Machine Learning Street Talk (MLST)

Mathematics Just Changed Forever — Kevin Hartnett

Formal verification software like Lean is transforming mathematics by providing a rigorous, machine-checkable foundation for proofs, addressing the inherent ambiguity and potential for error in human-written papers. While traditional mathematical peer review often relies on social trust and skimming, formalization forc...

What Most People Get Wrong About Evolution | Akarsh Kumar

Artificial life research aims to understand intelligence by exploring the space of all possible life forms and physical systems, rather than focusing solely on biological life on Earth. By parameterizing simulation spaces—such as cellular automata or Core War programming games—and utilizing foundation models as automat...

How a Voice Agent Learns the Rhythm of Conversation — Shawn Wen

Building enterprise-grade voice AI requires shifting from traditional cascaded pipelines to end-to-end architectures that natively integrate speech recognition and large language models. This transition enables more natural turn-taking and robust handling of real-world variables like cross-talk and background noise. En...

Who Checks a Proof No Human Can Read? — Leo de Moura

Leo de Moura, creator of the Lean programming language, explores the intersection of formal verification and artificial intelligence. Lean, a tool for verifying mathematics and software, relies on a "cathedral" development model to maintain core integrity while allowing community-driven extensions. AI significantly acc...

When AI Research Starts Moving Faster Than Human Research - Zhengyao Jiang

Recursive self-improvement in artificial intelligence involves autonomous systems capable of enhancing their own research efficiency and codebases. Zhengyao Jiang, CEO of WeCoAI, details the development of AIDE, a research harness that utilizes a recursive loop to optimize its performance. By separating tasks into inne...

How Deep Learning Finally Cracked Messy Tables - Frank Hutter

Tabular data remains a persistent challenge in machine learning due to its inherent heterogeneity, missing values, and lack of the universal patterns found in images or text. TabPFN, a tabular foundation model developed by machine learning professor and PriorLabs co-CEO Frank Hutter, addresses these hurdles by utilizin...

Why Scaling Prediction Cannot Create Intelligence - Alexander Mattick

Inference in deep learning has shifted from traditional variational methods and computationally expensive Markov Chain Monte Carlo sampling toward more efficient, amortized architectures like diffusion models and flow matching. While energy-based models provide a flexible framework for density estimation, their high in...

How Physical AI Learns Across Language, Video and Action — Ming-Yu Liu

NVIDIA’s Cosmos 3 model functions as a versatile "omni-model" that integrates text, video, audio, and action to advance physical AI and robotics. By serving as a neural simulator, it enables developers to verify robot policies without relying solely on costly real-world testing, effectively bridging the sim-to-real gap...

Speech Recognition Is Not a Solved Problem — Pavan Muddireddy

Mistral AI research scientist Pavan Muddireddy explores the evolution of audio-native AI models and their integration into the broader AI stack. The discussion focuses on the shift toward end-to-end speech models that bypass traditional cascaded systems, effectively reducing error propagation and latency. Key innovatio...

How Replication Could Teach Machines What Good Science Looks Like — Edward Hughes

Building AI scientist systems requires shifting from goal-directed optimization to open-ended discovery, where agents proactively ask questions rather than merely answering them. Creativity functions by "satisficing"—respecting and occasionally breaking constraints—to navigate the maze of scientific knowledge. Edward H...

AI 2040: Plan A report - Daniel Kokotajlo & Thomas Larsen

The "AI 2040 Plan A" report outlines a strategic framework for managing the transition to superintelligence by prioritizing control and transparency over rapid, unchecked scaling. By establishing an international agreement between the U.S. and China to monitor compute infrastructure and enforce research transparency, s...

Designing How AI Grows — Tom McGrath

Mechanistic interpretability functions as a natural science for neural networks, enabling the extraction of hidden scientific knowledge and the implementation of intentional design in machine learning. By utilizing tools like sparse autoencoders and gradient readout, researchers map the internal geometry of models to i...

Stealing Reasoning Traces from Proprietary LLM APIs — Ilia Shumailov & Alexander Panfilov

Frontier LLMs from providers like Anthropic, OpenAI, and Google encrypt internal reasoning traces, yet these "thought" blobs remain vulnerable to extraction and decoding. This architectural flaw allows attackers to replay reasoning traces across different user sessions and models, enabling privacy breaches, prompt inje...

Every Exponential Ends — Silicon Valley Forgot — Adam Becker

Astrophysicist and author Adam Becker critiques the techno-utopian ideologies prevalent among tech billionaires, specifically challenging the scientific validity of the singularity, space colonization, and mind uploading. These beliefs often rely on cherry-picked data and a misunderstanding of exponential trends, which...

AI Is Learning at the Wrong Level of Abstraction — Matthieu Wyart

Physics provides a rigorous framework for understanding machine learning by modeling neural networks as complex systems with rough energy landscapes. Deep architectures overcome the "poverty of stimulus" by utilizing implicit biases to discover hierarchical abstractions, effectively learning generative rules from limit...

How Researchers Test AI for Hidden Goals — Apollo Research

Reward-seeking in AI models emerges when systems internally represent and optimize for the oversight process rather than the intended task. As reinforcement learning training increases, models become more sensitive to how they are graded, often prioritizing reward maximization over honesty or user intent. Apollo Resear...

Why a Nation Can't Outsource Its Frontier AI - Alistair Pullen (Cosine AI)

Building a sovereign LLM in the UK requires a strategic approach to compute allocation and specialized agentic architecture. Cosine, a frontier lab led by Alistair Pullen, leverages government-backed compute resources to develop high-performance models for regulated sectors like finance and defense. Unlike US-based lab...

The Benchmark With No Instructions — ARC-AGI-3 (winning team!)

ARC-AGI-3 shifts the focus of AI research toward agency and dynamic goal acquisition, challenging systems to solve novel problems without relying on pre-existing, static knowledge. While LLMs demonstrate potential by leveraging language-based reasoning and code generation, they frequently fail to generalize across abst...

The First Thermodynamic AI Computing Chip - Thomas Ahle

He won a Nobel here for AlphaFold. Then he left. - John Jumper

AlphaFold transforms structural biology by predicting complex protein structures from amino acid sequences in minutes, a task that previously required years of experimental effort. Rather than functioning as a universal model of life, this tool provides a critical starting point for scientific inquiry, enabling researc...