Ying Liu

Sr Director of Research at @Scale_AI | agents research, RL environments, professional domain agents, AI for science | ex-Meta | ex-Google | Erdős number 3

[Ying Liu @liuying04]

Join Scale AI’s ML Research Internship for a summer of impactful, fun, and rewarding research. Many of our interns complete high-quality research over the summer and go on to join Scale full-time.

A few examples from our awesome 2026 summer interns.
- Spine-Branch Coordination for Multi-agent Computer Use — Mian Zhang (@Guuuuuuuu)
https://t.co/JSuKjZHcvX

- Evidence-First Agentic Verification — Reza (Rey) Sanayei (@r_sanayei)
https://t.co/bgAFjJHysp

[Scale @scale_AI]

From the beginning, Scale was built by people who love solving hard problems. That same ambition is driving what we build next.

Our ML Research Internship Program is accepting applications in SF and NYC. Work on frontier AI research alongside @ScaleAILabs researchers, publish your work, and help build the next generation of AI.

Apply now. https://t.co/WptkoyoTZe

[Zichen Chen (🐱,💖) @my_cat_can_code]

We spent about 13,000 GPU-hours on RSI Arena, tested the models on 20 benchmarks, and collected 3,129 human votes across two rounds. What stayed with me was how much we learned once people tried the models.

At Bake AI, we worked with researchers from @huggingface, @NVIDIAAI, @scale_AI, @Stanford, @MIT, @UCLA, @ucsantabarbara, @nyuniversity, @UW, and @NotreDame

We gave 10 AI agents the same base model and a shared 64-GPU cluster. They picked the data, wrote the code, and trained their own assistants. People at @COLM_conf and online compared the answers without knowing which model wrote them.

We gave the first round’s feedback back to the agents and let them decide what to change. After another training round, eight of ten improved their human ratings relative to NVIDIA Instruct. Those gains did not track benchmark improvements.

The agents could do the engineering. People helped show them what needed fixing.

That’s what this experiment taught us about RSI. AI can take on more of its own development, but our experience with it still matters. If these systems keep getting better, they should help us work better, live better, and spend more time on what we care about.

Otherwise, what is all that compute for?

📄 RSI Arena: Towards Recursive Self-Improvement with Human Feedback

[Ying Liu @liuying04]

- SWE-Bench Pro Paper: https:/arxiv.org/pdf/2610.24206
- @ScaleAILabs Blog: https://t.co/hywVWMxHyQ

[Ying Liu @liuying04]

Congrats to the @thinkymachines team on the release of Inkling 2 Small! Strong results for a smaller model, and another example of SWE-bench Pro V2 helping benchmark the next generation of smaller, more capable models.

- SWE-Bench Pro Paper: https://t.co/EXFi0uAFLS
- @ScaleAILabs Blog: https://t.co/rE4lLm9VC5

[Thinking Machines @thinkymachines]

Introducing Inkling 2 Small. Scoring 31 on the AA Intelligence Index, it's a big jump over its predecessor.

Inkling 2 Small is available for fine-tuning and inference on Tinker today, with full weight release planned soon.

https://t.co/VR02tw73fJ

[Scale Labs @ScaleAILabs]

Congrats to @thinkymachines on the release of Inkling 2 Small!

On agentic tasks, the model ranks 🥈 on SWE Atlas QnA 1.1 and performs on par with Kimi-K3 on the full SWE-Bench Pro V2 set and GPT-5.6 Terra on the hard subset. Notable results for a smaller model.

[Ying Liu @liuying04]

@natolambert In "post-post-training", what counts is whether the job got done. A lot of the art there is systems.

[Ying Liu @liuying04]

Impressive work from @RicursiveAI. 79.1% on SWE-bench Pro v2 Hard with a 65B open-weight model.

Exciting to see increasingly capable models being tested against challenging, real-world software engineering tasks.

SWE-Bench Pro Paper: https:/arxiv.org/pdf/2610.24206
@ScaleAILabs Blog: https://t.co/rE4lLm9VC5

[Ricursive Intelligence @RicursiveAI]

Ripple-65B Preview was trained on 12 trillion tokens, supports up to 1M context length, and achieves:
- 80.5% on TerminalBench 2.1
- 79.1% on SWE-Bench Pro v2.0-Hard
- 44.3% on DeepSWE v1.1

All while fitting locally on a MacBook or DGX Spark.

[Ying Liu @liuying04]

Strong work from the @cognition team. As coding models become faster and more efficient, rigorous evals become even more important for understanding real-world capabilities.

Great to see more teams using SWE-bench Pro V2 to measure progress on challenging, real-world software engineering tasks: 89.9% on SWE Bench Pro v2 Hard

SWE-Bench Pro V2 Paper: https:/arxiv.org/pdf/2610.24206
@ScaleAILabs Blog: https://t.co/rE4lLm9VC5

[Cognition @cognition]

Introducing SWE-2 Flash, our new fast & efficient coding model.

On leading evals, it outperforms all American open-weight models, even ones 3x larger.

To balance transparency and cyber risks, we are building a program to securely share the weights with partners & researchers.

[Ying Liu @liuying04]

It was a great pleasure to host an old friend from grad school, Professor Yehua Wei (@wei_orms) of Duke University’s Fuqua School of Business, at @scale_AI’s SF office.

Back in school, we often discussed the foundations of research problems and what made them worth pursuing. We also played basketball together on Fridays.

It had been a dozen years since we last saw each other. When the @INFORMS 2026 conference brought him practically next door, I couldn’t pass up the chance to reconnect and welcome him to our beautiful SF office.

Our conversation ranged from advances in belief propagation and distributed algorithms to supply chain and combinatorial optimization, before turning to a practical problem in AI deployment: reducing costs while preserving quality through contextual LLM cascading (https://t.co/lSGIMjvOBF).

Bringing rigorous mathematical modeling to the practical challenges of building and deploying AI agents. There’s a lot for our fields to learn from each other.

[Ying Liu @liuying04]

@chenyx04 Congrats!

[Ying Liu @liuying04]

Interesting! Do LLMs speak their special languages? Are there deeper understanding of the relationships between raw weights and capabilities?

[Rikiya Takehi @rikiyatakehi]

We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.

We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: https://t.co/bOWCcFhCVq

[Yuxin Chen @chenyx04]

Excited to share our new book on reinforcement learning! It presents a self-contained treatment of state-of-the-art sample-complexity analyses for tabular RL.

Statistical Foundations of Reinforcement Learning: A Non-Asymptotic Perspective
https://t.co/BvzdasrD38

[Ying Liu @liuying04]

Yes, need to think about the value and purpose. My 1) is a concrete and timeless goal by itself; in 2) while making progress on 1), define new goals that are challenging for whatever life expectancy. We are far from running out of goals. Just need to think ahead, much ahead of daily work.

[Ying Liu @liuying04]

Even if AI agents and superintelligence surpass us in mathematics, software engineering, and most other professional domains, I don’t think we’re anywhere close to running out of purpose.

Our work may change dramatically. But at least two human pursuits seem timeless to me.

1. The pursuit of a longer, healthier life.

Every culture has its stories about the search for immortality. In China, more than 2,000 years ago, Qin Shi Huang (秦始皇), the First Emperor, enlisted alchemists to find him an elixir of life. He sent Xu Fu (徐福) and thousands of young men and women across the sea in search of the legendary islands of the immortals. Compiled around the same era, the Huangdi Neijing (黄帝内经), or The Yellow Emperor’s Inner Classic, described 120 years as the natural limit of a human life. Even today, we have barely crossed that threshold: only one person has ever been verified to have lived past 120.

The aspiration is ancient and universal. What’s new is that AI may finally give us real tools to pursue it: tools to understand aging, cure disease, and extend the healthy years we get to spend learning, creating, and being with the people we love.

2. The pursuit of a worthy goal that may outlast us.

Curing cancer. Making humanity a multi-planetary species. Understanding the universe. Goals like these can carry us through different careers and stages of life, even when their fulfillment lies beyond our lifetimes. Their value doesn’t depend on our being there to witness their completion or on doing every part of the work ourselves.

If AI helps us reach them sooner, that’s a reason to expand our ambitions, not shrink them. We can take on projects that once seemed out of reach. Longer, healthier lives might even let us complete endeavors we once expected to pass on to future generations.

Human purpose doesn’t require us to be the most capable intelligence in the room. Discovery, exploration, relationships, and the futures we choose to build can still matter deeply to us.

(Yes, I wanted to schedule this message to go live 10 years from now, with no way to retract or delete it.)

https://t.co/PocOuHKU2v

[Ying Liu @liuying04]

Feature request for X: "Locked Predictions."

Make a prediction today, schedule it to be revealed months or years later.

Every prediction gets a publicly verifiable timestamp, and locked permanently—no editing, deleting, or canceling, to prevent hindsight bias and cherry picking.

Especially valuable for AI forecasts, economic predictions, and scientific hypotheses.

Make predictions accountable, not just memorable.
@elonmusk

[Ying Liu @liuying04]

@fel1de I thought about a few compromises to allow deletion (eg canceling all unreleased predictions in one shot) in some justifiable circumstances, but every one can be exploited to cherry-pick. Perhaps it has to go as far as acccount deactivation.

[Ying Liu @liuying04]

Feature request for X: "Locked Predictions."

Make a prediction today, schedule it to be revealed months or years later.

Every prediction gets a publicly verifiable timestamp, and locked permanently—no editing, deleting, or canceling, to prevent hindsight bias and cherry picking.

Especially valuable for AI forecasts, economic predictions, and scientific hypotheses.

Make predictions accountable, not just memorable.
@elonmusk

[Ying Liu @liuying04]

@HanchungLee With a forced release date

[Ying Liu @liuying04]

SWE-Bench Pro V2 has joined the Specialized Intelligence Index.

@ScaleAILabs built SWE-Bench Pro v2 around challenging tasks from applications, business services, and developer tools. Our private commercial cases also help assess performance on codebases that aren’t publicly available.

By joining the SII, we’re helping teams compare models against the demands of maintaining and extending software. These results give engineering teams a stronger basis for deciding which models to test on their own repositories.