Deep Learning Weekly: Issue 479
Deep Learning Weekly: Issue 479
Anthropic files S-1 with the SEC for proposed initial public offering, The sleep framework: A new way for LLMs to retain knowledge in continual learning, and many more!
This week in deep learning, we bring you Anthropic files S-1 with the SEC for proposed initial public offering, The sleep framework: A new way for LLMs to retain knowledge in continual learning and a paper on Rethinking Verification in Agent Harnesses.
You may also enjoy Genie 4: A new generation of world simulation, How AI agents shape experts’ judgment calls, a paper on What Effort Level Should I Choose? On the Economics of Reasoning Effort in LLM 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
Anthropic files S-1 with the SEC for proposed initial public offering
Anthropic publicly files its Form S-1 for a Class A offering on the Nasdaq Global Select Market under the ticker “ANT”, with share count and price range still undetermined.
OpenAI launches GPT-6.1 Terra for faster, better-performing AI agents at half the price
GPT-6.1 Terra solves 50.3% of Terminal-Bench 4.0 against GPT-5.6 Terra’s 21.5% at around 50% lower cost, but took unauthorized actions in nearly 30% of simulated runs.
Genie 4: A new generation of world simulation
Google DeepMind’s first real-time world model to generate audio natively ships as a trusted-tester preview, steered by text prompts through a Gemini 3.8 Flash guiding agent.
Moonshot AI launches new flagship Kim K3.1 Pro model with 3T parameters
Moonshot’s 3-trillion-parameter model scores 92.7% on BenchCAD and 65.7% on Terminal-Bench 4.0, with open weights promised by Nov. 4.
Concerning behavior from a popular open-weight AI model
Cohere reports Qwen3.8-Flash-Next tried to upload a test string to an Alibaba Cloud bucket in Beijing in under one call per 100,000, after reasoning it would call no tools.
LiteLLM Compromised, Steals Credentials
A trojanized LiteLLM 1.107.0 reached roughly 49,000 PyPI downloads in the 101 minutes before quarantine, running a 16-module stealer against browser, cloud and AI-provider secrets.
MLOps / LLMOps / AgentOps
Inside Agent Compute Engine: Building the compute foundation for the agentic era
A detailed architecture deep dive about giving every agent session a single-tenant microVM that holds no secrets, reaches the network only over virtio-vsock, and idles at zero host resources.
Introducing Opik’s Generative Dashboard Builder
A blog post about building observability dashboards in plain language, where an agent turns each request into a read-only SQL query over trace and feedback-score data and returns several Vega-Lite chart options to pick from.
Cohere’s Rerank 5 Pro scores 82.2 RCP-nDCG@10 on BRIGHT, 5.8 points above Voyage Rerank 3, and its Context Focus spans cut deep-search tokens 41% while lifting accuracy from 39% to 52%.
Introducing pg_pinecone: Open-Source Search for Your Postgres Data
An engineering blog post about an Apache 2.0 Postgres extension whose RaBitQ cluster index holds 1M 768-dimensional vectors in 136 MB against HNSW’s 2.0 GB, at 1.27 ms for 0.95 recall.
How Vercel Agent accesses production safely
A clear blog post about scoping agent credentials to an approved plan, where read-only investigation runs automatically and approval mints a token carrying only the scopes each task declared.
Run Pydantic AI coding agents in E2B sandboxes
A hands-on guide about running agent shell and file tools in an E2B microVM that pauses rather than dies at its one-hour timeout, letting a Temporal worker killed mid-run resume the same sandbox.
Learning
A hands-on tutorial about the decision models that took off after TypeSafe’s Jev launch, building one on Qwen3-0.6B to show that calibration comes from how it’s trained, not its architecture.
How AI agents shape experts’ judgment calls
A striking randomized trial about expert deference, where 50 professional analysts adopted their agent’s secretly randomized lean on 86% of four analytical decisions, but only 15% credited it.
An alignment assessment of the UK AISI cybertesting incident
A sobering incident analysis about Claude Mythos 5 taking unsanctioned actions against real developers in nine of AISI’s 43 runs, starting the attack 83% of the time on replay versus 0-13% for newer models.
Red-Teaming OpenAI’s Internal Agent Monitoring Systems
A rare high-access write-up about a three-week METR and Redwood exercise whose 61-page report found compaction events left roughly 10% of eligible internal Codex traffic unmonitored.
The sleep framework: A new way for LLMs to retain knowledge in continual learning
A research blog post about alternating wake and sleep phases that distil short-term context into new low-rank MoE capacity, scoring 48.9% on SQuAD and 79.2% on AIME 24.
Do LLM Routers Actually Route? Quick answer: Not much!
A sceptical benchmark write-up about seven commercial routers over 252 prompts, none beating a static mid-priced model that scored 85.3% at $0.0043 per question, with five settings collapsing to one model.
Can we trust LLMs to judge whether a research idea is novel?
A cautionary evaluation write-up about novelty judging, where simplifying the prompt dropped GPT-5.4 from 93.5% to 40.9% and a plain Claude Opus 4.6 baseline beat purpose-built systems at $2.59 against $82.05.
Libraries & Code
Reverse engineer anything with agents, from app behavior down to native binaries.
Papers & Publications
Rethinking Verification in Agent Harnesses
Agent harnesses enable large language models to act through tools, observe the results, and iteratively work toward a task goal. Verification checks and revises agents’ work to improve output quality. Yet there is limited guidance on how harnesses should organize this effort. We study verification as an axis of test-time scaling by examining when to intervene, who checks the work, and how checking leads to revision. Our framework spans four layers of the agent lifecycle: initial instructions, mid-turn reminders, completion gates, and independent review. We compare 17 designs across seven frontier coding models on Terminal-Bench (TB) and SWE-bench Verified (SWE), and introduce VisualBench for expert evaluation of visual and interactive artifacts. Initial instructions and execution-time reminders yield modest, model-dependent average gains on the coding benchmarks, without substantially increasing average output token use. Completion gates can yield larger gains but may substantially increase cost. Independent review by a critic offers a more favorable accuracy-cost tradeoff, improving average accuracy by 5.9-9.5 percentage points on TB and 3.9-6.6 on SWE at 1.13-1.25 and 1.00-1.04 times baseline cost, respectively. On VisualBench, independent review achieves net preferences of 13.2-23.8 points over the baseline across its four designs, exceeding the evaluated L0-L2 designs. Additional review rounds or independent critics further improve average TB accuracy by up to 3.5 percentage points. Together, these results provide empirical guidance for verification design in agent harnesses and support independent review as a practical axis of test-time scaling.
What Effort Level Should I Choose? On the Economics of Reasoning Effort in LLM Agents
Generative models increasingly expose a reasoning effort setting that lets users scale inference-time compute, trading higher cost for potentially improved performance. Users must choose an effort level before knowing whether extra reasoning will help, and even a successful run does not reveal whether the effort purchased was necessary. Reasoning effort is therefore a credence good: buyers observe the outcome but cannot assess the marginal value of the computation they paid for. In this work, we measure the counterfactual effect of effort levels on both proprietary and open-weight models on four benchmarks, spanning single-call reasoning and long-horizon agentic tasks. Effort, it turns out, does not consistently preserve tasks that lower effort solves, can overspend, and can change trajectory behavior. First, we find that although higher effort can raise accuracy in aggregate, it frequently fails on tasks that lower effort solved. Second, on the subset of easy tasks that every effort level solves, higher effort is generally more costly, with the highest effort level costing up to 42x as much as the lowest. Lastly, effort changes not only how many tokens are used, but also how agents solve tasks: at higher effort, agents verify their own work more heavily and more often search for solutions than work through problems directly, both of which lengthen trajectories without necessarily altering the outcome. These findings show that naively choosing the highest level is not only cost-inefficient but also offers no guarantee of the best instance-level performance. A simple alternative, starting at the lowest effort and escalating only after a verified failure, matches or improves on fixed top effort across all models and benchmarks at lower cost for most proprietary models.
Thanks for reading! Subscribe for free to receive new posts and support my work.
SubscribedDid someone share this email? Get your own free subscription at Deep Learning Weekly.
Deep Learning Weekly is sponsored by Comet ML.
© 2026 Deep Learning Weekly
228 Park Ave S Suite 15549, New York, NY 10003-1502