Atem Aguer
4 followers · 8 following · 1160 views
on the atlas — 88
- Superintelligence: The Idea That Eats Smart People3 savers
- [2309.10668] Language Modeling Is Compression4 savers
- My Approach to Building Large Technical Projects – Mitchell Hashimoto9 savers
- [2305.18290] Direct Preference Optimization: Your Language Model is Secretly a Reward Model8 savers
- Large Language Model: world models or surface statistics?4 savers
- New Book: AI is Good for You | Eric Jang5 savers
- The Techno-Optimist Manifesto - Marc Andreessen Substack3 savers
- The Secret to Good Writing: It's About Objects, Not Ideas - The Atlantic1 savers
- How to De-Risk a Startup - by Leo Polovets - Coding VC2 savers
- HALVA: Hallucination Attenuated Language and Vision Assistant1 savers
- Overleaf Example2 savers
- Efficient Guided Generation for Large Language Models1 savers
- 2301.066271 savers
- How and why we built our startup around small teams2 savers
- Laws of Tech: Commoditize Your Complement · Gwern.net26 savers
- Listen to Thomas Sowell | City Journal1 savers
- Solitude and Leadership by William Deresiewicz2 savers
- The History of Philosophy: From the Greeks and Christians to Moderns and Postmoderns – Discourses on Minerva1 savers
- Guest Blog Post - Attacking the DevTools | Microsoft Browser Vulnerability Research1 savers
- Sparse Matrix-Vector Multiplication with CUDA | by Georgii Evtushenko | Analytics Vidhya | Medium1 savers
- IEEE Xplore Full-Text PDF:36 savers
- Block Sparse Matrix-Vector Multiplication with CUDA | by Georgii Evtushenko | GPGPU | Medium1 savers
- Unwrap - 100 Exercises To Learn Rust1 savers
- Home | nand2tetris11 savers
- From Nand to Tetris: building a modern computer from first principles5 savers
- Ilya 30u3012 savers
- Extending Context is Hard | kaiokendev.github.io3 savers
- A history of NVidia Stream Multiprocessor2 savers
- Demystifying GPU Compute Architectures - by Babbage2 savers
- From Online Softmax to FlashAttention2 savers
- 2403.18103.pdf1 savers
- Diffusion models from scratch4 savers
- Understanding self-supervised and contrastive learning with "Bootstrap Your Own Latent" (BYOL) - imbue1 savers
- Building Diffusion Model's theory from ground up | ICLR Blogposts 20241 savers
- Jobs-to-be-Done: A Framework for Customer Needs | by Tony Ulwick | JTBD + Outcome-Driven Innovation2 savers
- On Christianity. An essay as a foreword for Tom… | by Nassim Nicholas Taleb | INCERTO | Medium3 savers
- Choose Good Quests - by Trae Stephens and Markie Wagner44 savers
- matejkajinic/the-manifestos: Curated overview of some of the most transformative manifestos that have shaped human thought over the last 300 years.1 savers
- How is LLaMa.cpp possible?4 savers
- Transformer Inference Arithmetic | kipply's blog17 savers
- The Little Book of Deep Learning7 savers
- The Only Way to Deal With the Threat From AI? Shut It Down | Time4 savers
- AI and the automation of work — Benedict Evans3 savers
- How Does a Database Work? | Let’s Build a Simple Database4 savers
- ELI5: FlashAttention. Step by step explanation of how one of… | by Aleksa Gordić | Jul, 2023 | Medium5 savers
- Why AI Will Save the World | Andreessen Horowitz8 savers
- OpenAI's plans according to Sam Altman3 savers
- How to Do Great Work80 savers
- The Bitter Lesson78 savers
- Thirty Observations at Thirty62 savers
- An Opinionated Guide to ML Research52 savers
- What I Wish Someone Had Told Me - Sam Altman52 savers
- Fast · Patrick Collison48 savers
- principles - Nabeel S. Qureshi41 savers
- Google "We Have No Moat, And Neither Does OpenAI"28 savers
- Moore's Law for Everything28 savers
- How to Learn Better in the Digital Age25 savers
- How to Optimize a CUDA Matmul Kernel for cuBLAS-like Performance: a Worklog24 savers
- ChatGPT Is a Blurry JPEG of the Web | The New Yorker22 savers
- What I Learned In 2023 - by Alexandr Wang20 savers
- Nadia Asparouhova | The independent researcher19 savers
- [2304.03442] Generative Agents: Interactive Simulacra of Human Behavior19 savers
- Simulators - LessWrong16 savers
- Why transformative artificial intelligence is really, really hard to achieve13 savers
- [REPOST] Epistemic Learned Helplessness | Slate Star Codex13 savers
- [1911.01547] On the Measure of Intelligence12 savers
- Understanding LSTM Networks -- colah's blog10 savers
- On Friendship and on Finding Your People - Alexey Guzey10 savers
- matrixcookbook.pdf9 savers
- Effective Spaced Repetition9 savers
- The Best GPUs for Deep Learning in 2023 — An In-depth Analysis7 savers
- Transformer Math 101 | EleutherAI Blog6 savers
- PyTorch internals : Inside 245-5D6 savers
- udlbook6 savers
- A student's guide to VC - Linda Tong6 savers
- pdf6 savers
- CUDA C++ Programming Guide5 savers
- The Case for American Seriousness5 savers
- [2106.09685] LoRA: Low-Rank Adaptation of Large Language Models5 savers
- Non-determinism in GPT-4 is caused by Sparse MoE - 152334H5 savers
- Non_Interactive – Software & ML4 savers
- Artificial General Intelligence Is Already Here4 savers
- Can LLMs Critique and Iterate on Their Own Outputs? | Eric Jang4 savers
- The implausibility of intelligence explosion | by François Chollet | Medium3 savers
- what the f* is e/acc - e/acc newsletter3 savers
- Sequence Modeling with CTC3 savers
- [2108.12409] Train Short, Test Long: Attention with Linear Biases Enables Input Length Extrapolation2 savers
- OpenAI's plans according to Sam Altman2 savers
highlights — 69
"Yes indeed -- you are going to write about things you can drop on your foot, and people, too. Green peppers, ears of corn, windshield wipers, or a grimy mechanic changing your car's oil. No matter how abstract your topic, how intangible, your first step is to find things you can drop on your foot."
The Secret to Good Writing: It's About Objects, Not Ideas - The AtlanticThinking means concentrating on one thing long enough to develop an idea about it. Not learning other people’s ideas, or memorizing a body of information, however much those may sometimes be useful.
Solitude and Leadership by William DeresiewiczWhat we don’t have, in other words, are thinkers. People who can think for themselves. People who can formulate a new direction: for the country, for a corporation or a college, for the Army—a new way of doing things, a new way of looking at things. People, in other words, with vision.
Solitude and Leadership by William DeresiewiczBut there’s one more thing I’m going to include as a form of solitude, and it will seem counterintuitive: friendship.
Solitude and Leadership by William DeresiewiczIntrospection means talking to yourself, and one of the best ways of talking to yourself is by talking to another person. One other person you can trust, one other person to whom you can unfold your soul.
Solitude and Leadership by William DeresiewiczSo solitude can mean introspection, it can mean the concentration of focused work, and it can mean sustained reading. All of these help you to know yourself better.
Solitude and Leadership by William DeresiewiczLeadership means finding a new direction, not simply putting yourself at the front of the herd that’s heading toward the cliff.
Solitude and Leadership by William DeresiewiczSo it’s perfectly natural to have doubts, or questions, or even just difficulties. The question is, what do you do with them? Do you suppress them, do you distract yourself from them, do you pretend they don’t exist? Or do you confront them directly, honestly, courageously? If you decide to do so, you will find that the answers to these dilemmas are not to be found on Twitter or Comedy Central or even in The New York Times. They can only be found within—without distractions, without peer pressure, in solitude.
Solitude and Leadership by William Deresiewiczthe idea that true leadership means being able to think for yourself and act on your convictions.
Solitude and Leadership by William DeresiewiczI try to build one or two demos per week
My Approach to Building Large Technical Projects – Mitchell HashimotoHumans often compete over perishable rewards (status, fame, etc.) versus compounding ones (technology, teams, etc.). If you avoid this mistake, you will be unstoppable in 5+ years.
What I Learned In 2023 - by Alexandr WangThe most practical way I know how to put this principle into action is by reading and rewriting the draft. Everything has a draft, even decisions about management. Step one is not to overthink the initial inspiration. Just get it out of your head and down on paper. Now you have something you can work with, something you can mold.
Commit to competence in this coming year“When men choose not to believe in God, they do not thereafter believe in nothing, they then become capable of believing in anything.”
Ayaan Hirsi Ali: Why I Am Now a Christian | The Free PressAtheism failed to answer a simple question: What is the meaning and purpose of life?
Ayaan Hirsi Ali: Why I Am Now a Christian | The Free PressPrinciples are usually the byproduct of success, not the cause of it.
principles are products of practice, not the reverse | Majd AlsadoThe core idea is trying my best to not kid myself: when my engagement with a piece of content is active and effortful then it’s learning, when it’s passive it’s entertainment. When I create I learn. When I consume I just relax.
How to Learn Better in the Digital Agedemanded an unambiguous mechanical procedure for determining whether any mathematical statement was true or false.
Complexity Theory’s 50-Year Journey to the Limits of Knowledge | Quanta MagazineJust a year later, Gödel delivered the first blow to Hilbert’s dream. He proved that a self-defeating statement like “this statement is unprovable” could be derived from any appropriate set of axioms. If such a statement is indeed unprovable, the theory is incomplete, but if it’s provable, the theory is inconsistent — an even worse outcome. In the same paper, Gödel also proved that no mathematical theory could ever prove its own consistency.
Complexity Theory’s 50-Year Journey to the Limits of Knowledge | Quanta MagazineThat was the motivation for Hilbert’s second condition, completeness: the requirement that all mathematical statements be either provably true or provably false.
Complexity Theory’s 50-Year Journey to the Limits of Knowledge | Quanta Magazineconsistency, was the essential requirement that mathematics be free of contradictions: If two contradictory statements could be proved starting from the same axioms, the whole theory would be unsalvageable.
Complexity Theory’s 50-Year Journey to the Limits of Knowledge | Quanta MagazineHe hoped to start from a few simple assumptions, called axioms, and derive a unified mathematical theory that met three key criteria.
Complexity Theory’s 50-Year Journey to the Limits of Knowledge | Quanta MagazineIf so, then P = NP: The two classes are equivalent. If that’s the case, there must be some algorithm that makes it trivial to solve enormous sudoku puzzles, optimize global shipping routes, break state-of-the-art encryption and automate the proofs of mathematical theorems. If P ≠ NP, then many computational problems that are solvable in principle will in practice remain forever beyond our grasp.
Complexity Theory’s 50-Year Journey to the Limits of Knowledge | Quanta MagazineIt implies that Google Deepmind knew, and found it trivial enough to write as a throwaway sentence in a paper. It implies that I should be a lot more bullish on them, and a lot more bearish against every other wannabe foundation model org that’s still working on dense models only.
Non-determinism in GPT-4 is caused by Sparse MoE - 152334HI think that misunderstands the problem. If a partner at a law firm wants a first draft of a paper, they want to be able to shape the parameters in completely different ways to a salesperson at an insurance company challenging a claim, probably with a different training set and certainly with a bunch of different tooling.
AI and the automation of work — Benedict EvansYou can rarely go to a law firm and sell them an API key to GCP’s translation or sentiment analysis: you need to wrap it in control, security, versioning, management, client privilege and a whole bunch of other things that only legal software companies know about (there’s a graveyard of machine learning companies that learnt this in the last decade).
AI and the automation of work — Benedict Evansif we make steam engines more efficient, then they will be cheaper to run, and we will use more of them and use them for new and different things, and so we will use more coal.
AI and the automation of work — Benedict EvansThe Lump of Labour fallacy is the misconception that there is a fixed amount of work to be done, and that if some work is taken by a machine then there will be less work for people.
AI and the automation of work — Benedict EvansAutomation plus the Jevons Paradox meant more jobs.
AI and the automation of work — Benedict EvansIn reality, more specific solutions from application-focused companies will come up and be adopted first because they work, then it'll be increasingly hard to convert enterprise customers.
LLM agents and integration dead-ends - by Nathan LambertThe fact of the matter is that digital products make it uniquely easy to trick yourself into thinking that you’re learning when you are actually being entertained.
How to Learn Better in the Digital AgeFirst, you need to know who your customers are (a more challenging question than it seems), then you have to figure out where to find them, and lastly, how to reach them efficiently and at scale.
Minimum Viable Distribution for StartupsWe envision that AI agents for marketing, however, will take a broader, more LAM-like approach to delivering results for marketing teams, by using an LLM interface to connect data, tools, and domain-specific agents in the pursuit of a high-level task.
Toward Actionable Generative AIThe picture is more complicated in practice, however, as a great deal of manual effort is necessary to integrate an LLMs output into the complete process of, for example, conceiving of a new campaign and rolling out the results.
Toward Actionable Generative AIWith that in mind, LAMs should focus on taking the reins on repetitive tasks and other busywork—the kind of thing most of us don’t want to do in the first place—that gets in the way of the kind of meaningful, high-value endeavors that we’re best at.
Toward Actionable Generative AIthe process that surrounds it.
Toward Actionable Generative AIWorkflows are a set of processes that we use to accomplish a goal. To make this concrete, I have a workflow around reading that includes the way I research, select, and collect reading material, from books to essays and articles, the times and places where I read them, the practices I use during and after reading. All of these processes fit together to form my reading workflow. There are a dozen or more tools I use, most of which don't interoperate in any meaningful way, and span the digital and the analog. My reading workflow spans days and weeks, and locations ranging from my office, my bedr…
Tools embody mediumsAI to help us perceive, think, understand, and do more. AI that you could use like an extension of your own mind. Today, some applications of machine learning fall into this category, but they’re few and far between. Yet, I believe this is where the true potential of AI lies.
AI is cognitive automation, not cognitive autonomyFor instance, if you take a model like StableDiffusion and integrate it into a visual design product to support and expand human workflows, you’re turning cognitive automation into cognitive assistance.
AI is cognitive automation, not cognitive autonomyOn acceleration, a main bottleneck will be autonomy.
What will GPT-2030 look like?The core mistake the automation-kills-jobs doomers keep making is called the Lump Of Labor Fallacy. This fallacy is the incorrect notion that there is a fixed amount of labor to be done in the economy at any given time, and either machines do it or people do it – and if machines do it, there will be no work for people to do. The Lump Of Labor Fallacy flows naturally from naive intuition, but naive intuition here is wrong. When technology is applied to production, we get productivity growth – an increase in output generated by a reduction in inputs. The result is lower prices for goods and serv…
Why AI Will Save the World | Andreessen Horowitzbuild only what you need as you need it and adopt your software as quickly as possible.
My Approach to Building Large Technical Projects – Mitchell HashimotoThese aren't "minimum viable products", because they really aren't viable, but they're good enough to provide an engineer some valuable self-reflection.
My Approach to Building Large Technical Projects – Mitchell HashimotoThe goal is to get to a demo.
My Approach to Building Large Technical Projects – Mitchell HashimotoSeeing the progression of "1 test passed", "4 tests passed," "13 tests passed" and so on is super exciting to me. I'm running some code I wrote and it's working. And I know that I'm progressing on some critical sub-component of a larger project.
My Approach to Building Large Technical Projects – Mitchell HashimotoSelfishness doesn’t require malice or even sentience. When an AI automates a task and leaves a human jobless, this is selfish behavior without any intent. If competitive pressures continue to drive AI development, we shouldn’t be surprised if they act selfishly too.
The Darwinian Argument for Worrying About AI | TimeAt first, the CEO carefully monitors the work, but as months go by without error, the AI receives less oversight and more autonomy in the name of efficiency. It occurs to the CEO that since the AI is so good at these tasks, it should take on a wider range of more open-ended goals: “Design the next model in a product line,” “plan a new marketing campaign,” or “exploit security flaws in a competitor’s computer systems.” The CEO observes how businesses with more restricted use of AIs are falling behind, and is further incentivized to hand over more power to the AI with less oversight. Companies t…
The Darwinian Argument for Worrying About AI | TimeYou could consider this a form of epistemic learned helplessness, where I know any attempt to evaluate the arguments is just going to be a bad idea so I don’t even try. If you have a good argument that the Early Bronze Age worked completely differently from the way mainstream historians believe, I just don’t want to hear about it. If you insist on telling me anyway, I will nod, say that your argument makes complete sense, and then totally refuse to change my mind or admit even the slightest possibility that you might be right.
[REPOST] Epistemic Learned Helplessness | Slate Star CodexAI risk is string theory for computer programmers. It's fun to think about, interesting, and completely inaccessible to experiment given our current technology. You can build crystal palaces of thought, working from first principles, then climb up inside them and pull the ladder up behind you.
Superintelligence: The Idea That Eats Smart PeopleThis business about saving all of future humanity is a cop-out. We had the same exact arguments used against us under communism, to explain why everything was always broken and people couldn't have a basic level of material comfort.
Superintelligence: The Idea That Eats Smart PeopleFor all we know, human-level intelligence could be a tradeoff. Maybe any entity significantly smarter than a human being would be crippled by existential despair, or spend all its time in Buddha-like contemplation.
Superintelligence: The Idea That Eats Smart People