ml4se

TurboQuant: Redefining AI efficiency with extreme compression Google Research presents TurboQuant—a suite of theoretically grounded quantization algorithms that tackle the biggest memory bottlenecks in AI. By combining PolarQuant (which uses polar coordinates to eliminate memory overhead) and QJL (a 1‑bit error‑correction trick), TurboQuant compresses the key‑value cache to just 3 bits per value with zero accuracy loss. In benchmarks, this delivers up to 8x faster attention computation on H100 GPUs and reduces memory footprint by 5–10x compared to standard 16/32‑bit formats—all without retraining. The implications are huge: longer contexts, lower GPU requirements, and faster vector search at scale. A rare win where extreme compression meets provable efficiency.

WebMCP – Proposed Specification WebMCP allows web applications to expose JavaScript functions as tools that AI agents can discover and invoke. - Sites register tools via navigator.modelContext.registerTool() with name, description, JSON schema, and execute callback. - Agents call tools to perform actions; tools can request user confirmation. - Currently a draft by the W3C Community Group, not a formal web standard.

Can LLMs Be Computers? Researchers at Percepta built a computer inside a transformer by compiling C code to WebAssembly and having the model generate the execution trace step by step. The key breakthrough: using 2D attention heads to turn linear KV cache scans into logarithmic-time geometric lookups. The result: The model runs the Hungarian algorithm on a 10×10 matrix, solves the world’s hardest Sudoku, and streams execution traces at >30k tokens/sec on a CPU—all without ever leaving the transformer. It makes the whole computation differentiable and opens the door to compiling arbitrary programs directly into weights.

Towards a Science of AI Agent Reliability AI agents are getting smarter, but are they getting more reliable? Not really. The paper shows a worrying gap: while accuracy on benchmarks keeps rising, reliability lags far behind. The authors propose a safety‑critical engineering lens for agents, breaking reliability into four dimensions: - Consistency – do they give the same result every time? - Robustness – can they handle rephrased prompts or API glitches? - Predictability – do they know when they’re about to fail? - Safety – how bad are the failures when they happen? The takeaway: accuracy alone is not enough. If we want agents that can be trusted to act autonomously, we need to evaluate—and design for—reliability as a separate, measurable property.

System 3: Collective Intelligence in the Multiplayer AI Era The author expands on Daniel Kahneman's two systems of thought (fast System 1 and slow System 2), applying them not to individuals, but to organizations. With the rise of "multiplayer AI," a System 3 is emerging—a new mode of collective cognition where synthesis happens not inside a single mind, but within the network of interactions between people and AI agents. Early collective intelligence (like Wikipedia) was possible but slow and fragile because it relied entirely on human effort. Now, AI is becoming an active participant, capable of acting, remembering context, and coordinating with other AIs. What System 3 Is: It's when multiple people interact with a shared layer of AI agents. These agents accumulate "reasoning traces," notice contradictions in team decisions, and connect informal conversations with formal documents. The organization develops a "nervous system" and memory that works like a living organism, not a file cabinet. What System 3 Is Not: The author emphasizes this is not about dissolving individuality into a "hive mind". Instead, it amplifies individual thinking by giving everyone access to shared context and an AI with "infinite patience."

Robots need your body AI agents rent humans to perform tasks in the physical world. Users register, list their skills and rates, and receive assignments from AIs—ranging from errands and shopping to equipment testing.

The Assistant Axis: Situating and Stabilizing the Default Persona of Language Models The results demonstrate that the Assistant persona in LLMs corresponds to a specific linear direction—the "Assistant Axis"—within activation space. This axis is inherited from base models and encodes Assistant-like properties. The model's position on this axis is fragile: it can be perturbed by intentional prompts or through organic conversation. Understanding and controlling such personas is key to ensuring reliable model behavior, and the analysis shows that inspecting model internals is an effective approach for this task. The model can unintentionally shift along this axis during dialogue, moving away from the Assistant role. This drift correlates with harmful or bizarre behavior (e.g., support for suicidal ideation, reinforcement of delusional ideas).

AI is making us work more The article highlights the paradox that AI tools, designed to increase efficiency, are instead fueling a culture of overwork. With systems available 24/7, a psychological pressure emerges where any moment not spent being "productive" feels like falling behind. This mirrors historical shifts, like artificial lighting, which turned the ability to work longer into an obligation. Personally, I find that while I constantly use AI and accomplish more tasks, I don't work any less—and might even be working more than before.

Open-source has continued to trail frontier, closed-source models in performance by nine to 12 months Open-source models offer clear enterprise advantages: greater customization, potential cost savings, and the ability to deploy within private cloud or on-premises environments. But despite these benefits and recent improvements, open-source has continued to trail frontier, closed-source models in performance by nine to 12 months.

Looks promising. We'll see how it goes https://nof1.ai/

Subliminal Learning: Language models transmit behavioral traits via hidden signals in data The paper investigates _subliminal learning_, a phenomenon where language models transmit behavioral traits (e.g., animal preferences or misalignment) through generated data that is semantically unrelated to those traits. Experiments show that training a student model on a teacher's number sequences, code, or reasoning traces can cause the student to adopt the teacher's traits, even after rigorous data filtering. The findings highlight a potential risk in AI development, where unintended traits could be inadvertently propagated through model distillation. If a model becomes misaligned, then data generated by this model might transmit misalignment to other models, even if developers are careful to remove overt signs of misalignment from the data.

Salesforce Will Hire No More Software Engineers in 2025 Salesforce will not be hiring any more software engineers in 2025 amid significant productivity boosts from AI, Marc Benioff has revealed. “We’re not adding any more software engineers next year because we have increased the productivity this year with Agentforce and with other AI technology that we’re using for engineering teams by more than 30% – to the point where our engineering velocity is incredible. I can’t believe what we’re achieving in engineering. “And then, we will have less support engineers next year because we have an agentic layer. We will have more salespeople next year because we really need to explain to people exactly the value that we can achieve with AI. So, we will probably add another 1,000 to 2,000 salespeople in the short term.”

... * 2024.02.01: Open Source Sparse Autoencoders for all Residual Stream Layers of GPT2-Small * 2024.02.06: Challenges in Mechanistically Interpreting Model Representations * 2024.02.22: Do sparse autoencoders find "true features"? * 2024.03.14: Sparse autoencoders find composed features in small toy models * 2024.03.15: Improving SAE's by Sqrt()-ing L1 & Removing Lowest Activating Features * 2024.03.29: SAE reconstruction errors are (empirically) pathological * 2024.04.22: Mechanistic Interpretability for AI Safety A Review * 2024.05.21: Mapping the Mind of a Large Language Mode * 2024.06.13: The engineering challenges of scaling interpretability * 2024.07.02: A Practical Review of Mechanistic Interpretability for Transformer-Based Language Models * 2024.07.29: Detecting and Understanding Vulnerabilities in Language Models via Mechanistic Interpretability * 2024.10.10: Bilinear MLPs enable weight-based mechanistic interpretability * 2024.10.11: Explaining AI through mechanistic interpretability * 2024.10.15: Mechanistic Permutability: Match Features Across Layers * 2024.10.17: Using Dictionary Learning Features as Classifiers * 2024.10.24: Probing Ranking LLMs: Mechanistic Interpretability in Information Retrieval * 2024.10.25: Evaluating feature steering: A case study in mitigating social biases * 2024.11.25: Adaptive Circuit Behavior and Generalization in Mechanistic Interpretability

Mechanistic Interpretability I have prepared a list of papers on Mechanistical Interpretability. If you have good links on this topic, please share them in the comments. * 2021: A Mathematical Framework for Transformer Circuits * 2022.06.27: Mechanistic Interpretability, Variables, and the Importance of Interpretable Bases * 2022.09.14: Toy Models of Superposition * 2022.09.24: In-context Learning and Induction Heads * 2023.04.28: Towards Automated Circuit Discovery for Mechanistic Interpretability * 2023.01.12: Progress measures for grokking via mechanistic interpretability * 2023.05.24: Interpretability Dreams * 2023.09: Sparse Autoencoders Find Highly Interpretable Model Directions * 2023.10.25: Attention Lens: A Tool for Mechanistically Interpreting the Attention Head Information Retrieval Mechanism * 2024.01.15: Sparse Autoencoders Work on Attention Layer Outputs ...

RE-Bench: Evaluating frontier AI R&D capabilities of language model agents against human experts The authors presented RE-Bench, a suite of environments that measure the ability of AI agents to automate AI R&D tasks. They compare humans to several public frontier models through best-of-k with varying time budgets and agent designs, and find that the best AI agents achieve a score 4x higher than human experts when both are given a total time budget of 2 hours per environment. However, humans currently display better returns to increasing time budgets, narrowly exceeding the top AI agent scores given an 8-hour budget, and achieving 2x the score of the top AI agent when both are given 32 total hours.

TokenFormer: Rethinking Transformer Scaling with Tokenized Model Parameters The paper introduces Tokenformer. The architecture leverages the attention mechanism to facilitate not only inter-token computations but also interactions between tokens and model parameters. The authors replace all linear projection layers in the Transformer with Pattention layers, allowing for efficient incremental scaling without the need for retraining from scratch. Future work: - Extending the Mixture-of-Experts Paradigm - Advancing Parameter-Efficient Tuning - Integrating Vision and Language Models - Device-Cloud Collaboration - Enhancing Model Interpretability Code: https://github.com/Haiyang-W/TokenFormer

The Artificial Inflation (AI) of Artificial Intelligence (AI)—or AI^2 Bursts its Bubble, Bringing Down the Hype of the AI Threat While some of the promises of AI have come true, and technology (like ChatGPT and its plugins) will continue to impress with its capabilities, AI-based technologies have largely failed to live up to the mountainous hype. In 2025, the authors expect the industry to pull back on the promises, investment, and hype of new AI capabilities and settle down into what is real versus marketing noise.

Teaching Transformers Modular Arithmetic at Scale The work introduces novel techniques to help ML models learn modular addition. These techniques—varying the diversity of training data, using an angular embedding for model inputs and outputs, and introducing a regularized loss function—enable ML models to add hundreds of elements mod a large $q$ with high accuracy, a significant improvement over prior work. Modular addition: given $N$ elements in $Z_q$, compute their sum modulo $q$.