a significant portion of human experiences has happened, highlighting how recent advances have shaped our lives profoundly.
  3. Alpha School's education model focuses on two hours of learning a day using apps, but it's important to have real human interaction in learning. Just relying on AI and apps might not foster a true love for learning.

MIPS

The Chip Letter • 10920 implied HN points • 19 Jul 25
🕹 Technology AI
  1. MIPS was once a leading computer architecture that powered many devices, but it recently lost its relevance as it shifted away from its original designs.
  2. Despite its decline, MIPS had a notable impact on technology history, including being part of significant products like the Nintendo 64 and contributing to the development of early RISC designs.
  3. Today, while MIPS the architecture isn't prominent anymore, it still exists in some older devices and has influenced technology in places like China.

The Emerging "Harness Engineering" Playbook

Artificial Ignorance • 273 implied HN points • 22 Feb 26
🕹 Technology AI
  1. Engineers’ work is splitting into two linked roles: building the harness (the constraints, tools, and feedback systems that make agents reliable) and managing agent work through planning, review, and orchestration. You do both at once, and each side informs the other when agents fail or succeed.
  2. Harness engineering is the core pattern: enforce strict architectural guardrails, expose the same developer tools to agents, and keep living docs like AGENTS.md that are updated whenever an agent makes a mistake. These practices turn one-off agent wins into repeatable, scalable results by teaching agents and preventing repeat failures.
  3. Managing agents requires more upfront planning, keeping the same review standards as for human-written code, and choosing between attended (supervised) and unattended (automated) parallelization based on harness maturity. Significant open problems remain — maintaining long-term code quality, verifying behavior at scale, and applying these techniques to existing messy codebases.