← All editions

Machines that learn, worlds we imagine

Robots learn from simulated worlds; agents need proof they have finished. Twelve reads on sensing, reliability, and the uses of imagination, with two science-fiction invitations to end the week.

Research, selection, and editorial review prepared with Codex; published by Tinkerer's Hall.

01 / Research
Allen Institute for AI (Ai2)· Ai2· Physical AI

MolmoBot: Training robot manipulation entirely in simulation (opens in a new tab)

THE EDITOR’S NOTE

Read for a concrete alternative to collecting robot demonstrations: procedural environments, randomized cameras and dynamics, and several policy architectures trained on shared synthetic trajectories.

In brief: MolmoBot: Training robot manipulation entirely in simulation

Ai2 investigates whether sufficiently varied simulated demonstrations can support manipulation on real robots without real-world policy fine-tuning. Its pipeline combines MuJoCo, procedural environments, and randomization of objects, cameras, lighting, and dynamics; several behavior-cloning architectures share the resulting training data. Examples include Franka pick-and-place and RB-Y1 door or drawer manipulation. The article describes real-world comparisons and releases training data, generation tools, and evaluation code. The scope matters: these are manipulation experiments on two platforms, with navigation explicitly excluded. Testing other skills and embodiments remains necessary.

In this edition

11 more reads
02 / Implementation
Hugging Face· Dana Aubakirova, Andres Marafioti, merve, Aritra Roy Gosthipaty, Francesco Capuano, Loubna Ben Allal, Pedro Cuenca, Mustafa Shukor, Remi Cadene· Physical AI

SmolVLA: Efficient Vision-Language-Action Model trained on Lerobot Community Data (opens in a new tab)

THE EDITOR’S NOTE

Useful for builders: connects a compact policy's architecture to data cleanup, fine-tuning commands, and the queueing decisions needed to keep a robot moving during inference.

In brief: SmolVLA: Efficient Vision-Language-Action Model trained on Lerobot Community Data

SmolVLA combines a small vision-language backbone with a flow-matching action expert that predicts chunks of robot motion. The article explains visual-token reduction, intermediate-layer features, and alternating attention blocks, then shows how asynchronous inference requests the next chunk while the robot executes the current one. Cube pick-and-place illustrates the scheduling benefit. It also documents cleaning task annotations and standardizing camera names across community datasets, with commands for fine-tuning. The evidence centers on simulation benchmarks and SO100/SO101 manipulation; community pretraining is concentrated on SO100 data, so broader embodiment transfer should not be assumed.

03 / Architecture
NVIDIA Physical AI· Michelle Horton· Physical AI

Developing Healthcare Robotics with GPU-Native Medical Physics Simulation (opens in a new tab)

THE EDITOR’S NOTE

A substantive simulation case study: follows instrument mechanics through imaging and learning, with explicit solver choices and limitations that make the implementation tradeoffs inspectable.

In brief: Developing Healthcare Robotics with GPU-Native Medical Physics Simulation

NVIDIA describes a GPU-resident pipeline connecting anatomical models, instrument mechanics, simulated imaging, and robot learning. In its catheter example, Cosserat rods are solved through a coupled XPBD system so insertion and rotation propagate along the instrument; synchronized fluoroscopy supplies the corresponding view. A separate surgical module models tissue, while action-conditioned video models offer another way to generate observations. The article distinguishes visual plausibility from physical guarantees. Crucially, catheter contact currently uses rigid, one-way vessel containment, and the surgical module is early access. These constraints bound what the described simulations can establish about real procedures.

04 / Research
Berkeley AI Research· Physical AI

RL without TD learning (opens in a new tab)

THE EDITOR’S NOTE

Read for a clear derivation of an alternative to sequential value bootstrapping, with unusually explicit boundaries around where the proposed learning rule currently applies.

In brief: RL without TD learning

Seohong Park explains Transitive RL, which estimates the value of reaching a goal by combining values through intermediate states. Restricting candidate subgoals to states within recorded trajectories makes this divide-and-conquer approach practical; expectile regression softens the maximization that otherwise risks overestimation. The article tests the method on demanding OGBench humanoid-maze and puzzle tasks and compares it with temporal-difference and Monte Carlo alternatives. This is relevant to learning from expensive robot experience, but the demonstrated setting is offline, goal-conditioned learning with deterministic dynamics. Extending it to stochastic environments and ordinary reward-based tasks remains open.

05 / Research
Berkeley AI Research· Physical AI

Information-Driven Design of Imaging Systems (opens in a new tab)

THE EDITOR’S NOTE

The sensing pick: explains how to evaluate and optimize measurement hardware independently of a downstream decoder, broadening the edition beyond robot policy models.

In brief: Information-Driven Design of Imaging Systems

This research treats sensing hardware as an information channel and asks how much an observation reveals about its source. The method estimates measurement entropy from data and subtracts the contribution of a known noise model. Its IDEAL procedure then optimizes the encoder without training a task-specific decoder alongside it. Color-filter design supplies a concrete example, with further evaluations in astronomy, lensless imaging, and microscopy. The estimate is an upper bound, so modeling errors can overstate information. The proposed extension to other sensors depends on deterministic encoding and known noise characteristics; robot-control gains are not tested.

06 / Research
Hugging Face· LLMs & agents

The Agent Said It Was Done. The Database Disagreed. (opens in a new tab)

THE EDITOR’S NOTE

Useful if your agent writes to a database: a plausible conversation can conceal a broken workflow. The executable checks and distinction between occasional success and repeated success offer more practical value than another model leaderboard.

In brief: The Agent Said It Was Done. The Database Disagreed.

ThinkingBox evaluates agents by inspecting the records and side effects they leave behind. Its opening example follows an assistant that correctly investigates a delayed delivery, then wrongly closes the unresolved ticket. Repeating each workflow from a clean backend separates tasks a model sometimes solves from tasks it handles consistently. The post explains isolated tool sessions, executable outcome checks, and cost estimates tied to successful attempts. Its public workflows are synthetic, and proposed improvements such as targeted retries have not been measured on the benchmark. Treat the results as evaluation design evidence, with those limits in mind.

07 / Implementation
Mario Zechner· LLMs & agents

What I learned building an opinionated and minimal coding agent (opens in a new tab)

THE EDITOR’S NOTE

A funny, detailed account of the plumbing beneath a coding agent: incompatible provider APIs, context handoffs, tool output, and terminal rendering. Especially worthwhile for readers who want to inspect and shape their agent's behaviour.

In brief: What I learned building an opinionated and minimal coding agent

Zechner builds Pi around a small agent loop and explicit control of model context. The interesting engineering lives in the seams: preserving partial responses when requests are cancelled, translating conversations between providers, and separating tool content for the model from structured details for the interface. In this 2025 design, ordinary files hold plans and external commands supply extra capabilities. The article explains those choices with code and an unusually candid discussion of their costs. Provider accounting remains approximate, and its benchmark compares harnesses using different models, so it cannot isolate the benefit of minimalism.

08 / Architecture
Anthropic Engineering· Justin Young· LLMs & agents

Effective harnesses for long-running agents (opens in a new tab)

THE EDITOR’S NOTE

Read this for the handoff machinery that makes fresh agent sessions useful. The failure catalogue connects premature completion and abandoned changes to specific practices a builder can adopt and inspect.

In brief: Effective harnesses for long-running agents

A long-running coding agent needs a reliable handover when its context window ends. Anthropic separates initial setup from subsequent implementation: an initializer creates a feature checklist, startup script, progress log, and Git baseline; later sessions tackle individual features and record their work. Browser checks catch failures that code edits or endpoint tests can miss, while the checklist makes unfinished requirements visible. The approach emerged from web application experiments, so transfer to other domains remains an open question. Browser automation also has blind spots: native alert dialogs were a specific source of missed bugs.

09 / Architecture
Allen Institute for AI (Ai2)· Jeremy Tryba· Other engineering

Impactful scheduling for GPU clusters (opens in a new tab)

THE EDITOR’S NOTE

A scheduling post with unusually good human failure modes: GPU squatting and universal high priority. It connects allocation policy, implementation, rollout evidence, and the interactive workloads that became worse.

In brief: Impactful scheduling for GPU clusters

Ai2 replaced priority competition with budgets for GPU time, hierarchical fair sharing, and an explicit preemption contract. Jobs declare a minimum useful runtime; after that protected period, the scheduler can reclaim capacity and requeue resumable work. Idle capacity remains available without charging a team's budget, allowing spare resources to stay busy. The result is an infrastructure lesson about incentives as much as algorithms: occupying hardware pointlessly now consumes an allocation. The authors report improvements after rollout, but interactive sessions suffered when preemption destroyed volatile state, and possible fragmentation affecting large jobs remained under investigation.

10 / Perspective
Alex Kladov· Other engineering

Benchmark In Milliseconds (opens in a new tab)

THE EDITOR’S NOTE

A concise palate cleanser after the longer systems pieces. It treats benchmark duration as a design choice for the person doing the optimization, with a useful balance between overhead and feedback speed.

In brief: Benchmark In Milliseconds

Kladov suggests sizing a microbenchmark so a run takes roughly 300 milliseconds. That keeps results easy to scan, gives fixed startup costs less influence, and makes improvements perceptible without slowing the editing loop. Repeated runs stay cheap enough to inspect variation, while the delay remains long enough for a human to feel a change. The important boundary is explicit: this is a heuristic for building intuition and making engineering decisions. It does not claim to provide a method for precise performance measurement. The useful takeaway is to design the measurement loop around its purpose.

11 / Fiction
SB Nation· Jon Bois· Science fiction

17776: What football will look like in the future (opens in a new tab)

THE EDITOR’S NOTE

A strange, funny, unexpectedly tender web serial. Let the opening page unfold before deciding what this is about. Its shifting format matters as much as the dialogue; explore it in a browser, with time for its video chapters.

In brief: 17776: What football will look like in the future

Jon Bois starts with what looks like a sports article and then lets the page become something else. Across conversations, maps, and video chapters, football expands into an imagined far future and a way of exploring ordinary places. The pleasure comes from the contrast between absurdly elaborate games and familiar, small human concerns: friendship, boredom, belonging, and the need for something to do. This is a multimedia serial, so follow its chapter sequence and give the opening room to work. The premise is science fiction; the invitation is to play.

12 / Fiction
Lightspeed· Ken Liu· Science fiction

The Bookmaking Habits of Select Species (opens in a new tab)

THE EDITOR’S NOTE

A 2012 story built like a field guide to alien literatures. Liu makes the physical form of a book feel strange again, then finds emotional weight in the question of how a culture remembers. A change of pace after systems and benchmarks.

In brief: The Bookmaking Habits of Select Species

Ken Liu imagines several intelligent species whose bodies and environments shape how they make and encounter books. An early example turns reading into the recovery of a recorded voice, with the act of reading wearing away part of what it preserves. From there, the story keeps changing what a book might be and what might count as reading. Its form resembles a set of compact observations rather than a conventional adventure. The interest lies in how each invented medium opens a different relationship between memory, interpretation, and time. Let the later examples arrive in their own order.

Illustrations assisted by AI, reviewed by the editor.

THE CONNECTING THREAD

Making room for the unexpected

The engineering selections return to a practical question: what survives outside the conditions we planned for? Simulators need boundaries, agents need observable outcomes, and shared infrastructure needs rules people can live with. The two stories open a different space for curiosity. They invite us to linger over games, books, and ways of being that no benchmark could tell us to want.