Pankaj Gupta

I just published my roadmap for building evals on highly subjective tasks. Here are the key steps involved and practical tricks for reducing noise / increasing confidence… Blog is here: cameronrwolfe.substack.com/p… Why is this difficult? Evaluating agents is already hard due to their broad scope, emphasis on expensive / long-horizon tasks, and non-determinism. Subjective tasks make this even harder because we cannot deterministically verify solutions and must rely on human judgments (or LLM judges). This introduces additional noise into an already difficult evaluation process. Evaluation roadmap. We can build reliable evaluations using a repeatable framework with three components: - Human evaluation: Our North Star for quality. Create quality guidelines that clearly define quality with the help of subject-matter experts, calibrate human evaluators to ensure agreement, and continually monitor evaluation quality. - Offline evaluation: Use calibrated LLM judges or task-specific rubrics as an efficient proxy for human evaluation. This allows us to scale experimentation without relying on expensive human evaluation for every new model. - Production monitoring: Track quality signals from real-world agent usage and use A/B testing to validate improvements before fully deploying a new agent. Multi-stage evaluation. These components form a tiered evaluation process: Offline evaluation → Human evaluation → Production A/B test → Full deployment We can test many ideas automatically, send promising candidates to human evaluation, and validate improvements through production A/B testing. Continual production monitoring then allows us to discover new use cases / failure cases in production that we can add into the earlier stages of evaluation to create a feedback loop. Sampling useful data. Evaluation quality depends heavily on which tasks we select. We want tasks that are diverse, realistic, difficult, discriminative, and accurate. Techniques like structured task taxonomies, model-based difficulty filtering, Item Response Theory (IRT), and LLM-based auditing help us construct more informative evaluation sets. Handling noise. We should leverage statistics throughout the evaluation process: - Power analysis: Estimate the number of evaluation samples needed to detect meaningful improvements. - Confidence intervals: Quantify uncertainty and provide clear confidence signals alongside evaluation results. - Re-sampling: Repeat agent / judge evaluations to reduce within-task variance and potentially tighten confidence intervals. - Paired comparisons: Evaluate models on the same tasks to more reliably detect performance differences. Key takeaway. An improved benchmark score does not always mean our agent is more capable. We need well-designed evaluation tasks, calibrated human / model evaluators, and statistical confidence signals to distinguish meaningful capability improvements from noise (especially on subjective tasks). By combining these techniques into a repeatable evaluation process / roadmap, we can make building and maintaining evaluation suites more systematic, reliable, and scalable.

In April 2009, Ashton Kutcher challenged Twitter's biggest account, CNN Breaking News, to a race to one million followers. Larry King replied: "CNN will bury you." In the video below, Kutcher records the message that started the race. "I currently only have like 50,000 less followers than CNN, which kind of makes me laugh." If he won, he promised to ring CNN founder Ted Turner's doorbell, run away and post the video. Kutcher also pledged to donate 10,000 mosquito nets if he won. CNN matched the pledge. CNN urged television viewers to follow its account. Kutcher went live on Ustream and asked his supporters to help. Early on April 17, Eastern Time, Kutcher became the first person on Twitter to reach one million followers, while CNN had just over 998,000. He opened champagne on the live stream and talked about people creating and broadcasting their own media. CNN had used its television airtime to win a race for Twitter followers. But an actor's account reached a million first. Was this a victory over the power of big media, or had a television star simply moved his fame to a new platform?

Many discussions about RL environments lately. Happy to share that we’ve just crossed $12M in annualized revenue from building RL environments for frontier labs! From open-sourcing RL environments in 2024 to securing our first contract with a frontier lab in January 2026, we’ve come a long way. Since then, we’ve been building 30+ types of RL environments across machine learning, coding, knowledge work, science, computer and browser use, tool use, and multimodality. And we’re just getting started. Would love to chat with folks working on model training and RL environments!

Xiaomi open-sourced the RL environments behind MiMo v2.6. In two-thirds of the coding tasks, the answer is still in the task's Git history, and MiMo finds it.

One year ago today, we formed @modveon. Today, we’re out of stealth with Sivar in El Salvador, our first national deployment. Grateful to @nayibbukele, his team, @coinbase, our partners and our investors for taking a big bet on us. And so proud of our team. Built. Shipped. Live!

⏱️⚙️Introducing *AutoBenchmark* 📊🏁 - Creating benchmarks automatically - Benchmarking benchmark creation - Studying the role & impact of humans in the loop Blog post: facebookresearch.github.io/R… Key takeaways: 1) We find human-agent collaboration gives big wins over agents alone - fine-grained feedback in ideation stage crucial - autobench can be used to measure this in the future with stronger agents 2) We show that it's possible to make *autoresearch benchmarks* for AI research using this recipe – full recursive improvement loop! 3) Important Ingredients: Autobenchmark creation works best with feedback from two sources: benchmark solvers + external verifiers (human+AI). 🧵1/5

The @BostonCollege Investment Committee (an LP and my beloved alma mater) asked for a few thoughts on what's happening in AI. I recorded a test run yesterday morning and then shared it with my partners, who encouraged me to share it more broadly... so here you go! This is not a sales pitch, it's just a reflection on what we're seeing. And it wasn't intended to be shared, so please pardon the rough edges. loom.com/share/c016702964a04…