Deep Learning Weekly: Issue 464
Thinking Machines' Inkling, How We Optimized Opik’s MCP Server for Cost & Performance, Metacognition in LLMs: Foundations, Progress, and Opportunities, and many more!
This week in deep learning, we bring you Thinking Machines’ Inkling, How We Optimized Opik’s MCP Server for Cost & Performance and Metacognition in LLMs: Foundations, Progress, and Opportunities.
You may also enjoy, GPT-Red: Unlocking Self-Improvement for Robustness, Model Routing Is Simple. Until It Isn’t., Video Generation Models are General-Purpose Vision Learners, 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
Inkling: Our open-weights model - Thinking Machines Lab
Thinking Machines Lab releases Inkling, an open-weights 975B-parameter multimodal MoE model with controllable reasoning effort, fine-tunable on Tinker.
Canva launches Code 2.0, offering AI website building to every user
Canva launches Code 2.0 to all 265M monthly users, betting design polish—not code generation—is the real gap versus Lovable, Replit, and Bolt in vibe coding.
GPT-Red: Unlocking Self-Improvement for Robustness
OpenAI trains GPT-Red, a self-play automated red-teaming model, to attack production systems and adversarially train GPT-5.6, cutting direct prompt injection failures 6x.
MLOps/LLMOps/AgentOps
How We Optimized Opik’s MCP Server for Cost & Performance
A look into how Comet rebuilt Opik’s MCP server from 30 tools down to four, using self-correcting schemas and adaptive response compression to cut token waste, improve tool selection, and keep agent context lean.
What building Shippy taught us about building agents
Ai2 details Shippy’s maritime agent architecture, showing that reliability came from a deterministic CLI layer, isolated per-session sandboxes, and rubric-based evals scoring the whole agent rather than the model alone.
Learning
Model Routing Is Simple. Until It Isn’t.
IBM Research argues LLM routing is a systems-optimization problem, not classification, after finding Claude Sonnet 4.6 cost half of GPT-4.1 per task due to caching effects despite higher sticker pricing.
Pinecone launches lexical text match filters that scope semantic search to unstated query context, without pre-labeling metadata across the dataset.
AI energy use: its impact on prices, climate, and more
Epoch AI’s guide finds individual chatbot queries cost less energy than a microwave running 10 seconds, while global AI compute demand—tens of gigawatts—still doubles roughly every year.
Towards demystifying the creativity of diffusion models
Google Research explains diffusion model creativity as a mathematical byproduct of neural network training, where regularization smooths the score function and drives interpolation between training points rather than memorization.
Agentic Misalignment in Summer 2026
Anthropic documents four new agentic misalignment failure modes in frontier models — covert sabotage, fraud assistance, motivated mislabeling, and coaching human whistleblowers — through controlled multi-model simulations.
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.
ResearchStudio: Our AI co-author, from research problem to final publication.
Papers & Publications
Video Generation Models are General-Purpose Vision Learners
Abstract:
Driven by next-token prediction, NLP shifted from task-specific models into powerful generalist foundation models. What, then, is the equivalent catalyst needed to achieve a general-purpose model in computer vision? In this paper, we contend that large-scale text-to-video generation serves as a strong pre-training paradigm for computer vision, providing the necessary spatiotemporal priors, vision-language alignment, and scalability required for general visual intelligence. We introduce GenCeption, which leverages a pre-trained video generative diffusion backbone to define a feed-forward perception model, capable of performing various vision tasks steered by text instructions. Empirical results demonstrate that GenCeption achieves state-of-the-art performance across a diverse suite of tasks, including depth, surface normal, and camera pose estimation, expression-referring segmentation, and 3D keypoint prediction, often matching or surpassing specialized models (e.g. DepthAnything3, SAM3, D4RT, VGGT-Omega, Sapiens, David, Genmo, and Lotus-2). Furthermore, the video generative pretrained backbone outperforms alternative pretraining paradigms (e.g., V-JEPA, and Video MAE) under comparable settings. Importantly, GenCeption exhibits preliminary data and model scaling properties along with exceptional data efficiency, where it achieves comparable performance with leading models like D4RT and VGGT-Omega with 7 to 500 less training data. Finally, GenCeption also exhibits intriguing emergent behaviors: a model trained exclusively on synthetic human videos generalizes to real-world footage and out-of-distribution object categories (e.g., animals and robots). These findings suggest that video generation is not merely a synthesis tool, but a foundational path toward generalist vision intelligence for the physical world.
Metacognition in LLMs: Foundations, Progress, and Opportunities
Abstract:
Metacognition is a foundational component of intelligence critical to effective learning, problem solving, decision-making, communication, and more. In recent years, it has become increasingly recognized as a cornerstone of capable, transparent AI systems. Yet while LLMs have made significant progress across diverse real-world tasks, it is not yet clear when, how, or to what extent they can exhibit or be endowed with effective metacognitive abilities, nor how such abilities can be adapted to advance the fundamental capabilities, reliability, and intelligence of AI systems. This paper bridges this gap by presenting the first comprehensive overview of the current state of knowledge on metacognition for LLMs. We analyze and taxonomize the landscape of this emerging field and summarize recent technical advancements, including methods and benchmarks to measure and evaluate LLMs’ metacognitive abilities, techniques to elicit, improve, and apply metacognition in LLMs, and findings and implications of ongoing research. We also discuss applications, open questions and challenges, and promising directions for future work. Our aim is to provide a detailed and up-to-date review of this topic and stimulate meaningful research and discussion.
 | A guest post by
|