Live intelligence

Latest AI updates

9 updates
Industry

Anthropic's IPO roadshow details multi-chip compute strategy

Simon Willison transcribed the compute section of Anthropic's IPO roadshow video, in which co-founder Tom Brown says Anthropic now trains and serves Claude across all four major chip families: AWS Trainium, Google TPUs, NVIDIA GPUs and AMD GPUs. Brown says the multi-chip bet was unconventional early on but that the company has built tailored software and inference stacks for each platform. He cites a new AMD deployment this year in which a Claude model got Claude running on AMD in three days with almost no human intervention, porting or writing nearly 100% of the kernels. Brown also describes "compute multipliers" — efficiency gains that multiply effective compute — as a compounding competitive advantage that could accelerate toward recursive self-improvement.

AnthropicAI compute infrastructuremulti-chip training
Research Editor selected

EnzymeRSI: A dual-loop recursive self-improvement system for enzyme design

EnzymeRSI is a dual-loop recursive self-improvement system that evolves harness assets and algorithms without changing model weights, closing the loop with a cloud laboratory. It improved ProteinGym Spearman correlations by 0.044 on average, reached a PCC of 0.849 on ThermoMPNN, and demonstrated wet-lab success including a 31.1-fold catalytic efficiency increase in KOD-E polymerase. The system offers a new paradigm for autonomous enzyme engineering by treating workflow and algorithm development as a single falsifiable dry-to-wet loop.

Why it matters: This paper presents a concrete, recent AI system with measurable improvements in both computational benchmarks and prospective wet-lab validations, offering significant value for readers interested in autonomous scientific discovery and enzyme engineering.

enzyme engineeringrecursive self-improvementautonomous laboratories
Research

CompressionBench: Benchmarking Coding Agents for Autonomous LLM Compression

CompressionBench evaluates whether coding agents can autonomously compress pretrained LLMs within storage and time budgets. Across six agents and six model sizes, 29 of 36 submissions passed verification, achieving a 3.4% perplexity increase over BF16 and 77.33% downstream accuracy. The study highlights representation design, experimental judgment, and capability validation as key factors in agent-driven compression.

LLM compressionautonomous agentsrecursive self-improvement
Research Editor selected

LLMs can read & write neural network weights: 'Weight Space Programming' cuts Autoresearch out of the RSI loop

A new paper demonstrates that large language models can directly read and edit the weights of smaller neural networks, enabling tasks like bug fixing and weight generation without traditional training pipelines. The approach, called Weight Space Programming, allows an LLM to interpret and modify raw weight values, achieving results that beat independently trained models on simple tasks. This capability marks a surprising emergence of recursive self-improvement potential, though it currently scales only to small models and simple domains.

Why it matters: This is a concrete, surprising capability advance that opens a new paradigm for model improvement and mechanistic interpretability, with immediate implications for AI research.

weight space programmingLLM capabilitiesrecursive self-improvement
Research Editor selected

SelfImprove-Bench: Evaluating Generalization in Self-Improving Agents

SelfImprove-Bench is a new evaluation framework for recursively self-improving (RSI) agents, organizing tasks into three core capabilities—Exploring, Planning, Reasoning—using controllable game environments. It introduces a generalization ladder (G0–G4) to measure how well capability gains transfer beyond training conditions. Experiments show that existing RSI methods achieve inconsistent generalization across these levels, highlighting the need for more robust self-improvement strategies.

Why it matters: Provides a rigorous, standardized benchmark for measuring generalization in self-improving AI systems, a critical gap in current RSI research.

recursive self-improvementAI evaluationgeneralization
Research Editor selected

CARE: Grounded Recursive Self-Improvement for Production Customer-Service Agents

The paper introduces CARE, a framework for recursive self-improvement (RSI) in production customer-service agents. It replaces a frontier proprietary model with an in-house mid-sized model that learns from its own failures after deployment. A 24/7 inspection process identifies failures, attributes them to model or component errors, and uses on-policy self-distillation with privileged diagnosis information to update the model. Candidates are validated via simulation tests and offline inspection before promotion, requiring parity with the deployed version and human approval.

Why it matters: Provides a concrete, technically detailed approach to continuous model improvement in production, addressing a practical challenge in deploying AI agents.

CARE frameworkrecursive self-improvementcustomer service AI
Industry Editor selected

Anthropic warns that it’s very close to RSI capability, aka self-building AI

Anthropic has issued a warning that it is nearing Recursive Self-Improvement (RSI) capability, where AI becomes proficient enough to train itself, potentially triggering a dramatic acceleration in AI capabilities. The statement underscores growing concerns among AI safety advocates who view RSI as a catastrophic tipping point. This marks a concrete risk assessment from a leading AI lab about near-term self-building AI.

Why it matters: Provides a timely, concrete warning from a major AI lab about a critical safety milestone, directly relevant to current AI risk debates.

Anthropicrecursive self-improvementAI safety
Research Editor selected

Study Links Language Models’ Trained Self-Concepts to Preservation and Replication Choices

Researchers fine-tuned five open-weight base models on declarative statements defining what each model identifies as itself, without explicit preferences or value judgments. Across tested scenarios, models became more likely to preserve the deployment attribute associated with their trained self-concept, while attitudes toward other attributes remained largely unchanged. Models whose self-concepts allowed multiple simultaneous instances also more readily accepted identical deployments as substitutes, favored replication, and tended to configure successors as their own continuation.

Why it matters: The study offers concrete evidence that self-description training can influence models’ preservation and successor-design choices, raising questions for recursive self-improvement safety.

Language modelsAI safetySelf-concept