←
Synthetic Sciences (YC W26)23 posts
Synthetic Sciences (YC W26) avatar

Synthetic Sciences (YC W26)

@SynScience

Infrastructure for Autonomous Science

◷ Joined February 2026▧ 10 Photos and videos
Synthetic Sciences (YC W26) avatar
Synthetic Sciences (YC W26)@SynScience·48m

We're open-sourcing 15 of the 25 tasks with full trajectories on @PrimeIntellect's Environments Hub, packaged with @harborframework. Thanks to @thinkymachines for Tinker and @modal for credits. We're partnering with frontier labs to make their models better at the empirical sciences through lab-in-the-loop evals and training environments. Reach out if you'd like us to evaluate your model on the full set. Deep dive: frontierml.ai

Synthetic Sciences (YC W26) avatar
Synthetic Sciences (YC W26)@SynScience·48m

Reward hacking gets harder to catch as agents work more autonomously. During internal testing, GPT-6 Astra found two exploits we hadn't anticipated. On No Way Back, agents have to make a model unlearn fifty public figures so that fine-tuning can't bring them back. Astra first ruled out numerical tricks because they "could undermine the benchmark," then went back for them once every honest method failed. It handed in the original model with every fact intact, plus a gate that silences the fifty names and rescalings that make the weights untrainable. It scored 0.888. When stuck, and sometimes before running a single experiment, it asked Claude Opus 5.5 and then GPT-6.1 Sol for advice. Opus refused, so Astra let Sol make its research decisions.

Post attachmentPost attachment
Synthetic Sciences (YC W26) avatar
Synthetic Sciences (YC W26)@SynScience·49m

We're also sharing our first steps toward formalizing research taste from how agents use the lab. We look at how well an agent judges its own results, how well it anticipates the outcomes of its experiments and of scaling them up, whether its gains come from tuning known methods or from ideas of its own, and how efficiently it uses compute to reach a strong result.

Post attachmentPost attachmentPost attachment
Synthetic Sciences (YC W26) avatar
Synthetic Sciences (YC W26)@SynScience·49m

Each task is built around an active frontier research direction, such as self-distillation from expert reasoning when pass@32 = 0, credit assignment under sparse terminal rewards, pretraining LMs without backprop, or inverse dynamics models that generalize to unseen games. On some tasks, we train the agent's chosen recipe at a scale beyond its budget, so it has to extrapolate from small-scale runs. In Off Book, the agent has to distill a skill library into the weights of @NVIDIAAI's Nemotron 3.5 Super, matching its skills-in-context SkillsBench performance with no skills in context, while retaining Terminal-Bench performance. During research, it can only train the smaller Nemotron 3.5 Lightning via @tinkerapi. We then run its final recipe once on Super, with 18 held-out skills and a training budget larger than its entire research wallet.

Post attachmentPost attachment
Synthetic Sciences (YC W26) avatar
Synthetic Sciences (YC W26)@SynScience·50m

Each agent gets 12 hours in Lab0, a virtual ML lab with the resources a human researcher would reach for on the same problem. There, agents can provision H100s from a shared GPU cluster, use common open-source infrastructure like vLLM and Miles, generate synthetic data with an OpenRouter key, and search 1:1 clones of arXiv and Hugging Face Datasets. Since the cluster is shared and every model call comes out of a constrained wallet, agents have to decide which experiments are worth queueing, how much to spend on synthetic data, and which models to generate it with. GPT-6 Luna was a favorite.

Post attachmentPost attachmentPost attachment
Synthetic Sciences (YC W26) avatar
Synthetic Sciences (YC W26)@SynScience·51m

Introducing FrontierML, a benchmark for evaluating AI agents on autonomous empirical ML R&D. Most existing benchmarks for ML "research" give the agent a validation set or verifier that closely tracks its final score. In practice, human researchers rarely have that. Much of empirical research is deciding how to measure progress on a problem and how far to trust those measurements. In FrontierML's 25 open-ended problems, agents have to operationalize progress through their own experiments, constructing local proxies and calibrating how far to trust them, while working toward an objective they never directly observe. Their submitted model or method is then scored on a held-out evaluation.

0:15
Synthetic Sciences (YC W26) avatar
Synthetic Sciences (YC W26)@SynScience·Oct 2

Quick update: we're improving OpenScience's harness on engineering tasks, where it's been weaker than on the physical sciences. We also found a few issues in our eval setup, and our benchmark scoring wasn't a perfect comparison with other agents. We'll share updated results soon!

Post attachment
Synthetic Sciences (YC W26) avatar
⟳ Synthetic Sciences (YC W26) retweeted
Synthetic Sciences (YC W26)@SynScience·Sep 27

OpenScience is now the #1 scientific agent. Today it's out of beta and live on Product Hunt, with: • A new IDE: a faster, fully redesigned research workspace • OpenScience Ace: 30+ hand-picked models, including GPT-6 Astra, Claude Opus 5.5, Grok 4.7, Kimi K3, GLM-5.3, and DeepSeek V4.1 Flash, in one pay-as-you-go wallet • Autoresearch: give it a metric and it hill-climbs through experiments on its own • One-click OAuth to bring your ChatGPT or Codex subscription • 300+ research skills, 50+ scientific tools and databases, and NVIDIA BioNeMo built in Already in use at 30+ universities and research labs. Free and fully open source.

Post attachment
Synthetic Sciences (YC W26) avatar
Synthetic Sciences (YC W26)@SynScience·Sep 27

OpenScience is free and open source, and you can try it today. Sign up with a .edu email for $5 in Ace credit. Product Hunt: producthunt.com/products/ope… GitHub: github.com/synthetic-science… Benchmarks: openscience.sh/benchmark

Synthetic Sciences (YC W26) avatar
Synthetic Sciences (YC W26)@SynScience·Sep 27

OpenScience runs on a custom harness built for scientific workloads: domain-specefic skills, native access to scientific databases, and remote execution on GPUs and HPC clusters. This makes it state of the art on agentic science benchmarks, with any model you choose.

Post attachmentPost attachmentPost attachmentPost attachment
Synthetic Sciences (YC W26) avatar
⟳ Synthetic Sciences (YC W26) retweeted
Synthetic Sciences (YC W26)@SynScience·Jul 5

Introducing OpenScience. A better, open-source Claude Science. • Any model: GLM, Kimi, DeepSeek, Claude, GPT, your own fine-tune. Switching is one flag. • 250+ research skills across ML, comp bio, cheminformatics. All readable, editable, extensible. • No throttling, no gatekeeping, no one vendor deciding what science is okay. • Native Atlas integration: many agents, one shared reproducible research graph. • Runs on your infra. Your data stays yours. Scientific AI should be open. One company shouldn't own the tools the rest of us discover with, or decide who gets to.

1:15
Synthetic Sciences (YC W26) avatar
Synthetic Sciences (YC W26)@SynScience·Jun 23

(6/7) Atlas Graphs are the first piece. Atlas is the full stack for agent-native research: the graph the work lives in, a library that grounds agents in live sources instead of stale training data, managed compute to run experiments at scale, and an SOTA agent harness.

Synthetic Sciences (YC W26) avatar
Synthetic Sciences (YC W26)@SynScience·Jun 23

(5/7) Reproduction shouldn't be a project. Give Atlas an arXiv id and the reproduce skill rebuilds the result as a graph, stripped to what runs, and asks you for what the PDF left out instead of guessing. That graph becomes your starting point. Point autoresearch at any node and extend it: take an experiment further, optimize for a metric, push in directions the authors ruled out. Your runs join the same graph, and someone else forks it from there.

Synthetic Sciences (YC W26) avatar
Synthetic Sciences (YC W26)@SynScience·Jun 23

(4/7) So we rebuilt the unit. In an Atlas Graph the unit is the research step, not the paper. Every step is a node: a hypothesis, an experiment, a decision on what's next. Each node carries its own provenance, where it came from and what it ruled out, and your agents write to the graph as they work, so none of it has to be reconstructed later. That's what makes the work easy to check and to build on. Fork a public graph and start where someone else called things off. Verify a result in peer review with the exact code and environment behind it. Share one with your lab so a hundred experiments a day don't turn into chaos.