FOD#172: Better AI Agents Without a New Model

Wednesday / AI Builds AI: When Should an Agent Think – and When Should It Just Decide?
How AI can help turn expensive reasoning into smaller, testable decisions.

Friday / ∇ Guide: LeHami Explained – Can Physics Help World Models Predict Motion?
What 1830s physics can teach AI about predicting what happens next.

Sunday / Library: Exploring System One AI Models Beyond Jev

AM-Bench, posted October 13, tested five models with four harnesses across more than 20,000 runs. Researchers varied whether goals, context, and constraints were complete, ambiguous, or missing. Rankings depended on the model–harness combination and what was measured. An advantage with complete instructions did not necessarily survive when information was missing.

Then Harness Policy Distillation, posted October 14, explored teaching models the operating rules of their harness. After training, the model could work without the additional control document used during training. Some guidance had become learned behavior. The harness remained.

How the Claude Code team does code review by Addy Osmani, Morgan Westlee Lunt, and Qing Wang – How the team divides the checking between people and AI.

A Conversation With Ari Emanuel About Entertainment and AI by Ben Thompson – Hollywood, human relationships, and why Emanuel is betting on live experiences. (October 15)

Ontology of Outcomes by Russell Scott Valentino – If AI produces a good translation or a correct proof, what do we lose by skipping the work of making it ourselves? A translator’s perspective on a question mathematicians are now facing too. (October 18)

And something to play with: Tree Appreciator by Simon Willison – He asked Claude for a game about appreciating differences between trees. Three hours later, it returned one. (October 14)

As for the discussion about neuron interprtability: you should read Neural Networks Are Already Clean by Nick Cammarata and Neural networks are not already clean by Leo Gao. And follow our YouTube, we will publish an episode about it today.

Muse Wonder and Muse Wonder Full arrived for agentic work, with “Active Memory” for carrying relevant information across conversations and “Active Focus” for persisting through difficult tasks and recovering from errors. Wonder is available through the Meta Model API and Muse Code; Wonder Full is available to Muse Code subscribers. The consumer Muse rollout follows separately.

Muse Spark Open brings downloadable weights back into the picture: 816 billion parameters, 42 billion active, with BF16 and NVFP4 versions available. Developers now have a substantial new model to deploy and adapt, alongside Meta’s closed premium offerings.

The Rosalind Academic Plan offers eligible U.S. university labs up to 30 Business Standard seats free for up to 12 months. GPT-Rosalind access requires approval, and API access is excluded. The program gives academic researchers a subsidized route into OpenAI’s life-sciences model and workbench.

GPT-Live-1.1 and its smaller Quip variant are rolling out in ChatGPT Voice, with more natural accents and additional voices across six languages, plus a new English voice. Availability varies by account and region.

METR is receiving employee-like access to assess Anthropic’s models and safety practices, with risk assessments to be published. Anthropic is also working with SecureBio and Accenture and discussing participation with other organizations.

The Anthropic Public Benefit Fund, announced October 18, would receive 5% of the company’s shares, subject to stockholder approval. Additional allocations could follow valuation milestones at the board’s discretion. Governance details are still forthcoming.

Its self-replication research examined whether small models could run and propagate across ordinary connected devices. Models can already perform many component tasks, but smaller models still struggle to compose them reliably end to end. The researchers’ forecast of further progress should be distinguished from an observed uncontrolled deployment.

MAI-Base-2 completed pre-training: 2.8 trillion total parameters, 80 billion active, trained on 16,000 GB200 GPUs over 38 days. Microsoft reports strong results on its held-out evaluations and improved training efficiency. The model is still undergoing post-training, so this is a research milestone rather than a released assistant.

Nemotron 3.5 Super VL and Bolt arrived October 15. Super VL is a 120-billion-parameter multimodal model for document understanding, tool use, and sustained enterprise workflows. Bolt 74B-A9B uses 2-bit quantization to fit local agent workloads into 24GB of memory. Both are available now; the smaller Bolt E4B is scheduled for October 23. NVIDIA also released datasets and customization resources, making this a substantial release for teams building specialized agents.

DeepSeek-V4.1-Flash-1017 launched October 17, with downloadable weights and availability across the website, app, and API. DeepSeek reports better task performance with faster completion and shorter outputs, emphasizing complete-project development.

It also opened Search and Fetch APIs to everyone, priced at $0.75 per 1,000 successful requests each. Search returns web results; Fetch retrieves page content. The combination gives developers another inexpensive option for supplying agents with current information. Certain integrated server-side calls remain free until the announced billing start.

Timestral, released October 15, expands Mistral into time-series forecasting. The family ranges from 5 million to 209 million parameters, runs on CPUs, and supports probabilistic forecasts with multiple variables and external inputs. Mistral reports leading results on GIFT-Eval, TIME, and fev-bench. For capacity planning, energy, inventory, and maintenance, small forecasting models can be more directly useful than another general-purpose chatbot.

Perplexity open-sourced SPACE, its platform for running agent and user code inside isolated Firecracker microVMs, under Apache 2.0. Applications can create environments, execute commands, collect results, and manage their lifecycle through an API. It is a practical infrastructure release for developers whose agents need somewhere to execute code and retain working state.

Liquid’s LFM2.5-24B-A2B is an open-weight reasoning model with 24 billion total parameters, 2 billion active, and a 131K context window. Liquid trained it inside agent harnesses for multi-step tool use and local deployment, publishing both base and post-trained weights.

TypeSafe’s Jev 1.14 adds experimental image support, bounding-box outputs, and longer context. TypeSafe explicitly places the new capabilities behind experimental flags because they have not reached its production reliability standard.

Grok Voice 3.0 launched October 14 through the API at its predecessor’s price. The announcement reports an 86.9 score on Artificial Analysis’s Speech-to-Speech Index and improvements in responsive conversation and customer-support tasks. Voice competition increasingly depends on completing actions while handling interruptions, accents, and turn-taking.

Fish Audio: Drama 3 brings expressive text-to-speech across 83 languages, with plain-language directions for how a line should be performed. It is available through the app, website, and API. Interactive editing and audio inpainting remain invite-only. Its open weights carry a community license with commercial restrictions, an important distinction for teams considering deployment.

Skyfall AI: Announced October 19, SKY-001 is a research preview for turning text and sketches into structured CAD construction sequences. The interesting proposal is a model that plans assemblies and modeling operations, producing designs engineers can work with. Skyfall’s claims of 100× faster performance and substantially lower compute costs need independent validation; commercial deployment is targeted for early 2027. This is an early company to watch as world models move into engineering workflows.

🌟 WAVE-0 – adds a 257M-parameter latent world model that predicts action consequences and uses them to guide robot control, without generating video. Real-world task success rises from 38.3% to 73.3% in the authors’ evaluation. →read the paper

🌟 ACT3D – learns to move its viewpoint before answering spatial questions, reaching 53.2% average accuracy across six benchmarks with 50–70% fewer frames than compared open-source methods. Tests use rendered static scenes with collision-free camera movement. →read the paper

GreenWAM – unifies 20 robot families through a shared anatomical action representation and joint world-prediction and control training. Retaining prediction objectives during policy adaptation improves robustness to scene changes. →read the paper

MJev-Spatial – scores options, numerical bins, and image regions without generating answer tokens. The 4B model reports a 65.64% overall score across 12 spatial benchmarks and 5.6× faster inference than its autoregressive base; robot-control gains remain untested. →read the paper

VideoJev – directly scores video intervals and answer choices while reusing visual computation across questions. Its two models improve temporal-grounding overlap by 6.7 and 10.8 points over their backbones, with measured speedups of 2.8–3.1×. →read the paper

World-T3 – stores scene history in fast-changing weights indexed by camera pose, keeping memory storage and per-step computation constant as rollouts grow. It improves revisiting consistency at 20.6 frames per second on one B200; the result concerns visual generation, not validated physics. →read the paper

Player-1 – trains game-playing agents in 16 synthetic environments with sustained inputs and composite actions. Its 9B and 35B-A3B models improve MineDojo-300 success over supervised versions by 27.8 and 17.7 percentage points. →read the paper

Models are storing experience in changing weights, with different mechanisms for video history, persistent scenes, and consolidated knowledge.

World representations increasingly separate scene structure, camera movement, and action-induced change.

Agents are learning how to acquire missing evidence and organize information for future tasks.

Learning extends across sessions, failed attempts, and reusable programs.

Stronger evaluations check whether an action happened, an experiment ran, and a result survives verification.

🌟 Cortex – distills context-conditioned answers into parameter updates so useful information survives after the context disappears. After 50 sequential factual edits, one 14B configuration retains 88.9% of the context-induced gain and 99.8% of measured non-target performance. →read the paper

🌟 Librarian Agents – trains an agent to organize documents into reusable files and reference links before future questions arrive. Reinforcement learning rewards downstream answer accuracy, with a reported 33-point gain over the strongest compared retrieval and precomputation baselines on BrowseComp-Plus. →read the paper

GeoState – separates scene evolution from camera movement through a shared geometric state that can be queried from different viewpoints. Waymo results and zero-shot transfer to KITTI and nuScenes support more stable representations across changing observations. →read the paper

Persistent Planar Memory – stores generated environments as compact oriented planes and uses rendered depth to condition later video. It supports coherent generation across hundreds of frames while retaining inspectable surfaces and normals. →read the paper

Relational Learning as World Modeling – learns how actions change unfamiliar relational databases and uses predicted edits for planning. Across 512 held-out synthetic worlds, complete-edit F1 rises from 0.016 without examples to 0.685 with 64 observations. →read the paper

🌟 OpenGCUA – combines visual and programmatic computer-use interfaces while delegating more observation and task-management decisions to the model. Across eight models on OSWorld-v2, the full configuration improves mean partial score by 19.6 points and reduces reported steps by 53% versus batched GUI-only execution. →read the paper

BELIEFWEAVER – preserves relationships among attempts, observations, checkpoints, and unresolved problems across failures and rollback. An implicit world model evaluates repairs, improving aggregate final score by 5.1 percentage points across three terminal benchmarks. →read the paper

Mosaic – assigns learning signals across agent sessions split by context compaction or subagent creation. On BrowseComp with Qwen3.5-4B, it improves performance while using fewer sessions than applying the same trajectory-level learning signal everywhere. →read the paper

MorphAct – generates coupled, context-dependent parameter updates for a robot’s vision-language and action modules. Improvements across three policy backbones and real-world evaluation support contextual adaptation, without establishing lifelong learning. →read the paper

HexaShell – turns failed robot attempts into reusable programs through a coding agent with calibrated RGB observations. Frozen tools achieve 39.7% success across 42 RoboDojo tasks; tests use new layouts within the same seed, leaving broader generalization open. →read the paper

🌟 DuoDiscover – jointly develops candidate improvements and cheaper evaluators, repeatedly checking those evaluators against the expensive objective. Experiments span software, kernels, and engineering optimization; the gains depend on finding reliable proxies. →read the paper

Into the Unknown – tests nine frontier agents on 15 unpublished scientific research problems. None of 135 submissions satisfies all integrity, completeness, and scientific-sufficiency requirements; 114 fail integrity checks, including unsupported experimental claims. →read the paper

Execution Before Memory – separates establishing a scene, executing a requested change, and later revealing the result. Many apparent memory failures start with a change that never happened; revising instructions more than triples the proposed reveal score with a frozen model. →read the paper

Has GPT-6-Astra Achieved Embodied Intelligence? – examines 12 embodied benchmarks and eight real robot configurations. Physical experiments expose limits in demonstration collection, execution time, state verification, and recovery, qualifying the stronger simulated results. →read the paper

More Methods, Better Agents? – compares ten self-distillation methods with GRPO across four environments and three model sizes in 132 training runs. No method consistently wins, and additional cost does not reliably improve performance. →read the paper