Writeups

Ryan Lopopolo’s experiment began with a constraint: write none of the code himself. Part 1 follows the resulting work on builds, observability, review, and Symphony’s agent orchestration—and the human attention those systems still require.

A roundtable on local tools, fast decisions, human review and the permissions that shape everyday assistants.

Elicit’s approach combines software engineering, curiosity about model capabilities, and product judgment about when to recover, experiment, or report a failure.

From enterprise search to coding agents, durable advantages depend on integration work, product habits, research progress, and the judgment needed to use the results.

Pokémon exposes the gap between choosing a goal and reliably acting on it—and makes improvements in long-running agents visible.

Ari Weinstein and Nikunj Handa explain the interface and API choices behind faster agents, and where their limits remain.

Jason Ma explains human-video pretraining, robot demonstrations, and the hardware and workflow decisions behind Dyna’s physical agents.

From asking questions of a textbook to generating two-person audio discussions, NotebookLM’s development joined model capability with editorial judgment—and a deliberately simple interface.

Inside WorldPrompt, real-time diffusion, robot simulation, and the limits of predicting what happens next.

From selecting wind-tunnel experiments to evaluating generative applications, Humanloop’s story follows a changing bottleneck: turning model behavior into evidence a team can act on.

Part 1 follows the VAE retrospective, Würstchen’s compressed image generation, concept-based controls for diffusion, the compression argument for unsupervised learning, and the opening of Christian Szegedy’s adversarial-examples retrospective.

Part 1 follows the path from rebuilding an NLP course after ChatGPT to web agents, social evaluation, code optimization and the compute constraints shaping academic research.

Part 2 follows tiny image perturbations, tokens that provide working space, selective training data, and cheaper context extension—ending with the distinction between a larger window and the ability to use it.

Part 2 moves from the experiments academia can afford to run into SWE-Bench, contamination detection, and GAIA’s tests of everyday assistance.

Part 3 of 4 moves from the limits of longer context to synthetic-data selection, adaptive KV caches and distributed training, then follows Sasha Rush’s tutorial on recurrent models for language and diffusion.

Part 4 follows stepwise reasoning feedback into deployment safety, browser automation, and multi-agent software development, then closes with the reversal curse: why learning a fact in one direction may not teach its reverse.

From image search to conversational characters, useful AI depends on matching hardware, model size and training data to the task—and meeting its response-time budget.

Part 2 of 2 follows inference engineering from distributing a model across GPUs to keeping video coherent, training for efficient serving, and preserving knowledge across requests.

Part 1 follows a long request through caching, speculative decoding and production serving, then examines quantization fidelity, meaningful speed comparisons and the constraints of local inference.

How Brex combines a small AI team with shared infrastructure—and handles the harder work of evaluation, institutional knowledge, and code ownership.

How Gemini turns a question into an adaptive investigation—and the model, interface, evaluation, and execution choices that make it useful.

Streaming data, reproducible experiments and persistent agents serve a common goal: shorten the path from a research idea to a trustworthy model improvement.

Jason Liu’s April 2024 conversation on turning model outputs into usable data, keeping libraries focused, and learning from experiments.

Intuitive prompt exploration, contested evaluations, open research, and the risks of agents that see only part of a task.

Reasoning helps an agent choose its next move. Reliable action also depends on usable tools, meaningful feedback, memory, and control over the workflow.

Part 1 follows AI engineering’s growing community into inference economics, enterprise agents, synthetic data and the businesses succeeding close to either the machines or the customer.

Part 2 follows the move from impressive models to usable products: simpler image editing, branching agent execution, portable memory, harder evaluations and falling inference costs.

What users keep, edit, share, or discard can help diagnose AI output. Amplitude’s early question-to-chart work connects those signals to product experiments and a longer-term ambition to model behavior itself.

How DNA-encoded screens, careful controls and an agent sandbox turn measured protein–molecule interactions into experiments for models.

Experiments have runtimes. Part 1 follows Lila’s effort to turn physical laboratories into training feedback for scientific AI, from callable instruments to catalysts, materials, and cell therapies.