Related posts

Chief Scientist at OpenAI. Worked on the research that turned scaled reinforcement learning into reasoning models, and writes about where he thinks that path leads.

• Our readingkorrents.com

  The fundamental challenge of alignment is generalization: holding values in situations the training never covered.

  AI alignment

  Their words

  The fundamental challenge of AI alignment is generalization.

  An Alien Mindopenai.com

Professor at Stanford working on robot learning and meta-learning; co-founded Physical Intelligence.

• Our readingkorrents.com

  It is the diversity of a robot data set, not its size, that produces generalization: dropping the most diverse slice hurts far more than dropping a random fifth.

  Their words

  we find that the performance on held out tasks decreases dramatically. Whereas if we um just take out a random 20% of the data that's less diverse than the most diverse subset, the performance um only decreases a little bit. And so this suggests that actually having really diverse data plays an important role in enabling it to generalize to new tasks.

  youtube.com

Swedish programmer with degrees in engineering and psychology, author of Your Code as a Crime Scene and founder of CodeScene.

• Likedrcmnd.app

  Multi-Paradigm Design for C++

  Their words

  Multi-Paradigm Design for C++ by James O. Coplien. Twenty years after reading this, the core techniques are still with me. These are commonality analysis, where the purpose is to identify families of systems, and variability analysis, which focuses on capturing the domain parameters that vary. In essence: the foundation of great software design.

  adamtornhill.substack.com

Professor of international affairs at Johns Hopkins SAIS and co-author, with Abraham Newman, of Underground Empire: How America Weaponized the World Economy.

• Our readingkorrents.com

  Ireland's colonial wars helped set the template Britain later applied elsewhere in its empire

  Their words

  The brutal war and its aftermath set some of the parameters for what Britain did later in other parts of the world.

  John Crowley has diedprogrammablemutter.com

Chief Scientist at Google DeepMind and Google Research, at Google since 1999, where he co-created MapReduce, Bigtable, TensorFlow and the TPU.

From one piece Jeff Dean: The 1% Rule for Building in AI 2 beliefs, in the piece's order there

• Our readingkorrents.com

  What sits in a model's context is far clearer to it than its training data, which is trillions of tokens stirred into a soup of parameters.

  Their words

  And the nice thing about that is that information is really clear to the model, unlike the training data the model was trained on where it's all kind of like trillions of tokens stirred together into a soup of of hundreds of billions or trillions of parameters, but it's all less clear than the actual context uh that the model sees directly for this particular problem or uses use case.

  Jeff Dean: The 1% Rule for Building in AIyoutube.com

• Our readingkorrents.com

  You improve a model you cannot retrain by writing better guidelines and skills for it, not by adjusting its parameters.

  Their words

  And often you can actually make the model work better and succeed at that kind of problem by not just adjusting the model parameters which is hard to do from the outside but from you know creating better guidelines for the model you know writing skills for the model to know how to use different tools that would be incredibly useful for solving this particular class of problem.

  Jeff Dean: The 1% Rule for Building in AIyoutube.com

Climate scientist and energy systems analyst; writes The Climate Brink with Andrew Dessler and leads climate research at Stripe.

• Our readingkorrents.com

  The US warming trend since 1970 is consistent across raw, adjusted, satellite, and surface temperature datasets

  Their words

  The overall US warming trend (around 0.5F per decade since 1970 in max, min, and mean temperatures) is robust in every dataset, raw or adjusted, satellite or surface.

  Hot days, cold thermometerstheclimatebrink.com

Co-founder and chief executive of Meta Platforms, the company he started as Facebook in 2004.

• Our readingkorrents.com

  There is no factory you can pay to produce biological data; each new dataset requires inventing a new scientific method first.

  Their words

  it's not just like there's some factory somewhere that you can pay to produce the the data like you actually need to invent new novel scientific approaches

  youtube.com

Co-founder and CEO of Anthropic; previously VP of research at OpenAI.

From one piece Dario Amodei — “We are near the end of the exponential” 2 beliefs, in the piece's order there

• Our readingkorrents.com

  RL environments exist to make a model generalise, not to teach it each skill one at a time — exactly as pre-training does.

  Their words

  the goal is not to teach the model every possible skill within RL just as we don't do that within pre-training, right? Within pre-training, we're not trying to expose the model to, you know, every every possible you know, way that words could be put together, right? You know, we're it's it's rather that the model trains on a lot of things and then and then it reaches generalization across pre-training, right?

  youtube.com

• Our readingkorrents.com

  Models already generalise substantially from tasks that can be verified to tasks that cannot.

  Their words

  We already see substantial generalization from things that that verify to things that don't verify. We're already seeing that.

  youtube.com

American social psychologist at NYU Stern School of Business; author of The Happiness Hypothesis, The Righteous Mind and The Anxious Generation.

• Our readingkorrents.com

  The share of American teenagers who say that life often feels meaningless surged as Gen Z entered adolescence.

  America

  Their words

  But as soon as Gen Z entered the dataset, around 2013, meaninglessness surged.

  The Devil's Plan to Ruin the Next Generationafterbabel.com

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

From one piece Andrej Karpathy — “We’re summoning ghosts, not building animals” 2 beliefs, in the piece's order there

• Our readingkorrents.com

  Models are enormous only because the internet is bad: most of that capacity does memory work, and a better dataset would need far less of it.

  Their words

  Um so I almost feel like because the internet is so terrible, we actually have to sort of like build really big models to compress all that. Uh most of that compression is memory work instead of like cognitive work. But what we really want is the cognitive part actually delete the memory

  youtube.com

• 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

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

  Emergent abilities come from generalisation turning compositional, not merely from a training set containing a lot of stuff.

  Their words

  Yeah. So there's a subtlety here. Emerging capabilities don't just come from the fact that internet data has a lot of stuff in it. They also come from the fact that generalization once it reaches a certain level becomes compositional.

  youtube.com

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.

From one piece Why A.I. Isn't Going to Make Art 2 beliefs, in the piece's order there

• Our readingkorrents.com

  Art is what results from making an enormous number of choices, and a prompt contains almost none of them.

  Their words

  But let me offer a generalization: art is something that results from making a lot of choices.

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

• Our readingkorrents.com

  Writing that deserves a reader's attention is always the product of effort by the person who wrote it.

  Their words

  Let me offer another generalization: any writing that deserves your attention as a reader is the result of effort expended by the person who wrote it.

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

Coffee consultant and author of books on espresso, roasting and brewing, including The Professional Barista's Handbook and The Coffee Roaster's Companion; writes about brewing gear and roast control at scottrao.com.

• Likedtheir ownrcmnd.app

  Decent DE1

  Their words

  In the early days, the DE1 was considered "unforgiving" by many. I didn't mind, as the DE1's ability to control shot parameters more than made up for its early challenges.

  scottrao.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

  In instruction tuning, a smaller set of high-quality training examples outperforms a larger, noisier dataset.

  Their words

  training on a small set of high-quality data outperforms instruction-tuning on larger, noisier data.

  An Overview of Instruction Tuning Dataruder.io

Deep-learning researcher and Turing Award laureate; founder of the Mila institute.

• Our readingkorrents.com

  Access to large-scale generalist AI systems that could be weaponized should be limited, meaning their code and neural network parameters should not be released open-source.

  Their words

  At the same time, in order to reduce the probability of someone intentionally or unintentionally bringing about a rogue AI, we need to increase governance and we should consider limiting access to the large-scale generalist AI systems that could be weaponized, which would mean that the code and neural net parameters would not be shared in open-source and some of the important engineering tricks to make them work would not be shared either.

  FAQ on Catastrophic AI Risksyoshuabengio.org

Aeronautical engineer and former fighter pilot; analyses aircraft design and airline economics at Leeham News.

• Our readingkorrents.com

  Battery-electric airliners do not work even on optimistic assumptions, and the arithmetic said so before anyone built one.

  Their words

  Having done the sums before hitting the keys, I was a skeptic. It just didn’t work, even with optimistic assumptions around the parameters for the aircraft and its batteries.

  leehamnews.com

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

• Our readingkorrents.com

  Solely measuring skill at any given task falls short of measuring intelligence, because skill is heavily modulated by prior knowledge and experience.

  Their words

  We argue that solely measuring skill at any given task falls short of measuring intelligence, because skill is heavily modulated by prior knowledge and experience: unlimited priors or unlimited training data allow experimenters to "buy" arbitrary levels of skills for a system, in a way that masks the system's own generalization power.

  On the Measure of Intelligencearxiv.org