Deep Learning Weekly: Issue 463
Introducing Grok 4.5, How Internal Optimizations Led to Comet Cost Intelligence, a paper on AlayaWorld: Long-Horizon and Playable Video World Generation, and many more!
This week in deep learning, we bring you Introducing Grok 4.5, How Internal Optimizations Led to Comet Cost Intelligence, and a paper on AlayaWorld: Long-Horizon and Playable Video World Generation.
You may also enjoy, Introducing GPT-Live, Intelligence is Free, Now What? Data Systems for, of, and by Agents, a paper on Can LLMs Reason Structurally? Benchmarking via the lens of Data Structures, and more!
As always, happy reading and hacking. If you have something you think should be in next week’s issue, find us on Twitter: @dl_weekly.
Until next week!
Industry
xAI launches Grok 4.5, its fastest and most token-efficient model yet (80 TPS, 4.2× fewer output tokens than Opus 4.8 on SWE-Bench Pro), priced at $2/$6 per million tokens.
OpenAI launches GPT-Live, a full-duplex voice model that listens and speaks simultaneously and delegates reasoning/search tasks to GPT-5.5 in the background, now powering ChatGPT Voice.
Introducing Mods: Enabling Agents to Self-Improve through Harness-Level Adaptation
Letta launches Mods for Letta Code, letting agents self-modify their own harness code rather than only learning through context, with a built-in learn/generate-env workflow.
Leanstral 1.5: Proof Abundance for All
Mistral open-sources Leanstral 1.5, which saturates miniF2F, solves 587/672 PutnamBench problems, and sets new SOTA on FATE-H/X while surfacing 5 previously unknown bugs across 57 tested repositories.
MLOps/LLMOps/AgentOps
Intelligence is Free, Now What? Data Systems for, of, and by Agents
UC Berkeley researchers argue near-zero inference costs will force data systems to be redesigned around agentic workloads — serving agent query swarms, sustaining multi-agent memory/coordination, and letting agents synthesize disposable custom systems.
NVIDIA argues agentic AI’s bottleneck is open, inspectable synthetic data rather than model weights, releasing Nemotron’s over 10 trillion pre-training tokens and millions of post-training samples alongside a Prompt Atlas and locally-grounded persona datasets covering more than 2.4B people across ten countries.
Learning
Engineering Insights: How Internal Optimizations Led to Comet Cost Intelligence
A look into how Comet’s engineering team built Cost Intelligence to expose where AI coding agents waste tokens and budget, turning opaque AI spend into actionable engineering metrics.
Harness Engineering for Self-Improvement
Lilian Weng’s post argues the “harness” wrapping a model is becoming as important as the model itself — and is now being optimized by the agents it contains.
Anthropic publishes a history of Claude Code’s origins, tracing its evolution from a 2022 internal coding assistant through “clide” to its February 2025 launch and shift to near-100% AI-written code by winter 2025.
Separating signal from noise in coding evaluations
OpenAI’s audit finds ~30% of SWE-Bench Pro tasks are broken, retracting its own earlier recommendation to adopt the benchmark as a replacement for the flawed SWE-bench Verified.
Modular Pretraining Enables Access Control
Anthropic and AE Studio introduce GRAM, a technique that isolates dual-use knowledge (virology, cybersecurity, nuclear physics) into switchable modules, letting one pretrained model approximate five separately data-filtered models at a fifth of the training compute.
Libraries & Code
An open-source AI observability tool used to debug, evaluate, and monitor LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards.
A context harness for AI agents: all your scattered context — code, memory, docs, databases, SaaS — in one searchable, browsable, file-like interface.
Papers & Publications
AlayaWorld: Long-Horizon and Playable Video World Generation
Abstract:
Game worlds have traditionally been built through labor-intensive production pipelines, making them costly to develop, difficult to customization, and expensive to modify after deployment. Recent advances in video world models offer a fundamentally different paradigm. Rather than explicitly authoring every component of a virtual environment, these models autoregressively synthesize future observations conditioned on the current world state and user interactions, enabling playable worlds to be generated online. Trained on both gameplay recordings and real-world videos, they can capture diverse visual appearances and physical dynamics, opening new opportunities for interactive applications beyond gaming, including embodied intelligence. In this paper, we present \textbf{AlayaWorld}, a full-stack open-source framework for building interactive generative worlds. AlayaWorld enables open-ended real-time interaction, allowing users to freely navigate and perform diverse actions such as combat, spell casting, and monster summoning. The framework unifies the complete development-from data preparation model architecture, model training, inference acceleration, and deployment-within a modular and extensible architecture. Alongside the framework, we release reproducible pipelines, reference implementations, evaluation tools, and comprehensive documentation, establishing a practical foundation for future research and real-time applications of generative world models.
Can LLMs Reason Structurally? Benchmarking via the lens of Data Structures
Abstract:
Large language models (LLMs) are deployed on increasingly complex tasks that require multistep decision-making. Understanding their algorithmic reasoning abilities is therefore crucial. However, we lack a diagnostic benchmark for evaluating these capabilities. We propose to use data structures as a principled lens: as fundamental building blocks of algorithms, they naturally probe structural reasoning—the ability to understand and manipulate relationships such as order, hierarchy, and connectivity that underpin algorithmic reasoning. We introduce DSR-Bench (Data Structure Reasoning Benchmark), spanning 20 data structures, 35 operations, and 4,140 problem instances. DSR-Bench features hierarchical task organization, fully automated generation and evaluation, and fine-grained diagnostics. Evaluating 13 state-of-the-art LLMs reveals critical limitations: the top-performing model achieves only 0.46/1 on challenging instances. Three auxiliary probes targeting more realistic usages expose further weaknesses: models perform poorly on spatial data and context-rich scenarios, and they struggle to reason over their own code.
 | A guest post by
|