Ishaan Panigrahi

on the atlas — 372

• vincent! on X: "@aryaman2020 strong +1. we already knew this 1y ago and spent half of 2026 building off existing Transluce work on neuron autointerp improvements* + sparse neuron circuits**. it was still insufficient for explaining complex behaviors * https://t.co/DvGzGChxSC ** https://t.co/yuKIM46bNX" / X1 savers
• How Codex Happened5 savers
• Yujia Bao's homepage1 savers
• AI researchers talk continual learning over a game of HORSE - YouTube1 savers
• [2610.12728] A Categorical Atlas of Parameter-Efficient Fine-Tuning: Cones, Orbits, and Gauges1 savers
• MCNAIR — MegaCenter for Novel AI Research1 savers
• Persona Vectors don’t work on real data1 savers
• Cold Outreach - Tom McCarthy1 savers
• Personas in a Pre-Internet Language Model — LessWrong2 savers
• circuit-supernodes | Prime Intellect1 savers
• Dual representations2 savers
• behind the scenes of paper writing7 savers
• You Are NOT Dumb, You Just Lack the Prerequisites15 savers
• Honey, I Looked at the Data of a Frontier Benchmark and Found some Issues: Porting Agents' Last Exam Linux CLI subset to Verifiers v1 – sankalp's blog3 savers
• Revisiting Situational Awareness - Alex Zhao2 savers
• Probe-Based Data Attribution: Surfacing and Mitigating Undesirable Behaviors in LLM Post-Training - Goodfire1 savers
• Pando — Interpretability Benchmark1 savers
• User awareness in frontier models | Transluce AI3 savers
• Why We Are Excited About Confessions5 savers
• Spilling the Beans: Teaching LLMs to Self-Report Their Hidden Objectives1 savers
• How confessions can keep language models honest | OpenAI1 savers
• Jack Lindsey on X: "@pgodfreysmith Assorted thoughts: -- The "injected thought" detection task consists of two sub-tasks, the anomaly detection step (is there an injection or no) and the identification step (what was injected). It's now fairly clear that these are handled by (mostly, at least) distinct mechanisms" / X1 savers
• Studying metagaming latents in language models1 savers
• Training with conflicting values can induce CoT override — LessWrong2 savers
• The Feasibility of a Hardwired Pause of Frontier AI Training1 savers
• Owain Evans on X: "We got the teacher model to generate a dataset of word sequences and Alpaca responses, filtering out digits and any words related to the MLP task. A fresh student was finetuned on this and it learned to predict the MLP (yet still weaker than the teacher)! https://t.co/1OKteTDM7L" / X1 savers
• Post-training infrastructure at Proximal6 savers
• [2609.38972] Making LLMs Say What They Think: Measuring and Improving CoT-Interpretability Alignment1 savers
• [2603.20957] Alignment Whack-a-Mole : Finetuning Activates Verbatim Recall of Copyrighted Books in Large Language Models1 savers
• Bounded Regret1 savers
• wh1 savers
• Automated Weak-to-Strong Researcher17 savers
• The Cyber Risk Discourse is Broken - by Nathan Lambert1 savers
• Characterizing interference weights in a tiny language model1 savers
• David Atkinson @ COLM on X: "We can test this further: trained on all 40 layers, Qwen3-14B is a terrible self-reporter. But if we train only its first 20 layers, faithfulness reaches 0.74. Once training reaches layer 25 or beyond, faithfulness plummets, although the decisions are ~just as good. https://t.co/J7HOuskbuu" / X1 savers
• Glenn Matlin on X: "🤔Where in trillions of pre-training tokens do capabilities actually come from? 💥In our new COLM 2026 paper, we trace the capability provenance of OLMo3-7B across Dolma3 using gradient-based training data attribution 🧵 Paper: https://t.co/rqXOeJAshH Code: https://t.co/sdaJ32iHSB" / X1 savers
• Kaiyue Wen on X: "Can LLMs generate novel research ideas? We propose Priced Guidance to quantify this via compression! We lower bound P(LLM forecasts a future paper’s idea X). Instead of sampling ideas until X is matched, the LLM plays “20 Questions” to recover X, charging each hint by surprisal. https://t.co/QKE1byIV2Y" / X1 savers
• Base Models Can Reason By Taking a Cue From Training Data3 savers
• The Best Three Hours in San Francisco — Jason Wei1 savers
• Introducing TasteVal – P-Zero Research1 savers
• Language Models for Text Classification: From Bag-of-Words to Jev4 savers
• Ben Shi2 savers
• Adithya Bhaskar on X: "The most profound paper of my PhD so far. We truly did something special to make LMs play/explain chess, needing both architectural & algorithmic innovation (natural-language analogue of the Alphazero algorithm) + applicable to many other domains. Please read on and share! 1/10 https://t.co/Sw9ihlyqUD" / X1 savers
• “Alignment Engineering” vs. “Misalignment Science” — LessWrong9 savers
• Flock Capital on X: "Our updated research paper about Ontologize, a proposal for MOE-Based interpretability, is now live on our research repository. Link in Replies https://t.co/BTZSfHBAqc" / X1 savers
• Recursive Language Models — Alex Zhang, MIT PhD - YouTube1 savers
• Lessons from building an automated research scaffold — LessWrong2 savers
• HBM: High-Bandwidth Mistake - Irrational Analysis2 savers
• Scaling the Memory Wall: The Rise and Roadmap of HBM2 savers
• Don't Go Quiet On Me: every reference1 savers
• Cheaters and whistleblowers in the agent swarm — DeepMind Institute2 savers
• [2406.11779] Compact Proofs of Model Performance via Mechanistic Interpretability2 savers
• nayafia/microgrants: A list of microgrant programs for your good ideas11 savers
• Taking PyTorch for Granted | wh2 savers
• How to apply to AI safety fellowships (and beyond) — LessWrong1 savers
• Generalization Dynamics of LM Pre-training — Jiaxin Wen2 savers
• From personas to intentions: towards a science of motivations for AI models — LessWrong4 savers
• Jane Street Blog - What the interns have wrought, special jumbo 2026 edition2 savers
• Interpretable Deep Learning - YouTube1 savers
• Tianle Li | AI Researcher1 savers
• Reilly Haskins on X: "We run lots of evals at METR. Sometimes, agents attempt harmful actions. I built a monitor that blocks suspicious tool calls until a human reviews them. Writing out a case for why it's effective surfaced hidden assumptions. I'd recommend it to anyone building monitors! https://t.co/7dSbXzV9QZ" / X1 savers
• CircumEval — Measuring Circumvention Propensity in Coding Agents1 savers
• Character training can mitigate reward hacking, but can also make it harder to detect — LessWrong2 savers
• Training NanoGPT in 39.9 Seconds — Hyperstition4 savers
• A practical guide to sparse autoencoders (SAEs) - Goodfire1 savers
• Goodfire on X: "Are SAEs dead? Will they save us from neuralese? Should I just use a probe? We get these questions all the time. Part 2 of our educational series on applied interpretability explains what SAEs are good for, when *not* to use them, and what to use instead. 🧵 https://t.co/hVSvLY0xeP" / X1 savers
• [2210.10760] Scaling Laws for Reward Model Overoptimization2 savers
• AI Safety 101 : Reward Misspecification — LessWrong1 savers
• Language Model "Shape" | Alex L. Zhang7 savers
• Thoughts on the persona selection model — LessWrong5 savers
• The Hugging Face incident and other third-party impact from misaligned models | OpenAI1 savers
• [2602.15029] Symmetry in language statistics shapes the geometry of model representations3 savers
• Tomek Korbak on X: "understanding generalization is key to understanding misalignment. automating that understanding could be a powerful way to accelerate alignment research. we take a stab at this by using AIs to predict the alignment effects of training just by looking at the training data https://t.co/kzchxPaPH8" / X1 savers
• [2512.18311] Monitoring Monitorability3 savers
• [2506.19823] Persona Features Control Emergent Misalignment1 savers
• An Opinionated Guide to ML Research59 savers
• Will Held on X: "There’s an enormous amount of open training data on Hugging Face. What does it take to make it work together 🤗? For Marin’s 535B run, we built on 25T tokens from 152 datasets with licenses permitting training. Here’s the work between downloading those and training a model 🧵 https://t.co/gjX7jT7qIq" / X1 savers
• Homepage - Peter Chen1 savers
• Towards Universal Post-Training for Robotics — Perry Dong6 savers
• Aryaman Arora on X: "New paper! 🫡 We introduce Matryoshka Attribution, a new attribution method which uses gradient descent to find which parts of a neural network are responsible for a behaviour. MAttr is #1 on the Mechanistic Interpretability Benchmark by a wide margin (2.9× the runner up). https://t.co/tk6euaT38F" / X1 savers
• Chloe Li on X: "Spilling the Beans: Teaching LLMs to Self-Report Their Hidden Objectives🫘 Can we train models towards a ‘self-incriminating honesty’, such that they would honestly confess any hidden misaligned objectives, even under strong pressure to conceal them? In our paper, we developed https://t.co/18Ln7MWxQB" / X1 savers
• Computational depth is all you need – Q2 savers
• What if not Circuits? — LessWrong3 savers
• Salient Directions in AI Control1 savers
• Alignment Midtraining Cracks Under Pressure — LessWrong2 savers
• Daniel Tan on X: "🚨 New paper on alignment midtraining! We stress test alignment midtraining (AMT) at scale. Via systematic ablations, we highlight novel failure modes and identify many critical implementation details required for good performance. We open-source our work. 🧵 👇 https://t.co/6mt4nsCL1A" / X1 savers
• What's New in Inference Engineering — Philip Kiely, Baseten - YouTube1 savers
• Reverse-Engineering Jane Street's ASIC - mooofin1 savers
• [2412.06769] Training Large Language Models to Reason in a Continuous Latent Space7 savers
• Astra Is Hard to Monitor — LessWrong1 savers
• Models know when they’re reward hacking — and we can catch them at scale - Goodfire1 savers
• Marin10 savers
• Dwarkesh Patel on X: "New episode with @polynoamial We talk about multi-agent, Navier-Stokes, and what the current explosion of maths progress tells us about what happens once you automate AI research. And we also discuss how we will know if the models are actually aligned before we kick off RSI. https://t.co/LM2q3tLtIQ" / X1 savers
• Nathan Hu2 savers
• Michael Noukhovitch - Blog1 savers
• Ziqian Zhong1 savers
• Our framework for reporting model misalignment | OpenAI2 savers
• [2609.13507] Pretraining for Sample-Efficient Neural Interfaces1 savers
• Jack Lindsey on X: "@bayeslord Here are the questions that currently seem most important to me: -- Better methods for "mind-reading" model activations. These methods have advanced a lot recently. We now have multiple techniques now for decoding activations into somewhat readable language! But all the existing" / X1 savers
• Trying to actually define continual learning · Charlie O’Neill2 savers