The Archive

Neural networks trained by gradient descent on a smooth cost function can nevertheless learn in steps: the cost holds on long plateaus and then drops abruptly. Meanwhile, training losses instead follow smooth power laws. Variants of both behaviors occur in architectures with very different microscopic structures, which is the signature of a few relevant collective variables. We show that a symmetry fixes what those variables are: a network layer is a sum over interchangeable units, so relabeling the units leaves it unchanged; given smoothness and the condition that a unit's gradient vanish at...

LLM-based agents increasingly rely on external memory to support long-horizon reasoning and interaction. However, the main bottleneck is not simply storing past experience, but recovering the right set of evidence when relevant information is distributed across many interactions. Existing approaches struggle with this access problem. Full-context methods require noisy long-context search, flat retrieval often returns isolated and incomplete records, and graph-based memory systems can be expensive to construct while compressing rich event context. We introduce RippleMem, a long-term memory sys...

The tech giant has considered a nine-figure budget for the payments, according to the WSJ.

Today on Decoder, I’m talking with Hayden Field, The Verge’s senior AI reporter, about a question that’s been rocketing around the tech industry for the past week: Is Google losing the AI race? That’s because last week Google announced a bombshell reorganization of its AI division, Google DeepMind. Jeff Dean, the company’s chief scientist, is leaving to form his own startup and DeepMind cofounder and CEO Demis Hassabis is stepping aside to focus on longer-term research. You can read these moves, and Google’s reaction, in a lot of different ways. So I really wanted to sit down with Hayden to d...

Rapid revenue growth fuels hope Claude maker's IPO is the biggest listing in history

The police-tech giant Flock is announcing today that it will change officers’ access to its nationwide network of license plate readers, in an apparent effort to quell a growing backlash and win back contracts lost amid concerns about mass surveillance and police abuse. Several changes aim directly at a problem that has made recent headlines:…

You soon won’t have to look at two Copilot apps in your Windows taskbar. | Image: Microsoft Microsoft is finally beginning to combine its consumer and commercial Copilot AI assistants into a single "super app" interface, starting with the Copilot and Microsoft 365 Copilot apps. Both personal and work accounts will be moved to the new unified app, which recycles the "Microsoft Copilot" name but features an updated app icon. The single app also means there won't be two annoying Copilot icons in the system tray or taskbar anymore. "Starting with a fresh look, the updated app combines the best of...

The mark flags anything Claude processed, even human writing it only edited.

Learn how startups use GPT-5.6 to build faster, more cost-efficient AI agents with smarter model selection and new Responses API capabilities.

Preview Ultrafast, a new OpenAI API service tier that runs GPT-5.6 Sol up to 14× faster. Powered by Cerebras, it delivers up to 750 output tokens per second.

OpenAI appoints Dali Rajic as Chief Revenue Officer to lead its global revenue organization and help businesses realize the full value of AI.

When we set out to talk to kids about artificial intelligence, we thought we knew what we’d hear. We expected some to tell us they were using it to cheat a little, the way Millennials and Gen Xers opened up CliffsNotes or programmed formulas into their TI-82s, and others to share inspiring ways they were…

Release: alchemy-utils 0.1a1 Performance boost for DuckDB exports and CSV imports, see here .

AI teammate category just had its most significant new entrant yet

DeepSeek V4 Pro 0813 (on OpenRouter) The latest DeepSeek Pro model is now available, via API only. I had to link to OpenRouter because DeepSeek don't have any obvious announcement page for their new model. I haven't been able to confirm if they plan to release the open weights, but given the weights are available for both April's deepseek-ai/DeepSeek-V4-Pro and July's deepseek-ai/DeepSeek-V4-Flash-0731 it seems likely. Interestingly I got very different looking pelicans for the three different reasoning levels of low, medium, and high. I've not noticed this kind of difference from any other m...

Is Anthropic's new watermarking system a travesty? Some have taken to social media to complain that it is.

“ShieldFont” aims to poison AI training data without making pages unreadable for people.

The data was scraped and exfiltrated from 2,500 users of a compromised AI package.

Streaming platform says user-generated content "may be used for future Gen AI model improvements."

"If this was opt-in, nobody would opt in," Twitch CPO Mike Minton said on a livestream responding to user feedback. "That's honestly the answer."

Release: alchemy-utils 0.1a0 I've long pondered what a database agnostic version of my sqlite-utils Python library and CLI utility might look like. This morning (literally a shower project) I tasked Codex and GPT-5.6 Sol Ultra with building a prototype: Do a research spike to see what it would take to build a library with the same core API as SQLite-utils - in particular the insert and upsert and insert_all and upsert_all and create and update methods, and the table introspection stuff - but backed by SQLalchemy so it works for multiple database engines Test against PostgreSQL and SQLite and ...

Alibaba released the open weights for Qwen3.8-2.4T-A95B (Qwen3.8-Max), its largest open-weight model, bringing near-frontier capabilities to the open... Alibaba released the open weights for Qwen3.8-2.4T-A95B (Qwen3.8-Max), its largest open-weight model, bringing near-frontier capabilities to the open ecosystem. It has 2.4T total parameters with 95B activated per token. It has 2.4T total parameters with 95B activated per token. It’s a fine-grained mixture of experts (MoE) architecture with a hybrid of full and linear attention, a context window of… Source

Cognition may be looking to raise another mega round just a few months after raising $1 billion at a $26 billion valuation.

Creative agents still lack an effective way to learn from high-quality human films, limiting their ability to produce cinematic-grade videos. A key challenge is the absence of a structured video representation that is both faithful to film content and directly usable for agentic reasoning and manipulation. To address the challenge, we propose the Agentic Video Auto-Encoder (AVA-Encoder), a framework for learning agent-native video representations via agentic auto-encoding. AVA-Encoder transforms a video into a knowledge graph (KG) representation and then reconstructs it back into video. Its h...

Aerial vision-language navigation (VLN) requires an embodied agent to integrate visual evidence over time, plan future actions, and determine when it has reached a navigation goal under partial observability. Although recent VLA models offer a promising perception-to-action paradigm, adapting them to aerial navigation remains challenging due to limited historical context, short planning horizons, and unreliable implicit termination. To address these challenges, we propose DreamFly, a diffusion-based aerial VLN framework built on Dream-VLA. DreamFly introduces a causally aligned historical mem...

Recent work on distillation transfers the capabilities of large models to smaller ones often by updating the latter's parameters, through teacher forcing, on-policy distillation, and related training-time methods. In this paper, we ask whether such transfer can instead occur at test time. We study strong-to-weak scaffolding: whether a stronger builder model can construct inference-time harnesses that help a weaker target model solve tasks more reliably without any parameter updates. Using four representative Theory-of-Mind benchmarks, each builder model uses 5% of the data as a validation set...

Safe offline RL typically assumes access to dense per-step cost annotations, but in practice supervisors provide only trajectory-level stop-feedback: a binary signal at the first unsafe transition, with no per-step attribution. We frame this as a temporal credit assignment problem and propose the Redistribution-based Cost Inference (RCI) framework, which converts sparse stop-feedback into dense per-step costs via return decomposition, then trains a constrained offline policy on the augmented dataset. We show that return-equivalent redistribution preserves the feasible policy set and the optim...

At Ai4, three of the world's most respected AI experts—Geoffrey Hinton, Fei-Fei Li, and Andrew Ng—debated regulation, open-source access, and how America can compete as China advances in Asia.

Dynamic Master Logic (DML) provides a hierarchical framework for representing system behavior by linking functional objectives to underlying structural elements. However, DML construction typically relies on expert interpretation of technical documentation, limiting scalability for complex systems. This study presents a framework for automated construction of DML models from system descriptions and their representation as Knowledge Graphs (KG-DML), using Retrieval-Augmented Generation and Large Language Models as enabling tools. Building on prior work with small-scale systems, the framework e...