This week in deep learning, we bring you Introducing GPT-6.2 Astra and Sol, External sandboxes for coding agents and a paper on Self-Improvement Memory Enables Continual Recursive Self-Improvement.
You may also enjoy Google releases Gemini 4 Hydrogen, Building strong safety evaluations, a paper on SWE-CoWork: Benchmarking Coding Agents as Digital Employees in Organizations, 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
Introducing GPT-6.2 Astra and Sol
OpenAI introduces GPT-6.2 Astra and Sol, with Astra scoring 5.2 points higher than GPT-6 Astra on a professional work benchmark at less than one-third the cost, and Sol nearly matching Astra on coding at about one-fifth the cost.
Google releases Gemini 4 Hydrogen, a lower-cost model said to outscore Argon
Google’s $1.50-per-million-input-token Gemini 4 Hydrogen scores 51 on the Artificial Analysis Intelligence Index against frontier Argon’s 50, and reaches 57% on Terminal-Bench 4.1.
Alibaba launches Qwen4-Max with training, attention optimizations
Alibaba’s Qwen4-Max activates 71 billion of 2.4 trillion backbone parameters and uses up to 49% fewer tokens than its predecessor on Terminal-Bench 4.0 coding tasks.
Cyberattack on Fireworks AI allowed hackers to access Notion customers’ chat transcripts
Attackers breached inference provider Fireworks AI through its cloud provider and reached Notion AI chat transcripts affecting around 1% of Notion’s customers, in Fireworks’ second incident in two months.
AlphaGenome 2: Deciphering how cells read DNA to better understand disease
DeepMind’s conditional DNA model, trained on over 107 million single-cell profiles across 68 tissues, identifies roughly 30% of causal variants on TenK10K against the original’s 16%.
EmbeddingGemma 2 technical report
Google extends its open EmbeddingGemma line to native multimodal retrieval, placing text, images, audio and video in one shared embedding space on a more compact backbone.
MLOps / LLMOps / AgentOps
A practical guide to testing model migrations ahead of GPT-5, Gemini 2.5 and Claude Sonnet 4.5 retirements, with an open-source kit for comparing outputs, identifying behavioral regressions and turning failures into repeatable tests.
Building an Agent Harness for OTel Instrumentation with GenAI Semantic Conventions and Weaver
A technical blog post about six coordinated agents that write framework instrumentation and gate it on Weaver live-check violations, covering 29 attributes across 14 LoongSuite Python instrumentations.
Adapting to AI penetration testing orchestrators: The Modern Security Model
A threat-intelligence post about daily AI pentest traffic growing 30x since October 1, with over 50 framework families scanning customers and more than 12,000 ARTEX-lineage repositories.
Learning
How we made Opik dataset and experiment ingestion 30x faster
A detailed engineering post about fixing eight bottlenecks, from a needless per-dataset lock to O(n²) parameter binding, which cut a 30K-item upload from 39 minutes to 80 seconds.
Building strong safety evaluations
A methodology guide about designing safety evals, noting that single-turn harmless response rates already sit at 94–98% and an API-versus-product configuration gap reaching 34 percentage points.
External sandboxes for coding agents
A post about running agent-executed code in a dedicated sandbox while the agent stays outside, with three reference servers built on v86, Apple Containers and Microsoft’s Quicksand.
Co-Designing Sparse Attention for Efficient Long-Context Inference
A deep dive about eight design rules for trained sparse attention, where the indexer and Top-k stage consume 47% of a DeepSeek V3.2 prefill layer at 128K tokens.
How good are Fable 5.2 and GPT-6.2 Astra at real work?
A benchmark report about grading six models on real Epoch work, where GPT-6.2 Astra averaged 65%, 17 points below Fable 5.2 despite a roughly tied ECI score.
Build Durable, Secure Long-Running Agents
An engineering write-up about an overnight eval-driven research agent with spend caps in plain code and credentials behind a metering proxy, whose winning 26-round run cost $74.63.
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.
The open-source agent for CI/CD reliability.
Papers & Publications
SWE-CoWork: Benchmarking Coding Agents as Digital Employees in Organizations
A coding agent working within a software organization must establish what to build, collaborate with colleagues, and verify what it delivers. We introduce SWE-CoWork, a benchmark comprising 101 development assignments across 101 real repositories, 11 programming languages, and 12 application domains. Each assignment places an agent in a simulated engineering organization with 26 colleagues who have distinct responsibilities, bounded knowledge, and controlled collaboration behavior. Starting from a brief handoff, agents must recover requirements distributed across colleagues and workplace records, reconcile conflicting evidence, and implement changes evaluated by private executable tests. We evaluate eleven models and find a substantial gap between partial implementation and complete task delivery. The two leading models satisfy approximately 63% of fail-to-pass (F2P) tests but complete fewer than 8% of assignments. Providing complete requirements upfront improves F2P by approximately 16 percentage points for Seed Evolving and GPT-6 Astra, yet leaves complete task success unchanged. Requirement-level trajectory analysis reveals failures throughout the delivery process: Claude-Opus-5.5 directs a mean of 93% of first questions about queried cards to their accountable owners across runs but obtains only 64% of colleague-held requirements. Deferred questions go unfollowed, implementations remain incomplete even after requirements are acquired, and 31% of requirements reported as completed fail their tests. These findings highlight sustained requirement tracking, from clarification through implementation, verification, and handoff, as a key challenge for reliable autonomous software engineering.
Self-Improvement Memory Enables Continual Recursive Self-Improvement
Harness optimization is often studied with a fixed policy model and benchmark. When either changes, the optimizer must decide which parts of an evolved harness to reuse or revise, but the harness alone may not explain why earlier edits were made or how they performed. In this work, we study Continual Recursive Self-Improvement (CRSI), a setting for continuing existing harness-optimization methods across compatible benchmarks or policy models. We propose Self-Improvement Memory (SIM) for the harness optimizer, rather than the task-solving policy. SIM groups edits by recurring harness problem and records their rationale, predicted and observed outcomes, and evaluation context. New attempts extend this history as optimization continues. In our experiments across four domains, SIM achieves higher final scores than harness-only and root-cause-only continuation in cross-dataset and pairwise cross-model comparisons. Across ten independent source runs in coding, paired continuations with SIM improve final scores by an average of 2.36 points over harness-only continuation.


