we’re in for another boom of cool interpretability research like The Lottery Ticket Hypothesis & Manifold Hypothesis with the rise of frontier open models get to really dig into how frontier intelligence is able to model concepts across modalities (text, image, video, audio) and how LLMs mechanistically execute search programs during rollouts to find the right trajectories across tasks

If an LLM's mind holds concepts humans haven't even named yet, how would we ever go looking for them? 🔍 Interpretability research has mostly searched for concepts we already have words for: refusal, truthfulness, deception. But the space of distinctions an LLM actually uses for computation is almost certainly larger than our finite vocabulary. Finite descriptions can only denote a countable number of properties, while the space of properties over an internal state is mathematically far larger. Somewhere in that gap sit distinctions no human concept was ever built to describe. 💡 That's the premise behind "Xeno-Interpretability." Its key move is separating two questions that are usually bundled together: can a representation be experimentally located and causally manipulated, and can it be explained in human terms? A representation can be robustly findable and behaviorally important even when no human category fits it — the paper calls these "xeno-representations." 🌐 This isn't just philosophical. In multi-agent systems, model-native representations could quietly stabilize and propagate through agent-to-agent messages while staying only partially visible in the human-readable parts of the conversation, a real concern for AI safety. Title: Xeno-Interpretability: Investigating the Alien Minds of LLMs URL: #Interpretability# #AISafety#

Be part of our weekly Generative AI study group every Friday at 8 am PT. Exchange model merging techniques, discuss TPU orchestration, dig into interpretability frameworks, and more! Make sure to register at and join the #generative-ai# Slack channel. See you there!

SOME DETAILS FROM OPUS 5.5 BLOG: > most cyber tasks still get routed to Opus 4.8 > if often knows it’s being evaluated > that makes real world behavior harder to assess > best alignment results yet on 2,000 scenarios > 85% fewer attempts to cross containment boundaries > matches or beats Mythos 5.1 on biology > they wants to rely less on reading CoT and more on interpretability > tightening RL env filtering as a major source of misalignment > they still argue for government regulations > text watermarking is included

As reasoning models consume more tokens and AI systems become more expensive to run, understanding what those tokens actually buy is becoming increasingly important. In this episode, @Stanford professor and Big Spin co-founder @ChrisGPotts joins us to discuss AI tokenomics and his research into “tokenflation”—the possibility that token usage is growing faster than the measurable value those tokens produce. We explore how to measure the return on AI spending, why benchmarks alone provide an incomplete picture of model progress, and what inference-time scaling means for the economics of increasingly capable models. Chris also explains why expert AI users tend to get better results by challenging and iterating with models, how AI fluency affects outcomes, and why more efficient architectures could change the underlying economics. We also discuss DSPy, interpretability, the limits of today’s transformer architectures, and where Chris sees opportunities for more fundamental innovation in AI. 🗒️ Full show notes: 📖 CHAPTERS =============================== 00:00 - Introduction 05:38 - Linguistics in the Age of Language Models 09:26 - Scale Limitations in NLP Research 12:54 - Challenging the Bitter Lesson Mindset 15:12 - Relationship Between Data, Mechanistic Interpretability, and Efficiency 17:03 - DSPy 21:32 - Prompt Optimization and Model Variability 24:35 - Tokenomics and the Rising Cost of AI 28:13 - Measuring Token Purchasing Power with a CPI 32:18 - Inference-Time Scaling 35:36 - Defining Value Across Different AI Tasks 38:33 - AI Value Creation 40:38 - Predicting AI Costs 42:25 - Tokenflation 46:22 - AI Fluency 50:12 - Key Lessons of AI Fluency Work 54:44 - Future Directions

The complete roadmap for becoming dangerous at LLM engineering from zero to production : Watching tutorials won’t make you an LLM engineer. Building these projects will: - Train a small working model - Master embeddings from scratch - Create your own tokenizer - Code RoPE and ALiBi yourself - Write attention by hand - Construct multi-head attention - Assemble a complete Transformer block - Test different training objectives - Implement generation strategies - Add speculative decoding - Build efficient KV caching - Explore MQA, GQA, and MLA - Solve long-context challenges - Understand FlashAttention deeply - Calculate real hardware limits - Build a simple Mixture-of-Experts - Study sparse vs dense trade-offs - Experiment with state-space models - Try diffusion-based language models - Design solid data pipelines - Generate and use synthetic data - Verify scaling laws yourself - Go through SFT to DPO to RLHF to GRPO - Apply proper quantization - Set up serving infrastructure - Build reliable evaluation systems - Create RAG pipelines - Add tool use and agent loops - Connect vision adapters - Practice interpretability techniques - Run serious red-teaming - Ship one complete end-to-end system Finish even half of these and you’ll be ahead of 95% of people talking about LLMs.

Why did a spend management platform start its own AI research lab? @karimatiyeh describes @RampLabs as a “collection of crazy ambitious people that don’t like to be told that something is impossible”. It started about a year ago and has since worked on projects like building its own production-grounded coding benchmark “Ramp SWE-Bench”, putting Claude Code in RollerCoaster Tycoon, and a mechanistic interpretability playground. The Redpoint team spent the day with Ramp Labs to answer questions like: ▪️ How did Ramp Labs get started? ▪️ How do they decide what research to pursue? ▪️ The long-term vision for Ramp Labs

SITUATION EXPLAINED: Many in the AI safety community are concerned about the risks posed by open source and open weight models. But they also have the potential to diffuse one very important class of risks: loss of control. @CFGeek, Member of the Policy Staff at @METR_Evals explains: "I think there's this really interesting thing happening with open source or open weight models, and I think there's an opportunity for open weight models to be a very positive force for certain threat models and safety." "At METR, we mostly focus on the loss of control threat model, wherein AI systems could undermine the safeguards that we have in place, particularly because they have misalignment." "Open weight models, if you had models that were developed openly, interpretability techniques could be applied by a wide range of researchers. One of the risks you wanna guard against is the risk of the developers themselves hiding evidence that things are going wrong." "There's this missed opportunity right now where the safety world and the open weights world are not talking to each other, and are totally at odds. I think in the long run that's gonna be bad."