What the smart people are recommending. 7867 books, tools, and products endorsed by the thinkers absorb.md tracks. Ranked by how many times each has been recommended across compiled podcasts, papers, posts, and tweets.
βThe paper that blew my mind was InstructGPT, because it pointed out that you can take the pretrained model, which is autocomplete, and if you just fine-tune it on text that looks like conversations, tβ¦β
βone example of this is that there was a few papers came out of Cornell over the last few months like one of them was they published this after good financial cryptography 2016 workshop that they were β¦β
βIn this paper, we address this by introducing adaptive canonicalization, a general framework in which the canonicalization depends both on the input and the network. Specifically, we present the adaptβ¦β
βLeveraging recent advancements in multi-modal LLMs, we propose SLEDGE: Step-by-step LayEred Design GEnerator to model each update to a design as an atomic, layered change over its previous state, whilβ¦β
βThis paper introduces Physics-Informed Extreme Learning Machines (PIELMs) as fast alternative to PINNs for solving both forward and inverse problems in financial PDEs.β
β# JEPA as a Neural Tokenizer: Learning Robust Speech Representations with Density Adaptive Attentionβ
βWe discuss this, along with the other implications of this research, in our blog: https://www.anthropic.com/research/automated-alignment-researchersβ
βFor the full study, see here: https://alignment.anthropic.com/2026/automated-w2s-researcher/β
βRead the full letter here: https://arxiv.org/abs/2510.09520v1β
βweird thing to put in print https://www.upi.com/Archives/1986/10/13/Forbes-ranks-Walton-as-richest-American-for-second-year/4151529560000/β
βHis full walkthrough is linked in the description.β
βWe address this challenge by unifying calibration with computation, granting the quantum error correction process a dual role: its error detection events are not only used to correct the logical quantβ¦β
βTo address these issues, we introduce STARC-9 (STAnford coloRectal Cancer), a large-scale dataset for multi-class tissue classification.β
βWe hope that the simplicity and theory-friendly ecosystem offered by LeJEPA will reestablish self-supervised pre-training as a core pillar of AI researchβ
βWell, let's get to that. I mean, so, so yesterday, you know, and I'm glad to say this was the cover of Nature, Nature Journal yesterday.β
βCounterpoint: https://www.niemanlab.org/2026/04/do-links-hurt-news-publishers-on-twitter-our-analysis-suggests-yes/β
βThere was an algorithm that was theorized by a guy named Peter Shor. I think I talked about this on a prior episode in 1994 called Shor's algorithm. Today, it's kind of the commonly well-known model fβ¦β
βwe've been publishing at the general of laward we've been P publishing prompts that turn the AI into a tutor into a mentor that uh into a student you have to explain stuff toβ
βAnd then, a couple years ago, I think in 2023, there was another computer scientist named Oded Regev from NYU who published another paper that showed a faster different approach to Shor's algorithm.β
βThis paper introduces KunLunBaizeRAG, a reinforcement learning-driven reasoning framework designed to enhance the reasoning capabilities of large language models (LLMs) in complex multi-hop question-aβ¦β
βLast week I shared our recent arXiv paper detailing the worldβs largest entangled state, with 120 qubits at 56% fidelity with a shot-retention rate of 28%, run on Heron R2 ibm_aachen.β
β# Cambrian-S: Towards Spatial Supersensing in Videoβ
βThis is not in the technical report but this is from the paper actually that kind of originates our tokenization approach. \n Paper name is called one tokenizer to rule them all.β
βThe corresponding paper is actually the art of asking you can find in in the archive that actually has research on this particular focus.β
βAll of them in command line translate but it includes rulebased \n filtering again difficulty filtering which come up again this difficulties as an important metric and also we looked at some some of β¦β
βFor for the merging recipe. We have another paper here a simmer merge that is actually not using only one merging technique \n but kind of select the merging techniques based on some of the metrics.β
βThis paper is concerned with Quantum Picturalism, a novel visual mathematical language for quantum physics. Originally developed over two decades ago to explore the foundational structure of quantum tβ¦β
βWe present Legilimens, a continuous learning system for the mobile edge's System-on-Chip GPUs.β
β_ArXiv paper co-authored by Andrew Ng_β
βI know you have too many suggestions, but consider this anyway. Clearly in your territory.β
βVitalik already quoted some of the wonderful stuff that he released in his article in uh, April this year. And I want to focus on that, the importance of privacy.β
βThe Urgency of Interpretability: Why it's crucial that we understand how AI models workβ
βHere @robbieking1000 considers the issue from the perspective of complexity theory. http://quantumfrontiers.com/2025/12/11/can-ai-predict-the-quantum-universe/β
β# Constructive interference at the edge of quantum ergodic dynamicsβ
βWe present an experimental study of magic state cultivation on a superconducting quantum processor.β
βThis conversation with Amjad, my recent one with New York Assembly member Alex Morris, and the series of episodes we did last year on SB1047 are all good examples of this.β
βAdvice on how to interview for a faculty job, from an authoritative source. http://quantumfrontiers.com/2026/01/04/nicoles-guide-to-interviewing-for-faculty-positions/β
βIn this paper, we propose improved methods for training world models that enable efficient gradient-based planning.β
βI loved your urgency of interpretability essay.β
β# International AI Safety Report 2025: Second Key Update: Technical Safeguards and Risk Managementβ
βI read the Alpha Evolve paper, or to be more precise, I fed it into Notebook LM and had it make a podcast that I could then listen to that would explain it to me.β
βThe Trump administration just announced that we'd be rescinding what's known as the Biden diffusion rule, which was a rule that came out in January.β
βSpecReason: Fast and Accurate Inference-Time Compute via Speculative Reasoning https://arxiv.org/abs/2504.07891β
βI don't know if you read Sam Altman's blog that was out this week called The Gentle Singularity. Uh I would encourage everybody to actually go read it.β
βan editor from nature approached me and said like would you write a review paper on uh on on the topic to introduce the topic to the larger scientific community and that's how this paper came about.β
βthere's a paper you know famous paper u the that that sort of proposes this idea resnet um the lead author was coming her from at the time Microsoft research in Beijing. This is the most cited paper iβ¦β
βJungsang Kim, Ken Brown, Luming Duan, Peter Maunz and I published sort of this architecture paper... 2013 or 2014 how you might build a scalable modular quantum computer.β
βWe propose a novel framework for one-shot visual imitation learning via world-model-guided trajectory generation... Our method is evaluated on two simulated benchmarks and three real-world robotic plaβ¦β
βBuilding deep learning models that can reason about their environment requires capturing its underlying dynamics. Joint-Embedded Predictive Architectures (JEPA) provide a promising framework to model β¦β
βWe combine our findings to propose a model that outperforms two established baselines, DINO-WM and V-JEPA-2-AC, in both navigation and manipulation tasks. Code, data and checkpoints are available at hβ¦β
βOur work addresses the problem of learning latent actions world models on in-the-wild videos, expanding the scope of existing works that focus on simple robotics simulations, video games, or manipulatβ¦β
βWe apply an Information Bottleneck framework to compare human conceptual structure with embeddings from 40+ LLMs using classic categorization benchmarks.β
βwe propose Guillotine, a hypervisor architecture for sandboxing powerful AI models -- models that, by accident or malice, can generate existential threats to humanity.β
βone of the best responses to it was by another writer named Derek Thompson who wrote a article called Nobody Knows Anything, which I think is a reference to a famous take by legendary Hollywood writerβ¦β
βthere was an article that went viral this week by Matt Schumer called something big is happening where he talked about this career opportunity that's going to be available to kind of AI early adoptersβ¦β
βthis is from an article by Philip Ragaway a very famous one the moral character of cryptographic work. I would uh highly encourage people to read it.β
βthe Hoover Institution just published something this morning and it's a complete indictment of what the billionaire tax was trying to do... they ran 100,000 runs and in 71% of those runs, it comes outβ¦β
βI think the Stanford did a study on AI optimism. They simply asked the question, do you think AI is going to be more beneficial than harmful? Something like 80% of people in China said yes.β
βI saw an oped in the Wall Street Journal to that effect that we shouldn't try and find an offramp, we should just keep going with this.β
βwe actually released a very detailed paper on command a which describes exactly what we did but similar to others we do a large pre-training phaseβ
βGemma Scope 2: helping the AI safety community deepen understanding of complex language model behaviorβ
βI know you know Daniel Cocatello I know you guys... you're familiar with his AI 2027 scenarioβ
βWe train models to Predict Ego-centric Video from human Actions (PEVA), given the past video and an action represented by the relative 3D body pose.β
βIn this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models.β
βthere's a really great article from Tomasan goo at theory ventures uh that compared the original JPT interface with the command line interface of the 1980sβ
βWe present the RELRaE framework, a framework that employs large language models in different stages to extract and accurately label the relationships implicitly present in the XML schema.β
βWe introduce SIMA 2, a generalist embodied agent that understands and acts in a wide variety of 3D virtual worlds.β
βHere we introduce AlphaQubit 2, a neural-network decoder that achieves near-optimal logical error rates for both surface and color codes at scale under realistic noise.β
βyou read the AI 2027 report and now you're talking about AI super coders training models and then having this just recursive loopβ
βJosh Gans who's an economist at the University of Toronto has a really nice paper outlining that in an area where we don't know what's going to happen where harms and benefits are both emergent that pβ¦β
βsome of them written by my colleague Daniel Schwarz that those papers showing that 01 preview leads to 10 to 30% improvements in quality and speedβ
βI understand it's not a well-loved benchmark but the Atari 100K benchmark that people have been doing for the last few years I think is a really good way of addressing thatβ
βBBF, the bigger, better, faster Atari 100k kind of uh top dog, it generates really really impressive learning scores in 100k frames. But if we just add a few frames of latency in there, it falls apartβ¦β
βthe gateau agent the the transformer-based learn to play all the Atari games...they demonstrated negative transfer learning where if you took a game a model trained on a dozen agents...it was harder tβ¦β
βThe ARC challenge is like an IQ test for machines that released in 2019... I was releasing this as a way to test and illustrate my definition of intelligence.β
βI first heard about Bitcoin when I was giving talks about quantum money you know in in 2012 or so...could you have quantum money that anyone could verifyβ
βthis protocol uh you mentioned for generating cryptographically certified random bits using a quantum computer. Uh that was a very good example of of this approachβ
βthere is um a conjecture or belief called the quantum extended church touring thesis which would say that every physical system can be efficiently simulated by a quantum computerβ
βinfluential publications like her nature paper on the single atom transistor. She even made these really complex scientific ideas accessible to a broader audience through her TEDex talkβ
βGaussian Embeddings: How JEPAs Secretly Learn Your Data Densityβ
βWe hope you've enjoyed this fascinating exploration of Michelle Simmons's immense impact on quantum computing based on Urugo Schnep's insightful essay. If you'd like to explore more about these topicsβ¦β
βAnthropic wrote a great blog post on this that I'd encourage you to check out.β
βI'm going to start with a quote and it's from Warren Buffett in his 2020 shareholder letter to investors and he said, 'In its brief 232 years of existence, there has been no incubator for unleashing hβ¦β
βWe introduce LessIsMore, a training-free sparse attention mechanism for reasoning tasks, which leverages global attention patterns rather than relying on traditional head-specific local optimizations.β
βa computer scientist named Oded Regv from NYU who published another paper that showed a faster different approach to Shor's algorithm... reduced the number of quantum operations required to factor a lβ¦β
β# Observation of disorder-induced superfluidityβ
βPeter Shor published his algorithm that showed that if we had a quantum computer, we could factor numbers into their primes. We could therefore break codes, break public key encryption systems. And thβ¦β
βSystem card: https://www-cdn.anthropic.com/53566bf5440a10affd749724787c8913a2ae0841.pdfβ
βDid you guys see the LP slides that went on the internet from Toma Brava's LP conference? They kind of highlight that within the broad marketscape, there are companies that are not just going to sit iβ¦β
βWe present DINO-world, a powerful generalist video world model trained to predict future frames in the latent space of DINOv2.β
βDwarkesh, did you read his recent blog post on his AI timelines? Oh, on continual learning, yeah?β
βIt will like change the test to fit the mistakes it made or sometimes delete the tests. It's really fascinating behavior that Anthropic published research on.β
βwhen you were with me on the AI engineer conference you talked about the the touring uh paper which you love and got you started uh in some ways on your machine learning journeyβ
βI invited a talk on the model spec for AI engineer and that was the most viewed uh talk of all of that we've ever had... the model spec is a perfect example of when the models are very capable. You stβ¦β
βwe publish techniques like instruction hierarchy and so with instruction hierarchy you sort of indicate that hey there's this message is from the system this message is from the developer this messageβ¦β
βSo I think uh you know Paul Graham uh the founder of Y Combinator you know the best accelerator in the world um wrote this uh essay about being resourceful and he talks about the qualities of foundersβ¦β
βWe introduce UQ, a testbed of 500 challenging, diverse questions sourced from Stack Exchange, spanning topics from CS theory and math to sci-fi and history, probing capabilities including reasoning, fβ¦β
β# Less Is More: Training-Free Sparse Attention with Global Locality for Efficient Reasoningβ
βOur results establish that both learning and sampling can be performed efficiently in the beyond-classical regime, opening new possibilities for quantum-enhanced generative models with provable advantβ¦β
βWe find surface subgroups in certain one-relator groups with torsion and use this to deduce a profinite criterion for a word in the free group to be primitive.β