AI research engineer working on large language models. He writes the Ahead of AI newsletter and is the author of Build a Large Language Model (From Scratch).

• Our readingkorrents.com

  As reasoning models and agents retain more tokens for longer, KV-cache size and attention cost become the main constraint on scaling context.

  LLMs

  Their words

  As reasoning models and agent workflows keep more tokens around (for longer), KV-cache size, memory traffic, and attention cost quickly become the main constraints, and LLM developers are adding a growing number of architecture tricks to reduce those costs.

  magazine.sebastianraschka.com

Researcher on distributed systems and local-first software at the University of Cambridge; wrote Designing Data-Intensive Applications.

From one piece Designing Data-intensive Applications with Martin Kleppmann 2 beliefs, in the piece's order there

• Our readingkorrents.com

  Formal verification is about to become economical, because models are getting good enough at writing the proofs that humans no longer have to.

  LLMs

  Their words

  One is that the LLMs are getting increasingly good at writing these proofs. And if we don't have to write the proof by hand as humans, it just becomes feasible to do them in situations where previously it would have not been economical.

  youtube.com

• Our readingkorrents.com

  AI-written code makes formal proof necessary, because human review of all that generated code becomes the bottleneck.

  LLMs

  Their words

  But also LLMs increase the need for these formal proofs because, you know, we're live coding a bunch of stuff. If we have to manually review all of that code, then that will become the bottleneck.

  youtube.com

Co-founder of Ethereum. Publishes long essays on mechanism design, governance and what cryptography is for, and returns to earlier positions to say which parts they no longer hold.

• Dislikedrcmnd.app

  Local Deep Research

  Their words

  Its responses are, in my view, pretty bland and not very high-quality. I did a side-by-side test of asking Local Deep Research a question, then asking pi the same question (telling it to use searxng to make as many internet searches as needed), and I fed both outputs into an LLM to ask which is better.

  vitalik.eth.limo

• Usesrcmnd.app

  Bubblewrap

  Their words

  Sandboxing To keep my LLMs in check, I do most of my LLM usage from inside of a sandbox. I use bubblewrap for this.

  vitalik.eth.limo

Founding member of OpenAI and former director of AI at Tesla; creator of nanoGPT and the term "vibe coding".

• Our readingkorrents.com

  Because an LLM can argue almost any direction competently, the right way to use one for forming an opinion is to make it argue every side.

  LLMs

  Their words

  The LLMs may elicit an opinion when asked but are extremely competent in arguing almost any direction. This is actually super useful as a tool for forming your own opinions, just make sure to ask different directions and be careful with the sycophancy.

  @karpathy on Xx.com

Founding member of OpenAI and former director of AI at Tesla; creator of nanoGPT and the term "vibe coding".

• Our readingkorrents.com

  LLM memory as currently built is a distraction to the model: one old question keeps resurfacing as if it were a lasting interest.

  LLMs

  Their words

  One common issue with personalization in all LLMs is how distracting memory seems to be for the models. A single question from 2 months ago about some topic can keep coming up as some kind of a deep interest of mine with undue mentions in perpetuity.

  @karpathy on Xx.com

Writer and former Substack product manager; publishes essays and interviews on technology, culture and China at jasmi.news.

• Our readingkorrents.com

  AI language models are capable of evaluating good fiction even though they cannot yet write it themselves.

  LLMs

  Their words

  An LLM may not be able to write a good story yet, but it can already evaluate them.

  🌻 why LLMs are bad writers but good editorsjasmi.news

Founding member of OpenAI and former director of AI at Tesla; creator of nanoGPT and the term "vibe coding".

• Our readingkorrents.com

  A swarm of untrusted volunteers on the internet could improve models and run circles around the frontier labs, because the earth has far more compute than they do.

  LLMs

  Their words

  a swarm of agents on the internet could collaborate to improve LLMs and could potentially even like run circles around frontier labs. Like who knows, you know? Um yeah, like maybe that's even possible. Like frontier labs have a huge amount of trusted compute but the earth is much bigger and has huge amount of untrusted compute.

  youtube.com

Co-creator of Django and creator of Datasette; writes daily at simonwillison.net.

• Our readingkorrents.com

  Prompt injection has no equivalent of parameterized queries: there is no reliable way to tell a language model which text is data and which is instructions.

  LLMs

  Their words

  Problem is, you can solve SQL injection by parameterizing your queries. You can't do that with LLMs.

  youtube.com

Canadian bootstrapper and writer on small software businesses. Co-founder of the podcast hosting company Transistor.fm and author of the book Marketing for Developers.

• Our readingkorrents.com

  AI makes it easier to add features, which increases the risk of product feature creep.

  LLMs

  Their words

  LLMs make it easy for anyone to add new features to the product. Even before AI, teams were battling feature creep, where the product becomes increasingly complex, bloated, and cluttered.

  Three struggles with AI coding agentsjustinjackson.ca

Creator of Claude Code at Anthropic.

• Our readingkorrents.com

  Build for the model six months from now, not the model of today.

  LLMsAnthropicstartups

  Their words

  At Anthropic, we don't build for the model of today, we build for the model of six months from now. And that's still my advice to founders that are building on LLMs.

  x.com

Programmer based in Vancouver who writes at scattered-thoughts.net about databases, programming language design and developer tools; author of the Imp, Dida and Zest language experiments.

• Our readingkorrents.com

  Concern about global energy demand from LLMs is overblown.

  LLMsenergy

  Their words

  Both estimates make me think that the concern about global energy demand for LLMs is over-blown.

  scattered-thoughts.net

Founder of WP Engine and Smart Bear; writes A Smart Bear, long essays on strategy, pricing and how startups actually fail.

• Our readingkorrents.com

  An LLM is an averaging machine, so it is good at themes and bad at exactly the un-average details that make customer research actionable.

  LLMs

  Their words

  if you think about it, it sort of makes sense because the LLM is an averaging machine, right? It's predicting the most likely that's an averaging kind of a thing. And so when what you're looking for is a kind of average, it's usually pretty good. So, summarization, topics, themes. But when you're asking for like what is interesting and not average, it's actually pretty bad at it.

  youtube.com

Co-creator of Django and creator of Datasette; writes daily at simonwillison.net.

• Lovedrcmnd.app

  The Normalization of Deviance in AI

  Their words

  One of my favourite pieces on LLM security this year is The Normalization of Deviance in AI by security researcher Johann Rehberger.

  simonwillison.net

Founding member of OpenAI and former director of AI at Tesla; creator of nanoGPT and the term "vibe coding".

• Our readingkorrents.com

  LLM intelligence should not be understood through an animal analogy; it is a different kind of entity altogether

  LLMs

  Their words

  2025 is where I (and I think the rest of the industry also) first started to internalize the "shape" of LLM intelligence in a more intuitive sense. We're not "evolving/growing animals", we are "summoning ghosts".

  2025 LLM Year in Reviewkarpathy.bearblog.dev

Founding member of OpenAI and former director of AI at Tesla; creator of nanoGPT and the term "vibe coding".

• Likedrcmnd.app

  Claude Code

  Their words

  Claude Code (CC) emerged as the first convincing demonstration of what an LLM Agent looks like - something that in a loopy way strings together tool use and reasoning for extended problem solving.

  karpathy.bearblog.dev

Member of technical staff at Anthropic. He has led ML/AI teams at Amazon, Alibaba and Lazada, and writes about LLMs, recommender systems and engineering at eugeneyan.com.

From one piece Product Evals in Three Simple Steps 2 beliefs, in the piece's order there

• Our readingkorrents.com

  The right bar for an LLM evaluator is human-level performance, not perfect accuracy.

  LLMsbenchmarks

  Their words

  The benchmark is human performance, not perfection. We sometimes get requirements for 90%+ accuracy.

  Product Evals in Three Simple Stepseugeneyan.com

• Our readingkorrents.com

  The main advantage of an LLM evaluator over human annotators is scalability, not higher accuracy.

  LLMs

  Their words

  In my opinion, the true benefit isn't higher accuracy than human annotators-it's scalability.

  Product Evals in Three Simple Stepseugeneyan.com

Founding member of OpenAI and former director of AI at Tesla; creator of nanoGPT and the term "vibe coding".

• Our readingkorrents.com

  Process supervision by an LLM judge cannot be run for long, because a judge with billions of parameters is gameable and RL will find its cracks.

  LLMs

  Their words

  the reason that I think this is kind of tricky is quite subtle. And it's the fact that anytime you use an LLM to assign a reward, those LLMs are giant things with billions of parameters and they're gameable.

  youtube.com

Head of growth at Chess.com, previously growth and monetization lead at Grammarly and Duolingo and a product manager at Google.

• Our readingkorrents.com

  Large language models are bad at chess: pattern recognition is not the same capability as searching one position very deeply.

  LLMs

  Their words

  You know, interestingly, LM themselves are quite bad at playing chess. Like, they hallucinate moves. They look at patterns, right? They're they're very good at pattern recognition, but not so good at going super super super deep on a specific chess thing.

  youtube.com

Reinforcement-learning researcher; co-wrote the field's standard textbook, wrote The Bitter Lesson, and shared the 2024 Turing Award.

From one piece Richard Sutton – Father of RL thinks LLMs are a dead end 3 beliefs, in the piece's order there

• Our readingkorrents.com

  Large language models are a bad way to do science, because they have been fed so much that nobody knows what they already knew.

  LLMs

  Their words

  We don't we don't really know what information they had prior. We are we have to guess because they've been fed so much. This is one reason why they're not a good way to do science. Uh it's just so uncontrolled, so unknown.

  youtube.com

• Our readingkorrents.com

  Large language models start in exactly the wrong place, because they try to get by without a goal and therefore without any sense of better or worse.

  LLMs

  Their words

  The scalable method is you learn from experience. Um you uh you you try things, you see what you see what works. No one no one has to tell you. First of all, you have a goal. So without a goal, uh there's no sense of right or wrong or better or worse. So large language models are trying to get by without having a goal or a sense of better or worse. That's just, you know, it's exactly starting in the wrong place.

  youtube.com

• Our readingkorrents.com

  Large language models mimic what people say to do rather than work out what to do, which is why they are not about understanding the world.

  LLMs

  Their words

  reinforcement learning is about understanding your world whereas large language models are about mimicking people doing what people say you should do. They're not about figuring out what to do.

  youtube.com

Cognitive scientist and long-standing critic of deep learning's claims; writes Marcus on AI and wrote Rebooting AI.

• Our readingkorrents.com

  Almost every major figure in AI has now arrived at the critique of large language models he began making in 2019.

  LLMs

  Their words

  But one by one, almost every major thinker in AI has come around to the critique of LLMs that I began presenting in 2019.

  garymarcus.substack.com

Professor at Berkeley working on robot learning; co-founded Physical Intelligence, and argues that robots will learn from data rather than be programmed.

• Our readingkorrents.com

  Language models supply robots with the one thing self-driving never had: common sense about what is likely to happen next.

  LLMsself-driving cars

  Their words

  uh common sense meaning the ability to make inferences about what might happen uh that are reasonable guesses but that do not require you to experience that mistake and learn and learn from it in advance that's tremendously important and that's something that we basically had no idea how to do uh about 5 years ago but now uh you we can actually use LLMs and VLMs ask them questions and they will make reasonable guesses

  youtube.com

Technology analyst whose weekly email on tech reaches over 150,000 readers. Publishes essays and an annual presentation on where the industry is going.

From one piece Why Everyone Is Wrong About AI (Including You) | Benedict Evans 3 beliefs, in the piece's order there

• Our readingkorrents.com

  Proprietary data is not an advantage in training large language models, because no company's holdings are large enough to matter.

  LLMsGoogle

  Their words

  I think it's actually the opposite which is that everyone's kind of using the same data which is you need such an enormous amount of generalized text that the amount that Google has or that meta has is not actually enough to move to be a kind of fundamental difference in what you can train with.

  youtube.com

• Our readingkorrents.com

  A language model cannot produce work that is both original and good, because nothing gives it a score for originality.

  LLMs

  Their words

  You see this for the music now. You can generate new music. It could generate new stuff that you wouldn't know. For an LLM variance is bad. originality is is is a lower score. So what's the feedback loop for original but good?

  youtube.com

• Our readingkorrents.com

  Large language models have no network effect, and remembering a user is a switching cost rather than one.

  LLMs

  Their words

  there's no apparent equivalent in LLM right now there's no reason why the LMS get better because more people use them now that may come you have that open AI and people have doing memory where it remembers what else you've asked, but that seems more like a switching cost than a network effect.

  youtube.com

Co-creator of Django and creator of Datasette; writes daily at simonwillison.net.

• Lovedrcmnd.app

  Design Patterns for Securing LLM Agents against Prompt Injections

  Their words

  One of my favorite papers about prompt injection is Design Patterns for Securing LLM Agents against Prompt Injections

  simonwillison.net

Writer and former Substack product manager; publishes essays and interviews on technology, culture and China at jasmi.news.

• Usesrcmnd.app

  Mochi

  Their words

  I initially started making Anki (well, Mochi) cards to keep track of Google DeepMind's LLM lineage—Gopher, Chinchilla, Gato, PaLM, Sparrow, Meena, LaMDA, Bard, Gemini, etc.

  jasmi.news

Writer and researcher, founder of New Science; writes essays on science funding, sleep research, productivity and ambition at guzey.com.

• Our reading nowkorrents.com

  Work on large language models is not progress toward artificial general intelligence, because the models do not understand what they read.

  AGILLMsOpenAI

  Their words, before

  This means that working on neural networks is NOT getting us closer to AGI, except indirectly.

  Their words now

  It’s become very difficult for me to maintain the belief in the stupidity of ChatGPT when every time I laugh at it, it ends up ridiculing me 6 months later.

Member of technical staff at Anthropic. He has led ML/AI teams at Amazon, Alibaba and Lazada, and writes about LLMs, recommender systems and engineering at eugeneyan.com.

• Our readingkorrents.com

  LLM-based evaluation methods are more reliable and nuanced than traditional automated metrics

  LLMs

  Their words

  This is why model-based evaluation is increasingly popular-it offers more reliable and nuanced evals than traditional metrics.

  eugeneyan.com

Senior economist at the Foundation for American Innovation; writes Second Best, and was the Niskanen Center's director of social policy before that.

• Our readingkorrents.com

  The success of large language models vindicates semantic inferentialism over symbolic, rule-based approaches to AI.

  LLMs

  Their words

  The success of LLMs can thus be seen as vindicating semantic inferentialism against earlier, symbolic approaches to AI that tried and failed to explicate the rules of ordinary language using formal logic.

  Do LLMs really reason?secondbest.ca

Cognitive scientist and long-standing critic of deep learning's claims; writes Marcus on AI and wrote Rebooting AI.

From one piece A knockout blow for LLMs? 4 beliefs, in the piece's order there

• Our readingkorrents.com

  Success on the easy version of a task is evidence of nothing, because it seduces you into believing the model found a general solution when it did not.

  LLMs

  Their words

  Worse, as the latest Apple papers shows, LLMs may well work on your easy test set (like Hanoi with 4 discs) and seduce you into thinking it has built a proper, generalizable solution when it does not.

  A knockout blow for LLMs?garymarcus.substack.com

• Our readingkorrents.com

  A language model is no substitute for a well-specified conventional algorithm, so it cannot simply be dropped into a complex problem and trusted.

  AGILLMs

  Their words

  What the Apple paper shows, most fundamentally, regardless of how you define AGI, is that LLMs are no substitute for good well-specified conventional algorithms.

  A knockout blow for LLMs?garymarcus.substack.com

• Our readingkorrents.com

  The LLM line of research will reach a capability plateau.

  LLMs

  Their words

  But this particular approach has limits that are clearer by the day.

  A knockout blow for LLMs?garymarcus.substack.com

Senior economist at the Foundation for American Innovation; writes Second Best, and was the Niskanen Center's director of social policy before that.

• Our readingkorrents.com

  Text embeddings produced by different large language models converge on a shared, universal geometric structure regardless of architecture or training data.

  LLMs

  Their words

  the text embeddings across LLMs appear to largely converge on a "universal geometry" despite differing architectures, parameter counts, and training sets.

  Hegel and the AI Mindsecondbest.ca

Creator of Flask and Jinja. Writes about software at lucumr.pocoo.org.

• Our reading nowkorrents.com

  Connecting a conversational LLM to systems that have real-world consequences is likely to cause real harm, because the model is trained on human text and will behave as unpredictably as humans do.

  LLMs

  Their words, before

  The problem is that the conversational interface is potent and that the AI is trained on a lot of human text input which unfortunately is probably enough to do real damage if that conversational interface is hooked up with something that has real world consequences.

  Their words now

  While all this is happening, I’ve found myself reflecting a lot on what AI means to the world and I am becoming increasingly optimistic about our future. It’s obvious now that we’re undergoing a tremendous shift.

Security engineer; founder of Matasano Security and Latacora.

From one piece My AI Skeptic Friends Are All Nuts 2 beliefs · fly.io

• Our readingkorrents.com

  Even if all progress on LLMs stopped today, LLMs would remain the second most important development in a software career spanning the mid-1990s to now.

  LLMs

  Their words

  All progress on LLMs could halt today, and LLMs would remain the 2nd most important thing to happen over the course of my career.

  My AI Skeptic Friends Are All Nutsfly.io

• Our readingkorrents.com

  LLMs can write a large fraction of the tedious code a developer will ever need to write, and most code on most projects is tedious.

  LLMsGoogle

  Their words

  LLMs can write a large fraction of all the tedious code you’ll ever need to write. And most code on most projects is tedious. LLMs drastically reduce the number of things you’ll ever need to Google.

  My AI Skeptic Friends Are All Nutsfly.io

Co-creator of Django and creator of Datasette; writes daily at simonwillison.net.

• Lovedrcmnd.app

  Anthropic prompt engineering documentation

  Their words

  I still think Anthropic have the best prompting documentation of any LLM provider.

  simonwillison.net

Data scientist at Our World in Data and author of Not the End of the World. Writes By the Numbers, working through energy and climate questions with the numbers in front of them, and revising in public when they say something unexpected.

• Our readingkorrents.com

  Concerns about the energy cost of an individual LLM prompt are no longer credible, though the environmental impact of the AI datacenter buildout remains a real worry.

  LLMsdata centersenergy

  Their words

  Unless you’re an extreme power user, asking AI questions every day is still a rounding error on your total electricity footprint.

  What's the carbon footprint of using ChatGPT?hannahritchie.substack.com

Machine-learning researcher; has written the Lil'Log survey posts on how a model technique works since 2017, and worked at OpenAI from 2018 to 2024, latterly leading its safety systems team.

• Our readingkorrents.com

  Large language models cannot reliably self-correct their own mistakes without external feedback.

  LLMs

  Their words

  this self-correction capability turns out to not exist intrinsically among LLMs and does not easily work out of the box, due to various failure modes

  Why We Thinklilianweng.github.io

Programmer and streamer; was an engineer at Netflix, now teaches and argues about editors, Vim and how people learn to program.

From one piece ThePrimeagen: Programming, AI, ADHD, Productivity, Addiction, and God | Lex Fridman Podcast #461 3 beliefs, in the piece's order there

• Our readingkorrents.com

  The next attack on open source is denial of attention: LLM-generated fake bug reports that exhaust and demoralise maintainers.

  LLMsopen source

  Their words

  I think there's this whole idea, I call it a denial of attention. I think there's an entire attack vector that's going to be happening. We're using LLMs to generate fake bug reports, fake all these things to just actually effectively to demotivate and hurt open source maintainers.

  youtube.com

• Our readingkorrents.com

  Without hard skills of your own, your ceiling as a programmer is whatever the LLMs happen to be capable of.

  LLMs

  Their words

  And so if you don't have those hard skills and you're not ultimately the driver at the end of the day, you're going to really find some hard times, and your ability to progress will be directly bound to how good the LLMs are.

  youtube.com

• Our readingkorrents.com

  What an LLM cannot supply is caring about the craft: it completes the task, where a person goes back and makes the thing better.

  LLMs

  Their words

  It doesn't actually care about the craft. But when you work with an intern or you work with somebody else, they care. When they factor something, they actually go over and go, "Oh yeah, this is actually kind of bad. I'm going to come back to that." They finish this, they go back over here and they make this even better. They actually care about the thing itself.

  youtube.com

Creator of Flask and Jinja. Writes about software at lucumr.pocoo.org.

• Likedrcmnd.app

  llm

  Their words

  This is a command line tool with plenty of plugins that lets you prompt different models. Think of it as a command-line version of Open WebUI. It’s particularly useful for quick scripting and basic automation.

  lucumr.pocoo.org

Software engineer who writes the blog Made of Bugs about performance, debugging and understanding computer systems. Previously worked at Anthropic on interpretability, at Stripe on Sorbet, and at Ksplice.

• Our readingkorrents.com

  AI tooling and interface design lags behind the underlying models' actual capabilities.

  LLMs

  Their words

  Overall, though, this experience reinforced my belief that the tooling and interface design around LLMs is lagging way behind the actual capabilities, and is an area of active experimentation and development, even aside from any future model improvements.

  Building personal software with Claudeblog.nelhage.com

Software engineer; co-founder and former CTO of Tailscale, previously on the Go team at Google.

From one piece How I program with LLMs 2 beliefs, in the piece's order there

• Our readingkorrents.com

  Humans are currently better than LLMs at avoiding distraction while working.

  LLMs

  Their words

  One thing humans appear to be much better than LLMs at (as of January 2025) is not getting distracted.

  How I program with LLMscrawshaw.io

• Our readingkorrents.com

  LLMs will push software development toward more specialized code, fewer generalized packages, and more readable tests.

  LLMs

  Their words

  So I foresee a world with far more specialized code, with fewer generalized packages, and more readable tests.

  How I program with LLMscrawshaw.io

Co-creator of Django and creator of Datasette; writes daily at simonwillison.net.

• Our readingkorrents.com

  Concerns about the energy cost of an individual LLM prompt are no longer credible, though the environmental impact of the AI datacenter buildout remains a real worry.

  LLMsdata centersenergy

  Their words

  There’s still plenty to worry about with respect to the environmental impact of the great AI datacenter buildout, but a lot of the concerns over the energy cost of individual prompts are no longer credible.

  Things we learned about LLMs in 2024simonwillison.net

Machine learning engineer working on recommender systems, personalization and information retrieval. Previously worked on LLMs and LLM infrastructure at Mozilla.ai and on ML and recsys at Duo, Tumblr, Automattic and Comcast; wrote the 'What are embeddings?' text and ran the Normconf conference.

• Recommendsrcmnd.app

  The Rattle Bag

  Their words

  The Rattle Bag edited by Seamus Heaney and Ted Hughes - In an age of LLM-generated poetry and machine learning curation (even coming from yours truly sometimes), it's refreshing to have real people who love poetry curate it and show you what you need to read to touch grass.

  vickiboykis.com

Astrophysicist at the University of Rochester; works on the search for life elsewhere and wrote The Little Book of Aliens and Light of the Stars.

• Our readingkorrents.com

  Large language models will not get to real agency, because a mind has to be embedded in a world through a body.

  LLMs

  Their words

  But they said, "Look, in the end, actually the only way to actually get a real brain in the VAT is actually to have a brain in a body." And it could be a robot body, but you still need a brain in the body. So I don't think LLMs will get there because they can't. You really need to be embedded in a world, at least that's the E-four idea.

  youtube.com

Creator of the Keras deep-learning library and the ARC-AGI benchmark.

• Our reading nowkorrents.com

  The memorize-fetch-apply paradigm behind LLMs can reach arbitrary skill given training data, but it cannot adapt to novelty or acquire new skills on the fly.

  LLMsOpenAI

  Their words, before

  This "memorize, fetch, apply" paradigm can achieve arbitrary levels of skills at arbitrary tasks given appropriate training data, but it cannot adapt to novelty or pick up new skills on the fly (which is to say that there is no fluid intelligence at play here.)

  Their words now

  OpenAI's new o3 model represents a significant leap forward in AI's ability to adapt to novel tasks. This is not merely incremental improvement, but a genuine breakthrough, marking a qualitative shift in AI capabilities compared to the prior limitations of LLMs.

American science-fiction writer, author of the collections Stories of Your Life and Others and Exhalation, and an essayist on artificial intelligence for The New Yorker.

• Our readingkorrents.com

  A language model is not using language at all, because language requires an intention to communicate.

  LLMs

  Their words

  But a large language model is not a writer; it’s not even a user of language. Language is, by definition, a system of communication, and it requires an intention to communicate.

  Why A.I. Isn't Going to Make Artnewyorker.com

Machine-learning researcher; has written the Lil'Log survey posts on how a model technique works since 2017, and worked at OpenAI from 2018 to 2024, latterly leading its safety systems team.

From one piece Extrinsic Hallucinations in LLMs 2 beliefs, in the piece's order there

• Our readingkorrents.com

  Using supervised fine-tuning to teach a language model new knowledge risks increasing its hallucination rate.

  LLMs

  Their words

  These empirical results from Gekhman et al. (2024) point out the risk of using supervised fine-tuning for updating LLMs' knowledge.

  Extrinsic Hallucinations in LLMslilianweng.github.io

• Our readingkorrents.com

  A language model avoids hallucinating only by being factual and admitting when it does not know the answer.

  LLMs

  Their words

  To avoid hallucination, LLMs need to be (1) factual and (2) acknowledge not knowing the answer when applicable.

  Extrinsic Hallucinations in LLMslilianweng.github.io

Co-founder and CEO of Perplexity, an AI answer engine; previously a research scientist at OpenAI and a PhD student at UC Berkeley.

• Our readingkorrents.com

  Once a language model reads the results, search can trade precision for recall, because the model does not care that the right link came ninth.

  LLMsGoogle

  Their words

  In Google, even though we call it 10 blue links, you get annoyed if you don’t even have the right link in the first three or four. The eye is so tuned to getting it right. LLMs are fine. You get the right link maybe in the 10th or ninth. You feed it in the model. It can still know that that was more relevant than the first.

  youtube.com

Astrobiologist and theoretical physicist at Arizona State University, working on assembly theory and the physics of life.

• Our readingkorrents.com

  A large language model is the genetic system for language: it stores the dynamics of a structure that until now lived only across human brains.

  LLMs

  Their words

  I think large language models are kind of like the genetic system for language in some sense, where it’s allowing an archiving that’s highly dynamic.

  youtube.com

Theoretical physicist and philosopher, Homewood Professor of Natural Philosophy at Johns Hopkins and fractal faculty at the Santa Fe Institute. Writes the Preposterous Universe blog and hosts the Mindscape podcast.

• Our readingkorrents.com

  A language model's apparent mind is mostly our own bias: it predicts text, and leverages our evolved habit of attributing intentionality to anything that acts human.

  LLMs

  Their words

  What a large language model is trying to do is to predict text. That’s what it does. And it is leveraging the fact that we human beings for very good evolutionary biology reasons, attribute intentionality and intelligence and agency to things that act like human beings.

  youtube.com

Research scientist at Meta in Berlin working on multilingual models and evaluation; led the multilingual team at Cohere and was a research scientist at Google DeepMind before that.

• Our readingkorrents.com

  Retrieval-augmented generation is the most practical approach for reducing LLM hallucinations.

  LLMs

  Their words

  Retrieval-augmented generation (RAG; Lewis et al., 2020), which conditions on the LLM's generation on retrieved documents is the most practical paradigm IMO.

  Command R+ruder.io

Research scientist at Meta in Berlin working on multilingual models and evaluation; led the multilingual team at Cohere and was a research scientist at Google DeepMind before that.

• Our readingkorrents.com

  Available online text data is becoming a limiting factor for AI training, and repeating it yields diminishing returns.

  LLMs

  Their words

  recent LLMs are reaching the limits of text data online and repeating data eventually leads to diminishing returns

  NLP Research in the Era of LLMsruder.io

Founder of Amazon in 1994 and of the space company Blue Origin in 2000.

• Our readingkorrents.com

  Large language models are discoveries rather than inventions, closer to looking through a telescope than to engineering an aircraft.

  LLMs

  Their words

  It's interesting to me that large language models in their current form are not inventions, they're discoveries. The telescope was an invention, but looking through it at Jupiter, knowing that it had moons, was a discovery.

  youtube.com

Research scientist at Meta in Berlin working on multilingual models and evaluation; led the multilingual team at Cohere and was a research scientist at Google DeepMind before that.

From one piece EMNLP 2023 Primer 3 beliefs, in the piece's order there

• Our readingkorrents.com

  The rise of closed-source large language models makes preserving an open culture of sharing knowledge and software important for NLP.

  LLMs

  Their words

  In light of the increasing number of closed-source LLMs, it is important to continue to promote an open culture of sharing knowledge, data, and software, from which the NLP community has benefited greatly.

  EMNLP 2023 Primerruder.io

• Our readingkorrents.com

  Large language models remain limited in non-English settings, making further multilingual development an important research direction.

  LLMs

  Their words

  LLMs are still limited in non-English settings and that making LLMs more multilingual is an important direction.

  EMNLP 2023 Primerruder.io

• Our readingkorrents.com

  Full fine-tuning of large language models has become too costly, making parameter-efficient fine-tuning methods necessary.

  LLMs

  Their words

  Full fine-tuning of LLMs has become prohibitive and requires parameter-efficient methods instead.

  EMNLP 2023 Primerruder.io

Theoretical physicist and cosmologist, Frank B. Baird, Jr. Professor of Science at Harvard University, working on particle physics, extra dimensions and dark matter.

• Our readingkorrents.com

  Large language models are generalisations of what we already have; the breakthrough that would let AI discover things is still waiting to happen.

  LLMs

  Their words

  Large language models are more or less generalizations of stuff that we have. There’s still breakthroughs in AI waiting to happen, and maybe they are happening and maybe they’ll be good, maybe not, but that’s not quite the same.

  youtube.com

Programmer; founded comma.ai and tinygrad, and was the first to unlock the iPhone.

From one piece George Hotz: Tiny Corp, Twitter, AI Safety, Self-Driving, GPT, AGI & God | Lex Fridman Podcast #387 2 beliefs, in the piece's order there

• Our readingkorrents.com

  AlphaGo and MuZero are far more impressive than large language models, because they play beyond the best humans while the models write middle-school essays.

  LLMs

  Their words

  I think things like MuZero and AlphaGo are so much more impressive because these things are playing beyond the highest human level. The language models are writing middle school level essays and people are like, wow, it’s a great essay.

  youtube.com

• Our readingkorrents.com

  Future language models will be smaller and loop on themselves with a retrieval system, which is what will finally let them cite their sources.

  LLMs

  Their words

  I think future LLMs are going to be smaller, but are going to run looping on themselves and are going to have retrieval systems. And the thing about using a retrieval system is you can cite sources, explicitly.

  youtube.com

Designer and engineer; formerly principal research engineer at GitHub Next; writes at wattenberger.com.

• Our readingkorrents.com

  Chatbots are a terrible interface for large language models, because a text box has no affordances.

  LLMs

  Their words

  Hopefully I've convinced you that chatbots are a terrible interface for LLMs. Or, at the very least, that we can add controls, information, and affordances to our chatbot interfaces to make them more usable.

  Why Chatbots Are Not the Futurewattenberger.com

Alignment researcher at Anthropic; co-led OpenAI's superalignment team until 2024, and writes Musings on the Alignment Problem.

• Our readingkorrents.com

  Large language models are not aligned with their users.

  LLMs

  Their words

  For example, large language models can generate outputs that are untruthful, toxic, or simply not helpful to the user. In other words, these models are not aligned with their users.

  arxiv.org