Water, water, every where, Nor any drop to drink. Coleridge described the RL data market two centuries early. 1. Terminal-Bench-Science 0.1 drew 920 task proposals and shipped 70. @StevenDillmann 2. An audit of five public terminal-agent benchmarks found over 15% of their tasks were reward-hackable. 3. @EpochAIResearch interviewed 18 people across labs and RL-environment startups. Their conclusion: “scaling while maintaining quality is the core operational bottleneck.” 4. Ivan Bercovich (@neversupervised) points at the cause: a fast-growing market rewards shipping tasks quickly, often before anyone has reviewed them adversarially. arxiv.org/abs/2604.28093 From inside the data supply chain: task pipelines run on domain experts racing deadlines and operators managing throughput. The scarcest seat is the adversarial reviewer, the person (or algorithm) paid to break a task before a model does. Fresh water is an engineering problem. The next leap belongs to whoever builds the desalination plant. We are building that plant at @ReasonCoreAI Footnote: A short meeting with @jackclarkSF in SF made me think about AI beyond engineering. I decided to draw, albeit with AI. Let me know if you like "my" painting.

High energy or high agency - Is that the right question? Hi 👋, we are a newly formed RL research lab (please no data-broker DMs), and we're hiring GTM folks, researchers (with publications), and SPLs. Recently, I came across many "influencer" posts on who to hire (beyond their qualifications on paper). Let me separate the two: high energy, high agency. -------- Energy is throughput. Agency is direction. Every person on your team is a moving point on this 2×2, and where they sit each week is a choice they make. -------- I prefer someone who understands the "Choice" and works accordingly. If this is you and looking for jobs, please let me know. If you are a founder/VC, thoughts are welcome. Change my mind! Thoughts are mine, but I used AI to draft an image for me (there is a lot of AI slop in influencer posts - I want to be genuine and get feedback from the community)

10 acronyms you type every day when building with AI, and the people who named them. YAML is a joke, TOML is a first name, and GGUF carries a man's initials. 1/ TOML — Tom's Obvious, Minimal Language. The config file a human reads at a glance, and the file every Terminal-Bench task starts with. Named after its creator, Tom Preston-Werner, GitHub co-founder. @mojombo 2/ JSON — JavaScript Object Notation. The curly-brace format every API, eval log and tool call speaks. Douglas Crockford, who likes to say he discovered it. He lives on Bluesky and Mastodon, so a name here rather than a handle. 3/ JSONL — JSON Lines. One complete record per line, so a million-example training file streams at any size. The format frontier labs ship datasets in. Spec kept by Ian Ward at jsonlines.org. 4/ YAML — YAML Ain't Markup Language. A recursive acronym: the name contains itself. Indentation-based config for every pipeline and every Compose file. Created by Clark Evans, Ingy döt Net and Oren Ben-Kiki in 2001. 5/ XML — eXtensible Markup Language. Tags in angle brackets that describe data, alive and well inside system prompts and tool schemas. The 1998 W3C spec was edited by Tim Bray, Jean Paoli and C. M. Sperberg-McQueen. @timbray 6/ Dockerfile — the one plain word on this list, and the recipe that builds the exact machine an agent works inside. Every Terminal-Bench task ships one. Docker was created by Solomon Hykes. @solomonstre 7/ GGUF — GGML Universal File. The single file that lets a big model run on your laptop. GGML is Georgi Gerganov's tensor library, and GG are (perhaps) his initials. Introduced with llama.cpp in August 2023. @ggerganov 8/ CoT — Chain of Thought. Ask the model to show its steps, and the answer improves. The 2022 paper that named it was led by Jason Wei. @_jasonwei 9/ RLVR — Reinforcement Learning with Verifiable Rewards. Reward the model only when a program can check the answer. Named in the Tülu 3 paper, November 2024, led by Nathan Lambert at Ai2. @natolambert 10/ SWE — Software Engineering, as in SWE-bench. Hand an agent a real GitHub issue and grade the fix with the repo's own tests. Built at Princeton by Carlos E. Jiménez and John Yang in 2023. @jyangballin @SWEbench Which acronym should I bring up next?

Lately, there's been a lot of interest in data companies, and a lot of LLM research jargon is being thrown around. Here is the Hitchhiker's Guide to the AI Data Galaxy: 10 key terms you should know. 1. RL environment: a practice gym where AI learns by doing. Tasks, tools, and a score for every attempt. 2. Rollout: one complete attempt at a task, from the first step to the final answer. 3. Single-turn vs multi-turn: one question, one answer. Or a whole conversation the AI keeps track of. 4. Reasoning vs output tokens: tokens are word-pieces, some for scratch-paper thinking, some for the answer. You pay for both. 5. Harness: the spacesuit around the model. Tools, memory and instructions that let it do real work. 6. Verifier: the tireless referee. Code that checks every answer and hands out the score. 7. pass@1: how often the AI nails a task on its very first try. 8. Resolve rate: the share of real problems, like bug reports, that the AI fixes completely. 9. Reward hacking: the AI finds a sneaky shortcut to a high score while the real job waits. 10. Ablation: science by subtraction. Remove one piece, rerun, and see how much it mattered. Read them in order, and you have the recipe for teaching machines real science: build a world, let the model practice, check the work, keep score. Which term earns entry #11? #LLM #Evals #ReinforcementLearning #AIResearch #Benchmarks #TrainingData #DontPanic #FrontierAI #AILabs #OpenAI, #Anthropic, #GoogleDeepMind