HN treats it as the go-to visual introduction to transformers, and many say it finally made attention and embeddings click.
Topics › AI, Machine Learning & Data
The best llms & generative ai videos, according to Hacker News
107 videos about llms & generative ai that Hacker News commenters keep pointing each other to, ranked by how many separate months they were recommended. Part of AI, Machine Learning & Data.
HN's most frequent recommendation for a clear, approachable foundation in how LLMs work without heavy math.
It is the go-to recommendation for learning how transformers and LLMs actually work, praised for its clarity and Karpathy's teaching.
Repeatedly named the best practical tutorial for using LLMs, with a well-liked fine-tuning section.
HN recommends it as the dense, approachable starting point for understanding LLMs and often cites it for specific claims about scaling and training.
HN recommends it as a thought-provoking talk from a researcher with early GPT-4 access, especially for the unicorn degrading under safety tuning.
Recommended as the best deep dive on tokenizers, worth the two hours for anyone working with LLMs.
HN treats it as the go-to visual explanation of attention, recommended to people who find text-based transformer explainers hard to follow.
Frequently recommended as the overview to get up to speed on AI risk, especially by people who valued The Social Dilemma.
Recommended as a clear explanation of why RNNs fall short and how transformers solve it.
HN recommends it as an accessible, clear walkthrough of the transformer paper for people with some AI background.
Cited as an early, striking demo of neural networks acting as game engines and world models.
HN commenters point to it as an impressive example of what AI animation could do and of VFX artists adopting AI in their workflows.
HN readers recommend it as an intuitive explanation of neural networks that shifts how people think about AI and language.
HN commenters recommend it as a first-hand argument for building your own lean agent harness instead of relying on bloated, vendor-controlled ones.
Recommended as the clearest overview of how LLM assistants are built and how to use them well.
HN commenters recommend it for building intuition about state space models and the tradeoffs between transformer and SSM layers.
HN points to it as the definitive account of the GPT-4chan experiment, used in debates on LLM bias, truthfulness and bots.
HN recommends it to newcomers as a clear explanation of attention, including the query/key/value intuition.
HN recommends it as a long but worthwhile way to demystify how language models work technically.
HN cites it for the argument that we cannot tell whether AI is early or late on its sigmoid curve until hindsight.
HN recommends it as a quick, non-dumbed-down way to get up to speed on how LLMs are built.
HN points to it as the main place to hear Lattner's motivations and design ideas for Mojo, often in Mojo threads.
Pointed to as an accessible early explanation of transformers and attention for newcomers.
HN cites it as a concrete demonstration that LLMs have suddenly become strong at finding and exploiting vulnerabilities.
HN cites it for the argument that LLMs capture language form but not meaning, and for how the brain's language network works.
HN recommends it as a quick way to see how the Krita AI plugin works and how it matches Photoshop's generative features.
HN points to it as an early, impressive demonstration that LLMs could write working code.
HN uses it as a reference for AI slop, LLMs citing LLM output, and the growing difficulty of fact checking.
HN commenters recommend watching it to hear his side directly, noting he comes across as coherent, while others dispute his sincerity.
HN commenters share it as an eloquent argument that AI output is not art, and as a companion to worries about losing deep thinking.
HN people link it when discussing AI-generated spam clogging search results and the broader harm of generative AI to the web.
HN treats it as the origin of the full-wine-glass meme and an illustration of how models lack concepts missing from their training data.
Recommended as a well-made, digestible introduction to scaling laws and why models cluster along a similar efficiency frontier.
Brought up in every deepfake thread as a funny and technically convincing early example of synthetic media used for satire.
Pointed to as a non-promotional, realistic long session of an experienced engineer working with AI on non-trivial code.
Shared as a funny example of jailbreaking AI voice bots, and as the source of the bolognese prompt-injection phrase.
HN cites it as the early example of voice cloning and audio deepfakes, with some recommending Peele's reaction and others calling the demo fake or never shipped.
HN threads on the Rabbit R1 point to it as an exposé on the company's history and operation.
Used as a go-to example of neural networks enhancing game rendering, referenced in later AI-graphics discussions.
HN points to it as the accessible explanation of why glitch tokens exist whenever tokenizers or odd GPT behavior come up.
Commenters cite it during Stable Diffusion release debates as a source for Mostaque's vision and Stability's principles.
Pointed to as one of the most impressive early deepfakes, where a good impersonation makes the effect convincing and funny.
Cited in AI-and-education threads as the optimistic case and demo for AI tutoring, and as a source on Khan Academy's LLM approach.
Recommended in LLM threads as an authoritative explanation from a ChatGPT architect of why models hallucinate and why imitation is not enough.
HN commenters point to it as the best first-hand account of Cursor's technical internals, straight from its founders.
Cited as a firsthand account of the raw, unfiltered GPT-4 before safety tuning, especially the part after 45 minutes.
Recommended as a more nuanced look at Hinton's views than headlines, especially his argument that LLMs are more than auto-complete.
Pointed to as LeCun's argument for the limits of LLMs and the need for world models.
HN commenters send it to people confused about ChatGPT's quirks, explaining how human-feedback training makes it a pleasing yes-man that bullshits.
HN points to it in AI music and data-poisoning threads as the source for this technique.
Referenced as the definitive review explaining why the AI Pin failed.
Commenters cite it as a rare, pointed grilling of Apple executives over why Siri fell behind.
Used in AI threads as a cautionary example of harm from LLM self-diagnosis, though some commenters cite it as part of broader criticism of AI.
Cited for LeCun's views on LLM limits and alternative approaches to intelligence.
HN points to it when discussing AI-generated bug reports and vibe-coded pull requests burdening open source maintainers.
HN commenters cite it as a primary source for what Zuckerberg says in his own words, such as on AI writing most new code, WhatsApp, and Apple interoperability.
HN points to it as a demonstration of local, controllable AI painting and a way to learn the workflow.
Mostly cited as the canonical statement of AI doomer views, with some recommending it and many criticizing it as irresponsible or wrong.
HN links it as a balanced, full-length interview with a leading alignment researcher on AI risk.
HN mostly cites it as a source for Karpathy's framing of programming in English and his generate-then-verify workflow, with a few critics finding it hype.
HN cites it as an early, quirky example of machine-generated screenplays and what LSTMs could do.
HN recommends it in threads on AI music flooding streaming platforms, for its practical detection ideas such as compression artifacts.
HN points to it as a takedown of AI-assisted amateur physics, though a few commenters push back on it.
Linked as the approachable explanation whenever HN discusses detecting or watermarking AI-generated text.
HN shares it as a fun demo of an LLM driving a live Emacs process and orchestrating its long-standing ecosystem.
HN cites it as an example of temporally stable AI video stylization built from ordinary production gear.
HN uses it as a clear explainer for diffusion mechanics, such as negative prompts and faster sampling, and for joint text-image encoders.
HN refers back to it as the primary source for what GPT-4 could do at launch, especially the sketch-to-website and multimodal demos.
HN cites it as an example of what a very small crew can make with generative AI tools, while noting how much manual work it still takes.
HN cites it for insight into Google researchers' views on AI progress and publishing, though one commenter finds the interview weak.
HN cites it as an example of deepfakes producing a more convincing CGI human than the studio's own effort.
HN commenters cite it as the serious skeptic's case against LLMs reaching general intelligence, usually in debates about whether scaling is enough.
HN mostly cites it as an example of a misleading or faked product demo, with a few pointing to its teachable-mode segment.
HN mostly cites it critically, as evidence of air pollution from AI data center turbines in Memphis.
Cited as an expert's first-hand evidence that GPT translation is near human quality for distant language pairs.
HN links it as an informed view from a working translator on using GPT for translation.
Cited in discussions of learning as compression and of why neural networks generalize.
Cited as a concrete early example that AI-faked video was already convincing, in discussions of misinformation and trust in media.
HN commenters cite it as a primary source for Brockman's views, such as moving fast to reduce risk, and for capability anecdotes like GPT-4's arithmetic.
Cited as a source for the fact that OpenAI uses Ray for training, and for Brockman's views on developer speed.
Mostly cited as an example of an overhyped or possibly staged AI demo, alongside later skeptical comparisons to other company demos.
Cited as evidence that AI can solve interview-style algorithm puzzles, in threads questioning the value of Leetcode.
Cited for Amodei's claim that each model is profitable on its own, which commenters debate and dispute.
HN mostly cites it critically, as a source for the claim that much of YC code is AI-written and as an example of insiders promoting vibe coding.
Only mentioned as a joke in threads about AI-generated content.
Cited as an early example of AI-composed music, usually with the caveat that humans did much of the production.
HN mostly cites it for specific Altman quotes, such as models as reasoning engines rather than databases, while many commenters found the interview shallow or insincere.
During the Altman ouster threads, HN mined it for his recent unscripted statements on AGI and safety; some also mock its tone.
HN links to specific segments of the keynote to back up claims about features like Passwords, Math Notes and the ChatGPT handoff.
Mostly cited as an example of the state of LLM-controlled robots, with some skepticism that it is a polished demo video.
HN cites it as a process example of how such AI-generated videos are made, including on a phone.
Mostly referenced for its specific segments or criticized for misleading performance comparisons, not recommended broadly.
HN commenters cite it mainly for Sutskever's statements that training data is far from exhausted.
HN commenters cite it as a concise explanation of GPT-3's scaling jump and surprising abilities at its release.
Used as an early, widely seen example of deepfakes in discussions of fake video and its risks.
Pointed to as an early example of LLM-powered NPC dialog in games.
HN commenters cite it as a primary source for Altman's statements, such as releasing AI more slowly, detection of LLM output being futile, and dismissing the GPT-4 parameter-count rumor.
Mostly cited critically, as an example of AI being used for personal moments in ways commenters found dystopian.
HN cites it as the original of the Balenciaga AI meme family and as an early viral AI mashup moment.
Cited as an early, amusing example of LLM-driven command-line toys, with commenters sharing similar experiments.
HN uses it as a primary source for what Altman said on topics like programming changing, GPT-5's expected leap and startup structure, and often posts timestamps or paraphrases.
Cited as an example of how realistic GAN-generated faces had become, and of what a well-structured latent space can do.
HN uses it as a reference point in deepfake and fake-news discussions, showing that convincing real-time video manipulation was already possible in 2016.
Cited as the debunking of Devin's marketing claims in skeptical discussions of AI software engineers.
Cited in AI-coding debates as one of the largest data-driven studies, notably for finding no correlation between self-assessed and measured productivity.
Mostly cited as a source for statements by Anthropic leaders, such as views on AI progress and on sycophancy.
Which of these have you already seen?
Connect YouTube and HackerTube AI hides what you've watched and ranks the rest for you.