HN isn't swamped yet, just obsessed with AI

A few days ago I posted Product Hunt is swamped now, which put one number on how much AI has changed the internet: a single comparable day of launches on Product Hunt went up roughly 17x between 2023 and 2026.

Hacker News is the other feed I open every day, so the obvious next question was whether it tells the same story. I ran the same kind of count. It does not, it gets about the same number of posts it got ten years ago.

The input stream didn’t get bigger

/newest is the Hacker News input stream: every story submission is waiting there, for votes to decide what the lazy folks actually sees, by people who are spending a lot of time curating the feed (thank you, to those who read the newest feed!).

I counted every post that went through it over a fixed two-week window, 8 to 21 May, for each year from 2016 to 2026, using a similar approach as the one for ProductHunt.

New Hacker News /newest submissions per day, averaged over the 8 to 21 May window each year. Orange marks the May 2020 outlier.

It sits a little over 800 a day, and it has for a decade. The only real outlier is May 2020, when many of us were stuck at home with nothing to do. 2026 is the busiest year since, at 940 a day, which is about 16% above 2019, not 16x. Whatever happened to Product Hunt did not happen here.

A quarter of it is now about AI

I also attempted to look at what those posts are about.

This is easy to do badly. Search only for ChatGPT-era words like LLM, GPT and agentic, and you will conclude that AI barely existed on Hacker News before 2023. That is wrong: Hacker News has been talking about machine-learning for a long time. So I built a classifier with more than 300 terms covering the whole field, from random forest, gradient boosting, convolutional, LSTM, word2vec and AlphaGo to the modern ChatGPT, Claude, RAG and agentic vocabulary. A post counts as AI if any of those terms appears in its title. I hesitated to build a LLM-based classifier, maybe I’ll do it if other stats-obsessed nerds ask for it in the HN-comment section.

Share of Hacker News /newest submissions (8 to 21 May each year) with an AI or machine-learning term in the title. Orange marks the post-ChatGPT years.

From 2016 to 2019 there was a deep-learning wave: AI titles climb from 3.3% to a pre-ChatGPT peak of 5.8% in 2018, the AlphaGo years. Then it fades and by 2022 it is back down to 3.5%, lower than it was in 2017. My best guess is that crypto and web3 were occupying Hacker News in 2021 and 2022, and AI’s slice of a fixed-size feed shrank to make room.

Then ChatGPT shipped in November 2022, which led the 2023 window to reach 15.2%, increasing 4x, and it has risen every year since, up to about a quarter of submissions in 2026. Nearly one in four new Hacker News submissions now carries an AI or machine-learning term in its title.

Why Hacker News holds and Product Hunt didn’t

Both ProductHunt and HN are frequently used to launch (and promote) products, so why are we not getting 700 new messages per day on HN related to product launches alone?

Hacker News is not a launch platform, and it has three defences a launch platform lacks: it rate-limits how quickly an account can submit, it is moderated, and the community flags hard enough that flagged-dead posts never reach /newest at all. I would be surprised if the same kind of defenses weren’t to appear on PH, I don’t see how they could keep going the way they currently are.

Method

The data comes from Hacker News’s Algolia search API, which indexes every story. tags=story covers normal link posts, Ask HN and Show HN; it excludes posts that were flagged dead.

I used the same 14-day window, 8 to 21 May, for every year, so the weekday mix and the season stay fixed and only the year varies. For each window I downloaded every story title and classified it locally with word-boundary regular expressions. Short all-caps terms like AI, ML and CNN are matched case-sensitively, so ML does not match 500ml and BERT does not match somebody called Bert.

Caveats

It only reads titles. A post counts as AI only if a term is in its title. Many AI posts have plain descriptive titles and are missed, and they are missed more often in the early years, because in 2017 a machine-learning paper often had a modest title while in 2026 “AI” is a word people put in the title on purpose. So the early baseline is, if anything, undercounted.
Dead posts are invisible. If AI spam is flooding /newest and being flagged to death, neither this count nor Hacker News’s own /newest page ever sees it. The flat volume line is partly Hacker News’s moderation working as intended.
The term list is broad, not hand-audited. A few terms have innocent older meanings, like the pre-AI Gemini protocol or “grok” as ancient hacker slang. But the same list runs against all eleven years, so that noise is a roughly constant background. It cannot invent a four-fold jump between two consecutive years.

Reproduce it

Two scripts, both pasted into the browser console on https://hn.algolia.com.

The first walks the API and downloads every story title for the eleven windows into window.HN, and saves a hn-titles.json backup so you never fetch twice.

Show the downloader

(async () => {
  const YEARS = [2016,2017,2018,2019,2020,2021,2022,2023,2024,2025,2026];
  const WINDOW_DAYS = 14, DAY = 86400;
  const END_MONTH = 4, END_DAY = 22; // May 22; window is the 14 days before it
  const search = async (a, b) => {
    const r = await fetch(`https://hn.algolia.com/api/v1/search_by_date`
      + `?tags=story&hitsPerPage=1000&numericFilters=created_at_i>=${a},created_at_i<${b}`);
    if (!r.ok) throw new Error(`HTTP ${r.status}`);
    return r.json();
  };
  const collect = async (a, b, out) => {
    const j = await search(a, b);
    if (j.nbHits > 1000 && b - a > 1) {
      const m = Math.floor((a + b) / 2);
      await collect(a, m, out); await collect(m, b, out);
    } else {
      for (const h of j.hits)
        out.push({ id: h.objectID, t: h.title || '', u: h.url || '', ts: h.created_at_i });
    }
  };
  const all = [];
  for (const year of YEARS) {
    const end = Math.floor(new Date(year, END_MONTH, END_DAY) / 1000);
    const before = all.length;
    for (let d = 0; d < WINDOW_DAYS; d++)
      await collect(end - (d + 1) * DAY, end - d * DAY, all);
    console.log(`${year}: ${all.length - before} stories`);
  }
  const seen = new Set();
  window.HN = all.filter(s => !seen.has(s.id) && seen.add(s.id));
  console.log(`TOTAL: ${window.HN.length} unique stories saved to window.HN`);
  const dl = document.createElement('a');
  dl.href = URL.createObjectURL(new Blob([JSON.stringify(window.HN)], { type: 'application/json' }));
  dl.download = 'hn-titles.json'; dl.click();
})();

The second classifies them, prints the table behind the chart above, and leaves three helpers on window: hnAudit() for per-term match counts, hnSamples("term") to see exactly what a term catches, and hnUnmatched() to eyeball the posts it missed. The full term list is in the script.

Show the classifier

(() => {
  if (!window.HN) { console.error('Run the downloader first - window.HN is empty.'); return; }
  const TERMS = [
    // Umbrella / general
    'artificial intelligence','machine learning','deep learning','neural network','neural net',
    'artificial neural','data science','data mining','predictive analytics','predictive modeling',
    'pattern recognition','statistical learning','computational intelligence','cognitive computing',
    'expert system','knowledge representation','knowledge graph','symbolic AI','automated reasoning',
    'machine intelligence','AGI','artificial general intelligence','superintelligence','AI','ML','A.I.',
    // Classical ML
    'random forest','decision tree','gradient boosting','gradient boosted','gradient descent',
    'stochastic gradient','XGBoost','LightGBM','CatBoost','support vector machine','support vector',
    'SVM','k-means','nearest neighbor','kNN','naive bayes','logistic regression','ensemble learning',
    'ensemble model','AdaBoost','Markov chain','hidden Markov','Markov decision','Bayesian network',
    'Bayesian inference','Bayesian optimization','Gaussian process','Gaussian mixture',
    'expectation maximization','principal component','dimensionality reduction','t-SNE','UMAP',
    'collaborative filtering','recommender system','recommendation system','recommendation engine',
    'anomaly detection','outlier detection','feature engineering','feature extraction','hyperparameter',
    'cross-validation','overfitting','regularization',
    // Learning paradigms
    'supervised learning','unsupervised learning','semi-supervised','self-supervised learning',
    'reinforcement learning','transfer learning','multi-task learning','active learning',
    'online learning','metric learning','representation learning','curriculum learning',
    // Neural net architectures
    'convolutional neural','convolutional network','CNN','recurrent neural','recurrent network','RNN',
    'LSTM','long short-term memory','GRU','gated recurrent','transformer model','transformer architecture',
    'transformer network','attention mechanism','self-attention','multi-head attention',
    'generative adversarial','GAN','autoencoder','variational autoencoder','ResNet','residual network',
    'U-Net','encoder-decoder','sequence-to-sequence','seq2seq','multilayer perceptron','perceptron',
    'backpropagation','batch normalization','graph neural network','spiking neural','capsule network',
    'diffusion model','denoising diffusion','Boltzmann machine','deep belief network',
    'mixture of experts','neural architecture search','state space model',
    // NLP / language
    'natural language processing','natural language understanding','natural language generation',
    'computational linguistics','language model','large language model','LLM','word embedding','word2vec',
    'GloVe','fastText','sentence embedding','named entity recognition','part-of-speech','sentiment analysis',
    'text classification','topic modeling','latent dirichlet','machine translation',
    'neural machine translation','speech recognition','speech-to-text','automatic speech recognition',
    'text-to-speech','speech synthesis','voice cloning','question answering','text summarization',
    'tokenizer','byte-pair encoding','BERT','RoBERTa','ELMo','XLNet','GPT-2','GPT-3','GPT-3.5','GPT-4',
    'GPT-4o','GPT-5','ChatGPT','generative pre-trained','instruction tuning','RLHF',
    'reinforcement learning from human feedback','prompt engineering','prompt injection',
    'in-context learning','chain-of-thought','few-shot','zero-shot','retrieval-augmented generation',
    'retrieval-augmented','hallucination','context window','foundation model','frontier model','multimodal',
    // Computer vision
    'computer vision','image recognition','image classification','object detection','object recognition',
    'semantic segmentation','instance segmentation','image segmentation','optical character recognition',
    'OCR','facial recognition','face recognition','face detection','pose estimation','image generation',
    'text-to-image','super-resolution','style transfer','ImageNet','YOLO','depth estimation',
    'visual question answering','image captioning','neural radiance','Gaussian splatting','3D reconstruction',
    // Reinforcement learning / agents
    'deep reinforcement','Q-learning','deep Q-network','policy gradient','actor-critic',
    'proximal policy optimization','Monte Carlo tree','multi-armed bandit','contextual bandit',
    'reward model','imitation learning','inverse reinforcement','AlphaGo','AlphaZero','AlphaFold',
    'MuZero','AlphaStar','AI agent','autonomous agent','agentic','multi-agent','agent framework',
    'function calling','tool use','ReAct',
    // Models / labs / products
    'OpenAI','Anthropic','DeepMind','Mistral AI','Cohere','Hugging Face','Stability AI','Midjourney',
    'Stable Diffusion','DALL-E','Imagen','Sora','ElevenLabs','Suno','Claude','Gemini','Google Bard',
    'Llama','LLaMA','Mixtral','Gemma','Qwen','DeepSeek','Command R','Grok','GitHub Copilot','Copilot',
    'Cursor','Windsurf','Devin','Perplexity','character.ai','Inflection AI','xAI','Together AI',
    'Fireworks AI','Groq','Cerebras','Whisper','CLIP','Segment Anything','GPT-J','GPT-NeoX','BLOOM',
    'Falcon LLM','Vicuna','Alpaca','WizardLM','Codex','Runway ML','Phi-3','Llama 2','Llama 3',
    // Frameworks / tools / infra
    'TensorFlow','PyTorch','Keras','scikit-learn','Theano','Caffe','MXNet','ONNX','LangChain','LlamaIndex',
    'vLLM','Ollama','llama.cpp','TensorRT','DeepSpeed','Megatron','Weights & Biases','MLflow','Kubeflow',
    'fine-tuning','fine-tune','LoRA','QLoRA','quantization','model distillation','knowledge distillation',
    'model pruning','model inference','model training','pretraining','model serving','vector database',
    'vector search','vector embedding','semantic search','embeddings','TPU','tensor processing unit',
    'MLOps','AutoML','MMLU','Kaggle','MNIST','training data','labeled data','synthetic data',
    'GPU cluster','neural processing','inference engine','JAX',
    // Safety / society
    'AI safety','AI alignment','alignment problem','AI ethics','AI governance','AI regulation','AI risk',
    'existential risk','algorithmic bias','AI bias','fairness in machine learning','explainable AI',
    'interpretability','mechanistic interpretability','responsible AI','AI doom','AI act','model collapse',
    'deepfake','synthetic media','AI policy','machine ethics','AI winter','data labeling','AI hype',
  ];
  const buildRegex = (term) => {
    const isAcro = /^[A-Z0-9][A-Z0-9.+]*$/.test(term) && /[A-Z]/.test(term) && term.length <= 6;
    const body = term.split(/[\s-]+/)
      .map(t => t.replace(/[.*+?^${}()|[\]\\]/g, '\\$&'))
      .join('[\\s-]*');
    return new RegExp(`(?<!\\w)${body}s?(?!\\w)`, isAcro ? '' : 'i');
  };
  const RX = TERMS.map(t => ({ term: t, re: buildRegex(t) }));
  const byYear = {};
  for (const s of window.HN) {
    const y = new Date(s.ts * 1000).getUTCFullYear();
    (byYear[y] ??= { total: 0, ai: 0 }).total++;
    if (RX.some(({ re }) => re.test(s.t))) byYear[y].ai++;
  }
  const result = Object.keys(byYear).sort().map(y => ({
    year: +y, total: byYear[y].total, aiPosts: byYear[y].ai,
    perDay: Math.round(byYear[y].total / 14),
    sharePct: +(100 * byYear[y].ai / byYear[y].total).toFixed(1),
  }));
  window.HNRESULT = result;
  console.table(result);
  window.hnAudit = () => {
    const rows = TERMS.map(term => {
      const re = buildRegex(term);
      let pre = 0, post = 0;
      for (const s of window.HN)
        if (re.test(s.t)) (new Date(s.ts * 1000).getUTCFullYear() <= 2022 ? pre++ : post++);
      return { term, total: pre + post, pre2022: pre, from2023: post };
    }).filter(r => r.total > 0).sort((a, b) => b.total - a.total);
    console.table(rows);
  };
  window.hnSamples = (term, n = 30) => console.log(
    window.HN.filter(s => buildRegex(term).test(s.t)).slice(0, n).map(s => s.t));
  window.hnUnmatched = (n = 40) => {
    const miss = window.HN.filter(s => !RX.some(({ re }) => re.test(s.t)));
    console.log(`${miss.length} unmatched; random sample:`);
    console.log(Array.from({ length: n }, () => miss[Math.random() * miss.length | 0].t));
  };
  console.log('Helpers: hnAudit() - hnSamples("Gemini") - hnUnmatched(40)');
})();

Sources

Hacker News Algolia search API: https://hn.algolia.com/api
The Hacker News input stream: https://news.ycombinator.com/newest
The companion post: Product Hunt is swamped now