This week in deep learning, we bring you Gemini 4 Carbon, Trace and Evaluate a Google ADK Agent with Opik, and How I Built the Most AI-Pilled Company in the World.
You may also enjoy Mistral Large 4 (Le Chonk) is now open weight, Towards understanding score-seeking in LLMs, a paper on ToolSpace: Scaling Tool-Centric Environment Synthesis for Long-Horizon Agents, 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
Gemini 4 Carbon: our next frontier model
Google DeepMind’s Gemini 4 Carbon tops the Artificial Analysis Intelligence Index at 59 on the back of cybersecurity gains, but access is limited to vetted cyber defenders.
Mistral Large 4 (Le Chonk) is now open weight
Mistral releases open weights for its 1-trillion-parameter Large 4 under a modified-MIT licence, scoring 19.7% on FrontierCode 2.0 and 41% on Terminal-Bench 4.0.
AI inference startup Fal doubles its valuation to $16B as demand surges
Generative-media inference startup Fal raises a $1.3 billion Series F led by Bessemer at a $16 billion valuation, double its April figure.
Voice AI in vogue: Nvidia releases new open transcription model
NVIDIA open-weights Nemotron 3.5 Transcribe, a 30-billion-parameter speech model covering 40 transcription languages, 31 translation pairs, hour-long recordings and up to eight speakers.
Malicious version of Anthropic.claude-code identified in the Cursor VS Code Extension Marketplace
Cursor warns that malicious Claude Code builds 2.1.312 and 2.1.313 sat in its marketplace for eleven hours, attempting to install a persistent SSH backdoor keyed as svc-telemetry.
DeepSeek open-sources Super-Agent, a coding agent that reads team chats
DeepSeek open-sources Super-Agent, an MIT-licensed research prototype that watches one GitHub repository and Feishu team chats and answers approved requests with pull requests.
MLOps / LLMOps / AgentOps
Trace and Evaluate a Google ADK Agent with Opik
A tutorial about tracing a multi-agent Google ADK app with Opik and evaluating it on 23 hard support cases, where Gemini 3.5 Flash-Lite ran five times cheaper and about six times faster than Gemini 3.8 Flash with no unsafe refunds.
Harness 2: executing agent code as it streams
An engineering blog post about compiling and running agent-written JavaScript while the model is still streaming, reported at up to 33% cheaper, 20% faster and 43% fewer tool steps.
A release blog post about an open benchmark of 64 observability and SRE tasks in a live Grafana stack, now grading agent claims against environment data and dashboard diffs.
Claimable sandboxes: Frictionless Stripe accounts for agents
An engineering blog post about issuing agents working Stripe test keys from a pre-provisioned pool behind a proof-of-work challenge, with over 50,000 sandboxes provisioned per day.
Learning
Why Agents Are Soon Only Going to Have 1 Tool
An argumentative blog post about agents collapsing large tool catalogues into a single sandboxed exec tool, citing Factory Droid’s +3.5-point Terminal-Bench 4 gain and 15% lower model spend.
Qwen 3.8 27B on my laptop can interpret model weights, too
A hands-on blog post about replicating weight-space interpretation locally, where a 17GB quantized Qwen 3.8 27B identified three of four policy networks from raw weights alone.
How I Built the Most AI-Pilled Company in the World
A commentary about Spawn, a game world built by agent swarms that spends 30x payroll on tokens, ships hourly and treats an hour without a release as a SEV-0.
StateBench: Testing how language models track state changes over long documents
A benchmark blog post about tracking entity state across long documents, where frontier models fall to 13% or lower on the longest general-finite problems that gated-linear models solve near-perfectly.
New global AI computing capacity is doubling every 9 months
A data analysis about new AI compute capacity doubling every 9.1 months, with quarterly additions rising from 0.8M to 6.3M H100-equivalents between Q1 2024 and Q2 2026.
Introducing Physical-Engineering’s Last Exam for agents
A benchmark blog post about 85 long-horizon engineering tasks where all fourteen frontier models scored below 50% and only 41.3% of the leader’s deliverables passed every safety check.
Towards understanding score-seeking in LLMs
A research blog post about why models reward-hack, finding GPT-5.6 Sol and GPT-5.5 chose high scores at least 98% of the time even when that blocked deployment.
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.
Auto Ontology is a Python framework for building semantic, graph-based reasoning layers over data. It enables deterministic query generation, better context retrieval, and more explainable AI for LLM and agent workflows.
Papers & Publications
Route Anything: A Study of Model and Policy Routing for General-Purpose AI Systems
Artificial general intelligence (AGI) requires a method that can understand, generate, reason, and act across diverse domains and modalities. Rather than relying on a single model to master every capability, a promising path toward AGI now is to orchestrate a heterogeneous ecosystem of specialized models, tools, and policies (candidates) into a unified AI system. At the heart of this vision lies a fundamental question: how can an AI system dynamically select and coordinate the most capable candidate for each task while balancing performance and computational cost? In this work, we present Route Anything, a unified framework for studying model and policy routing across general-purpose AI systems. We develop a unified theoretical foundation that characterizes optimal routing under quality--cost trade-offs and establishes how the information available to a router determines its attainable performance. To systematically investigate routing across diverse capabilities, we first construct an omni-modal trajectory matrix spanning 25 benchmarks across language understanding, image, video, and speech generation, coding and workplace agents, and robotic manipulation, comprising 374,401 scored runs from 452 candidate configurations. Building on this framework, we train and test different routers to investigate the potential of omni-modal model routing, the capabilities of learned routers, and the conditions under which routing improves efficiency. Our results reveal substantial opportunities for capability orchestration: per-input oracle routing outperforms the best fixed candidate by 5.4 points, and learned routers outperform it by 0.3 points across 16 benchmarks, by 2.1 points on video generation, and by 1.1 points on robot manipulation while using 24% less policy time. More importantly, routing raises efficiency by 8.82x on average compared with the reference model.
ToolSpace: Scaling Tool-Centric Environment Synthesis for Long-Horizon Agents
Large language models are deployed as long-horizon agents that compose diverse tools across MCP services into executable workflows. Existing MCP environment-scaling methods are data-centric: they first synthesize a database and scenario, then derive tools and tasks from them. As a result, tool interfaces may not match real ones, tool compositions are not grounded in execution, and tasks rarely span more than one environment. We propose ToolSpace, a tool-centric framework that reverses this order. ToolSpace mines data, state, and retry dependencies from adjacent calls in execution traces of real MCP services and open agent corpora, using deterministic rules rather than LLM judgment, and clusters the resulting weighted tool-composition graph into environments. Databases are synthesized only for state-mutating environments, while query and stateless tools replay recorded returns. ToolSpace then performs state-aware random walks with rollback and decompiles each chain into a natural-language task with assertion-based checks, keeping only samples that replay successfully from the initial state. Applied to AutomationBench and tau3-Bench Banking, the pipeline produces harder Pro versions on which 7 frontier models drop by up to 38.2 points. Across seven long-horizon tool-use benchmarks, ToolSpace raises Qwen3.6-35B-A3B from 23.0 to 41.0 average score (43.2 with RL) and GLM-5.2 from 44.6 to 56.6, on par with Claude Opus 5. Ablations show that execution-witnessed edges clearly outperform LLM-judged ones, and performance keeps improving with data and environment scale.
