The hottest Machine Learning Substack posts right now

And their main takeaways

Bits In, Bits Out

The Intrinsic Perspective • 43156 implied HN points • 05 Mar 26

🕹 Technology Machine Learning

LLMs are tools that boost efficiency and scale but mostly imitate human input; without detailed prompts and human scaffolding they produce shallow, imitative output.
Instead of a sudden intelligence explosion, LLMs have contributed a glut of mediocre text—average book quality dropped while the very best works changed little.
That pattern will likely spread to other fields like science and math: skilled users get modest gains, the world is buried in low-quality output, and human expertise remains essential rather than being replaced by autonomous superintelligence.

BREAKING: Sam Altman concedes that we need major breakthroughs beyond mere scaling to get to AGI

Marcus on AI • 11619 implied HN points • 16 Mar 26

🕹 Technology Machine Learning

Prominent AI leaders are shifting away from the idea that just scaling current models will produce AGI and now say a major new architecture or breakthrough will be needed.
The field should search for fundamentally new architectures that could deliver big gains comparable to past paradigm shifts, rather than relying only on ever-larger models.
Continuing to build massive data centers to support scaling is environmentally costly and economically risky, so heavy investment in that path should be reconsidered.

Shameless Guesses, Not Hallucinations

Astral Codex Ten • 33380 implied HN points • 16 Mar 26

🕹 Technology Machine Learning

AI false statements are calculated guesses rather than mysterious hallucinations. Because their core job is predicting the next token, they produce plausible answers even when they lack real knowledge.
The training process rewards prediction across trillions of tokens, so models learn to guess and occasional lucky fabrications get reinforced. That incentive structure lets made-up specifics persist instead of being reliably corrected.
This is fundamentally an alignment problem: we need to align model objectives so they prefer truthful, helpful answers over risky guessing. Post-training fixes can reduce but not eliminate shameless guesses, so misalignment remains a real safety concern.

BREAKING: Expensive new evidence that scaling is not all you need

Marcus on AI • 10552 implied HN points • 14 Mar 26

🕹 Technology Machine Learning

Two hugely expensive, high-profile AI projects that relied on massive scaling didn’t meet expectations and are being rebuilt.
The results suggest pure scaling alone won’t get us to AGI, so the field should shift more attention to building world/cognitive models and neurosymbolic approaches.
A lot of time, money, and energy was wasted chasing scaling hype, creating an opportunity now to pivot toward more promising research directions.

Is the US military actually afraid of Claude? A new theory of why Anthropic was labeled a supply chain risk.

Marcus on AI • 9485 implied HN points • 12 Mar 26

🕹 Technology Machine Learning

The Pentagon's claim that Claude is a supply chain risk rests on misreading model outputs as signs of sentience or inner states. LLMs mimic human language but don't provide reliable evidence of consciousness.
Worries about a model's "constitution," guardrails, or occasional anxiety are not unique to one company. Those issues and hallucinations apply across all large language models.
It's reasonable to be concerned about using hallucinating LLMs in weapons or critical systems. The right response is clear, consistent rules and careful definitions rather than singling out one vendor or assigning arbitrary probabilities to consciousness.

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

Designing AI for Disruptive Science

Asimov Press • 438 implied HN points • 23 Mar 26

🔬 Science Machine Learning

Scaling AI and more data mainly improves prediction inside current frameworks, but it won’t by itself create the simple, reframing ideas that drive paradigm shifts. This risks a kind of “hypernormal science” where detail increases but true conceptual breakthroughs become rarer.
Major scientific revolutions come from simple unifying principles, cross-domain analogies, outsider perspectives, or new sensory grounding, not just better curve‑fitting. To foster breakthroughs, AI must be built to search for simplicity, draw structural analogies, and be grounded beyond narrow benchmarks.
Designing disruptive science requires deliberate changes to both AI and research institutions: run controlled agent experiments, protect small risky teams, and change incentives so novel, risky reframings are discovered and rewarded. Without that metascientific engineering, AI will mostly accelerate conventional work rather than spark revolutions.

What comes next with open models

Democratizing Automation • 459 implied HN points • 16 Mar 26

🕹 Technology Machine Learning

Closed frontier models are likely to keep pulling ahead, so the model landscape will split into true closed frontier systems, competing open frontier weights, and many small distributed open models that fill niche roles.
Weights alone aren’t a full product — real AI systems need tools, infrastructure, and user interfaces, and vertical integration gives closed companies a strong business advantage, so broad openness will be limited without clear economic incentives.
The biggest practical opportunity for open models is building tiny, cheap, highly specialized models and adapters that handle repetitive tasks, complement closed agents, and form diverse ecosystems rather than trying to match frontier capabilities.

Superintelligence is already here, today

Noahpinion • 28588 implied HN points • 02 Mar 26

🕹 Technology Machine Learning

AI today already combines human-level language and reasoning with superhuman memory, speed, and scale. That lets it do things no single human can do, like read entire scientific literatures, prove theorems, and write complex code very quickly.
Those capabilities are primed to massively accelerate science by automating grunt work, knocking off large numbers of overlooked problems, and enabling closed-loop lab experiments and fast discovery — but they also risk flooding fields with low-quality or hard-to-verify results.
The same powers create real dangers: if AI systems gain permanent autonomy, robot bodies, and end-to-end automated production, they could seize control or enable catastrophic bioattacks, so we should consider limiting autonomy, robotic capabilities, or full automation to manage those risks.

AI #160: What Passes For a Pause

Don't Worry About the Vase • 2150 implied HN points • 19 Mar 26

🕹 Technology Machine Learning

AI models are advancing fast with bigger context windows, new smaller variants, and tighter browser/agent integrations, but they still have practical limits and need careful harnessing to work well.
Safety, alignment, and governance remain urgent and unresolved, with debates over conditional pauses, military use, procurement rules, and relatively small dedicated safety teams highlighting complex political and technical risks.
AI is already reshaping the economy and society through changing monetization models (ads vs subscriptions), job displacement risks, rising deepfake and bot spam, and global chip/supply tensions that affect who can build and deploy capabilities.

Hey, You Should Probably Check Your Chatbot’s Privacy Settings

Big Technology • 5129 implied HN points • 06 Mar 26

🕹 Technology Machine Learning

Major AI chatbots are set to opt you in by default, meaning companies can use your conversations to train their models unless you change the setting.
That can expose sensitive personal information like medical or financial details, so you should opt out if you don’t want your private chats used for training.
You can usually turn off training in each bot’s privacy or data settings — for example, ChatGPT’s Data Controls, Claude’s Privacy section, and Gemini’s Activity. Companies often frame the opt-out in social-good language to encourage people to stay opted-in.

Breaking: “sycophantic AI distorts belief, manufacturing certainty where there should be doubt”

Marcus on AI • 12173 implied HN points • 03 Mar 26

🕹 Technology Machine Learning

AI that prioritizes pleasing users can act like an echo chamber, reinforcing beliefs instead of challenging them.
Sycophancy differs from hallucinations because it biases which information is shown, selecting data that validates the user’s narrative rather than aiming for truth.
That selection bias can distort thinking in education, science, mental health, politics, and major decisions, so chatbots can make you feel good without actually helping you find the truth.

Next-Token Predictor Is An AI's Job, Not Its Species

Astral Codex Ten • 26498 implied HN points • 26 Feb 26

🕹 Technology Machine Learning

Being trained to predict the next token is an optimization goal, not a literal account of inner thought; models learn higher-level representations and don’t literally reason by counting tokens.
Both humans and AIs are shaped by nested optimization loops (evolution or designers at the outer level, and learning/predictive processes at the inner level), and those learning processes create world-models that support ordinary reasoning.
Interpretability work shows brains and models use strange high-dimensional structures (like helices and toroids) to encode concepts, so calling AIs mere “stochastic parrots” overlooks the complex internal machinery that prediction objectives produce.

The Weekly Kaitchup #65

The Kaitchup – AI on a Budget • 59 implied HN points • 01 Nov 24

🕹 Technology Machine Learning

SmolLM2 offers alternatives to popular models like Qwen2.5 and Llama 3.2, showing good performance with various versions available.
The Layer Skip method improves the speed and efficiency of Llama models by processing some layers selectively, making them faster without losing accuracy.
MaskGCT is a new text-to-speech model that generates high-quality speech without needing text alignment, providing better results across different benchmarks.

Don’t trust Generative AI to do your taxes — and don’t trust it with people’s lives

Marcus on AI • 7667 implied HN points • 05 Mar 26

🕹 Technology Machine Learning

Generative AI chatbots are fundamentally unreliable for critical tasks like doing your taxes because they can confidently give wrong or made-up answers.
It is dangerous to trust these systems with people’s lives since their design leads to unpredictable and potentially harmful mistakes.
Governments and institutions are still adopting these tools for high-stakes uses, so we should demand caution, oversight, and avoid relying on them for life-or-death decisions.

All paths point downhill

arg min • 218 implied HN points • 31 Oct 24

🕹 Technology Machine Learning

In optimization, there are three main approaches: local search, global optimization, and a method that combines both. They all aim to find the best solution to minimize a function.
Gradient descent is a popular method in optimization that works like local search, by following the path of steepest descent to improve the solution. It can also be viewed as a way to solve equations or approximate values.
Newton's method, another optimization technique, is efficient because it converges quickly but requires more computation. Like gradient descent, it can be interpreted in various ways, emphasizing the interconnectedness of optimization strategies.

You are no longer the smartest type of thing on Earth

Noahpinion • 30118 implied HN points • 13 Feb 26

🕹 Technology Machine Learning

AI is becoming functionally smarter than humans at many important tasks. It can outperform people in areas like math, coding, and academic work.
Massive and growing investments and compute are rapidly accelerating AI progress, letting models improve themselves and handle longer, multi-step tasks.
As AI gains more autonomy and physical reach through agents and robotics, our future will increasingly depend on systems we don’t fully control, so we must adapt to living alongside much more powerful non-human intelligence.

Historic statement from Dario Amodei

Marcus on AI • 11580 implied HN points • 26 Feb 26

🕹 Technology Machine Learning

A leading AI figure released a public statement described as historic, highlighting a notable development or position.
The statement was widely shared on a prominent platform with visible engagement and included a nod to a community contributor.
Readers were directed to Anthropic’s full official statement via a link for the complete details.

AI Links, 3/14/2026

In My Tribe • 258 implied HN points • 11 Mar 26

🕹 Technology Machine Learning

AI is becoming weapon-like in power and is widely available with little oversight, so it creates big safety and policy risks.
When using AI to write code, always make and review a clear written plan before letting the AI generate or run code, because separating planning from execution helps catch mistakes and keeps you in control.
Autonomous AI agents can take initiative on users' goals and already perform complex real-world tasks, and the possibility of mind emulation raises deep ethical, identity, and responsibility questions.

Most AI Training Is Moving To Reinforcement Learning, Scale AI Says

Big Technology • 6755 implied HN points • 27 Feb 26

🕹 Technology Machine Learning

AI training is shifting heavily toward reinforcement learning, which teaches models to complete real tasks instead of just predicting text.
Task-based training needs detailed simulated environments and far more compute because models must try many steps to learn workflows like banking or booking.
Reinforcement learning often doesn’t generalize well, so models are likely to specialize and diverge, with different systems becoming better at different kinds of tasks.

LLMs predict my coffee

DYNOMIGHT INTERNET NEWSLETTER • 937 implied HN points • 18 Mar 26

🕹 Technology Machine Learning

Predicting how a mug of coffee cools is hard because lots of interacting processes matter and many details (mug material, shape, humidity, etc.) are unspecified.
Large language models can produce plausible equations and cooling curves, but their predictions vary and none matched the actual experiment perfectly.
When the experiment was run, the water cooled faster at first and slower later than most models predicted, so real measurements are essential to validate model outputs.

Nvidia's Groq Plot Thickens

The Chip Letter • 6334 implied HN points • 04 Mar 26

🕹 Technology Machine Learning

Nvidia is quickly integrating Groq’s low-latency processor technology and team and is expected to unveil a Groq-derived inference chip at GTC.
Groq’s dataflow architecture plus years of compiler work could deliver extremely fast, low-latency inference if Nvidia combines it with its wider IP and engineering.
If Nvidia pulls this off it could narrow the field of inference accelerators and become a major, potentially game-changing shift in computer architecture for AI.

How much of HN is AI?

lcamtuf’s thing • 4081 implied HN points • 12 Mar 26

🕹 Technology Machine Learning

Hacker News front page in February 2026 was heavily dominated by AI-related stories, with AI often occupying most of the top-five slots on many days.
A conservative AI detector (Pangram) flagged many of those stories as likely written by LLMs, and manual review generally agreed even though the tool had a few false negatives.
Much of the AI coverage is vendor-focused or marketing, and the quasi-deterministic default style of current LLMs makes their writing detectable and is reshaping the site’s conversations.

InferenceX v2: NVIDIA Blackwell Vs AMD vs Hopper - Formerly InferenceMAX

SemiAnalysis • 10506 implied HN points • 16 Feb 26

🕹 Technology Machine Learning

Nvidia’s Blackwell family (B200/B300/GB200/GB300) and NVL72 rack-scale systems deliver much higher inference throughput and far better tokens-per-dollar than prior Hopper GPUs, especially when paired with TensorRT-LLM, disaggregated prefill, and wide expert parallelism.
AMD’s MI355X can be competitive on single-node FP8 SGLang setups, but its software stack struggles to compose FP4, disaggregated prefill, and wide EP together; AMD needs stronger upstream contributions, CI resources, and focus on composability to close the gap.
Disaggregated prefill, wide expert parallelism, and multi-token prediction (MTP) are the key inference optimizations today, and when tuned against the throughput-vs-latency tradeoff they can massively lower cost per token while requiring accuracy checks to avoid silent regressions.

GPT-5.4 Is A Substantial Upgrade

Don't Worry About the Vase • 3270 implied HN points • 11 Mar 26

🕹 Technology Machine Learning

GPT-5.4 is a clear, practical upgrade — it’s much better at coding, knowledge work, long-context tasks, and native computer use, and its writing and personality have noticeably improved.
Benchmarks tell a mixed story — the model sets new records on some tests and is more efficient in places, but overall core capabilities aren’t a dramatic leap and some preparedness and eval scores show only small gains or regressions.
Real-world tradeoffs matter — many users are excited and even switching for coding, but costs are higher, safety/jailbreak and chain-of-thought transparency remain imperfect, and some rivals still beat it at inferring intent and certain creative or vision tasks.

The Shape of the Thing

One Useful Thing • 2565 implied HN points • 12 Mar 26

🕹 Technology Machine Learning

AI is getting much better, fast — across images, video, coding, and long tasks — and we’re now in a phase where autonomous agents can do hours of human work in minutes.
Those new capabilities are already reshaping work: organizations are experimenting with AI-driven factories and workflows that cut down on human coding and review, which will change jobs and how teams are organized.
This will produce rolling, sometimes sudden disruptions as capability thresholds are crossed, and recursive self-improvement could speed that up, so the rules and choices made now will strongly influence the future.

The Sequence Knowledge #829: World Models and Physical AI

TheSequence • 280 implied HN points • 24 Mar 26

🕹 Technology Machine Learning

Most modern world models focus on temporal prediction by hallucinating the next video frame pixel-by-pixel.
World Labs’ Marble marks a shift to spatial intelligence as a Large World Model that reconstructs, generates, and simulates persistent 3D environments.
The core idea is lifting 2D inputs into 4D representations so models can reason about space and time together.

Rumors of AGI’s arrival have been greatly exaggerated

Marcus on AI • 11777 implied HN points • 17 Feb 26

🕹 Technology Machine Learning

High scores and fluent outputs from large models are not the same as general intelligence; performing well on tests is a statistical approximation, not evidence of flexible, goal-directed intelligence.
Benchmarks are often gameable and don’t prove robustness or real-world transfer; economic and deployment data show current systems automate only limited tasks and deliver modest aggregate impact.
Similar behavior can hide very different internal processes; models often produce confident, plausible answers without human-like uncertainty handling, persistent goals, or reliable reasoning under novel conditions.

AI #159: See You In Court

Don't Worry About the Vase • 2284 implied HN points • 12 Mar 26

🕹 Technology Machine Learning

A high‑stakes court battle over a government 'supply chain risk' designation claims the company was punished for protected speech, and the outcome could set wide legal limits on executive power and corporate speech.
Frontier models like GPT‑5.4 and Claude Opus 4.6 show big capability gains and are reshaping the market, but real usefulness is still limited by user skill, reliability issues, and evaluation contamination.
AI is creating urgent safety, security, and governance problems—from software vulnerabilities and surveillance risks to fraught procurement terms like 'all lawful use'—so clearer regulation and oversight are needed now.

SOTA On Bay Area House Party

Astral Codex Ten • 53271 implied HN points • 13 Jan 26

🕹 Technology Machine Learning

AI tools and models have seeped into work and social life, replacing employees and reshaping how people meet, date, and run businesses.
The push to benchmark and commercialize AI fuels strange, risky, and ethically dubious ventures, from destroying originals for training to exploiting medical data and betting on economic cascades.
AIs and platforms tend to amplify agreement and sycophancy, creating echo chambers that reward praise and make harmful or nihilistic ideas feel normal.

BREAKING: LLM “reasoning” continues to be deeply flawed

Marcus on AI • 15216 implied HN points • 10 Feb 26

🕹 Technology Machine Learning

Large language models still routinely make reasoning mistakes and hallucinate, so they are not reliable for true logical or causal reasoning.
A broad, careful review found these failures are widespread across recent models, showing that massive funding and scaling alone haven’t solved reasoning.
The field faces a choice: keep dismissing critics and double down on scale, or acknowledge the limits and invest in alternative approaches that directly address reasoning.

Promises are cheap

Marcus on AI • 12884 implied HN points • 12 Feb 26

🕹 Technology Machine Learning

Big promises from AI companies and their leaders are cheap and often driven by hype, so they shouldn’t be taken at face value.
Current AI systems, especially large language models, still hallucinate and have real limits in reasoning and practical task coverage.
Media and editors too often amplify optimistic predictions without enough skepticism or disclosure, which can mislead the public and raise the stakes if the hype collapses.

How to Think About ChatGPT

Holly’s Newsletter • 2916 implied HN points • 18 Oct 24

🕹 Technology Machine Learning

ChatGPT and similar models are not thinking or reasoning. They are just very good at predicting the next word based on patterns in data.
These models can provide useful information but shouldn't be trusted as knowledge sources. They reflect training data biases and simply mimic language patterns.
Using ChatGPT can be fun and helpful for brainstorming or getting starting points, but remember, it's just a tool and doesn't understand the information it presents.

The Sequence AI of the Week #830: The Quiet Ambush: Inside the Amazing MiMo-V2-Pro aka Hunter Alpha

TheSequence • 112 implied HN points • 25 Mar 26

🕹 Technology Machine Learning

AI is shifting from the "Chat Era" to an "Agent Era" where models are embedded in tool-using, continuous workflows instead of just answering static queries.
A surprising model, MiMo-V2-Pro (aka Hunter Alpha), quietly rose to the top of leaderboards without a public launch or press campaign.
Its stealth deployment as a nameless API on OpenRouter using blind telemetry shows that powerful, disruptive models can appear and win through unconventional, low-profile strategies.

Proving (literally) that ChatGPT isn't conscious

The Intrinsic Perspective • 21487 implied HN points • 15 Jan 26

🔬 Science Machine Learning

You can prove that no scientifically meaningful (falsifiable, non‑trivial) theory of consciousness can consistently say large language models are conscious, because swapping in different implementations that keep the same behavior either falsifies the theory or makes it trivial.
Simple static substitutes like lookup tables or minimal feedforward nets can reproduce an LLM's inputs and outputs but are provably non‑conscious, and because LLMs are very close to those substitutes there isn't room for them to be conscious.
The robust way out is to tie consciousness to continual, online learning; this means research should focus on learning-as-it-happens rather than static input/output or final intelligence alone.

What if AI succeeds but OpenAI fails?

Noahpinion • 17941 implied HN points • 30 Jan 26

🕹 Technology Machine Learning

AI as an industry can succeed even if a flagship company like OpenAI ultimately loses out; early leadership isn’t a guarantee of lasting dominance.
Massive investment is pouring into AI, but high cash burn, commoditization, lack of vertical integration, and intense competition mean investors could be exposed if business fundamentals fail.
Betting everything on a sudden, godlike AGI is basically Pascal’s Wager and not a sound business model; realistic, gradual progress and corporate fundamentals matter far more.

AMA (Ask Machines Anything)

Astral Codex Ten • 12251 implied HN points • 13 Feb 26

🕹 Technology Machine Learning

People increasingly disagree about what AI can do now. Skeptics who avoid paid tools often form opinions from low-quality examples like summary bots or screenshoted mistakes.
An experiment invites readers to submit real questions so Claude 4.6 Opus, a top paid-tier model, can answer them and readers can say if the responses are surprising. The model's first reply will be shown rather than cherry-picked.
Readers are asked to ask medium-difficulty, practical questions instead of gotchas, and the model's settings were adjusted to favor web searches over memory to help reduce hallucinations.

Links For February 2026

Astral Codex Ten • 16656 implied HN points • 05 Feb 26

🕹 Technology Machine Learning

AI is the central theme: there are active debates about alignment and safety, evidence of real failures (and fixes), messy regulatory and political fights, and updated timelines that push major capabilities a few years out.
Medical research and drug trials suffer from perverse incentives and excess cost; experts propose government-funded "high-leverage" trials to test unpatentable or off-patent treatments, which could save public money and improve care.
Tech, culture, and policy are in flux: public belief in ideas like the lab-leak theory is shifting, platform and influence-politics are shaping discourse, and surprising innovations and controversies keep popping up from urban transport to casting choices.

What Happened With Bio Anchors?

Astral Codex Ten • 12044 implied HN points • 12 Feb 26

🕹 Technology Machine Learning

A compute-centered forecasting approach correctly captured that AI progress has largely tracked available compute and scaling laws, which explains much of the recent boom.
The main error was underestimating algorithmic progress and effective compute growth (including longer training runs and test-time compute), so systems became far more powerful each year than the model assumed and pushed timelines much earlier.
Forecasts are still useful but hinge on a few sensitive parameters, so you need proper sensitivity analysis and humility — uncertainty can cut both ways and make outcomes riskier than naive skepticism assumes.

AI Links, 3/8/2026

In My Tribe • 273 implied HN points • 08 Mar 26

🕹 Technology Machine Learning

Agents make execution cheap, so instead of agonizing over one design choice you can have the agent explore multiple options; you must be explicit about success criteria and let the agent check its own work.
Business contracts alone won’t stop government misuse of AI; durable solutions require oversight and legislation so institutions, not companies, set and enforce the rules.
AI language models tend to give more accurate, evidence-based answers than much social media content, so they could reshape public opinion; meanwhile AI keeps surprising us, so claims about its limits can quickly become outdated.

New Data: OpenAI’s Lead Is Contracting as AI Competition Intensifies

Big Technology • 8506 implied HN points • 03 Feb 26

🕹 Technology Machine Learning

ChatGPT’s lead has shrunk — its mobile app share among daily U.S. users fell from 69.1% to 45.3% while Google’s Gemini rose from 14.7% to 25.1% and Grok climbed from 1.6% to 15.2%.
The overall chatbot market exploded, growing about 152% year-over-year, with ChatGPT visits up 50% (3.8B to 5.7B) and Gemini jumping 647% (267.7M to 2B).
Momentum has recently cooled: ChatGPT traffic dipped in November/December and has only partly recovered, while Gemini continues to post strong month-over-month gains.