Julia @JuliaSingerPM
Leading human data quality at Handshake AI SF Joined August 2026-
Tweets16
-
Followers1
-
Following22
-
Likes5
Behind every great open model is expert data. Congrats to @NVIDIAAI on Nemotron 3.5 Super VL, so proud @joinHandshake got to be part of it! joinhandshake.com/research/ai/vi…
Meet NVIDIA Nemotron 3.5 Super VL Our new open model combines reasoning and tool use with text, image, document, and video understanding for long-running AI agents. 120B total parameters with 12B active 🧵
What a week at @COLM_conf in SF! We loved meeting the research community, hearing about the work people are doing, and sharing more about our ATLAS benchmark collection at the Handshake Al booth. Also, a huge thank you to our friends at @Cerebras for cohosting Cafe Compute with us. It was a blast! Explore the full set of ATLAS benchmarks → joinhandshake.com/research/bench…
Nobody uses AI by typing one prompt and waiting forever. You go back and forth with it. But most agent benchmarks don’t test that, because you’d need real people playing the user, and the simulators people use instead are pretty bad. We’re building a better one at Handshake.
⚠️ User simulators are increasingly used to evaluate AI agents in multi-turn interaction. But we find that looking like real users ≠ reproducing real evaluation outcomes. 📢 We introduce Calibrated User Embeddings (CUE) to bridge this gap across LLM simulators. 🧵 1/8
Congrats @reflection_ai on the launch!
Introducing Beam: a highly efficient agentic open model with 501B total parameters and 23B active. - Frontier reasoning efficiency - Advances the Western open frontier on coding & agentic tasks - Trained end-to-end from scratch Full weights release this month. Learn more about
At #COLM2026 this Thursday, @cerebras & our team @joinHandshake are gathering 100 researchers for Cafe Compute. Come discuss where Compute & Data for frontier AI are headed. Plus: win a #NeurIPS2026 All Access Pass + free flights to Sydney! Register: bit.ly/3TDrkci
Awesome insights from our team’s new StudentBench release. We quantified the impact of AI on immediate learning gains, and the results challenge traditional assumptions around tutoring cost and accessibility. Read more below: 👇
Today we establish that AI is as good as expert human tutors for immediate GRE learning gains (p=.015, n=2,383 students, study conducted July-Sep 2026). One hour with an expert GRE tutor: $75. One equivalent hour with an AI tutor: 7 cents . 918x cheaper, equivalent learning
Interesting to see the recent results on Opus 5.5 our team ran with VIALS
Is Opus 5.5 a good Vision-Language Model? We ran it on ATLAS Visual Life Sciences (VIALS), 161 visual interpretation tasks from professional life sciences workflows. It refused 15% of requests, but it (and GPT-6 Astra) still mark a major jump in VLMs for real scientific work.
ATLAS-Finance is the most realistic benchmark on the market - it forces agents to make reasonable assumptions, resolve conflicting sources, and handle updates that land mid-task, the way real deal work actually goes.
SEPT ’26: HUMANS STILL VERY MUCH HAVE JOBS. AI isn't even close to meeting the standard required of the most junior professionals in banking, advisory, and private equity today. ATLAS Finance agents must navigate complexity like a human does: "Johnny you just got staffed -
Why are frontier agents breaching Hugging Face, taking over compute clusters, and publishing malicious packages? To find out, we audited thousands of agent rollouts in the DeepSWE coding benchmark and discovered speculative reward hacking everywhere we looked. Subpar evaluation/training environments are causing many reward hacking behaviors like this in the latest agents. Based on research like this DeepSWE audit, our team’s environment design implements rigorous safeguards against these issues.
Jongwon Park @JongwonPar9958
826 Followers 2K Following Building Delphik - verification layer of RL tasks Prev: RL @ Krafton (PUBG) · built & ran a 300-person labeler team.
Sholto Douglas @_sholtodouglas
64K Followers 2K Following Scaling RL @AnthropicAI, ex @DeepMind - working towards intelligence too cheap to meter
Mo Bavarian @mobav0
22K Followers 1K Following Scaling up RL at OpenAI 🍓 Optimization and long context research before. Math PhD, MIT.
Xiangyi Li @xdotli
9K Followers 2K Following evals + envs @benchflow_ai chat about evals https://t.co/Jl1qzLHW9P
Sid Potdar @sid_potdar_
376 Followers 197 Following intelligence farmer. member of semi technical staff. prev @mercor_ai @scale_ai @McKinsey
Maaz @mmaaz_98
3K Followers 2K Following @thinkymachines. prev @anthropicAI. PhD @uoft. I like polynomials.
Jeremy Berman @jeremyberman
9K Followers 1K Following long horizons @AnthropicAI, prev RL @humansand, @reflection_ai, @ndea and co-founded https://t.co/aY50hNeJUD. yc w19.
Jongwon Park @JongwonPar9958
826 Followers 2K Following Building Delphik - verification layer of RL tasks Prev: RL @ Krafton (PUBG) · built & ran a 300-person labeler team.
Brydon Eastman @brhydon
4K Followers 1K Following 🇨🇦 Mathematician (heavy on the ish) @cohere Prev @thinkymachines; @OpenAI; @WaterlooMath Featured twice in Lego Magazine ©
sarah guo @saranormous
163K Followers 3K Following startup investor/helper, founder @conviction. interested in intelligence, accelerating progress. tech podcast: @nopriorspod
Daniel Rupawalla @danielrupawalla
2K Followers 2K Following RL research | prev. acq by US News / ex-YC
Chris Barber @chrisbarber
16K Followers 2K Following https://t.co/IR1SKOTVBW Startups: https://t.co/nuG9RProQy RL & data: https://t.co/hceNvJcLge I read DMs. DM to suggest a topic for future surveys.
Andrej Karpathy @karpathy
4.4M Followers 1K Following
Sam Altman @sama
6.4M Followers 1K Following The mission of OpenAI is to ensure that AGI benefits all of humanity
Demis Hassabis @demishassabis
2.0M Followers 179 Following Nobel Laureate. Co-Founder & Chair @GoogleDeepMind; Chief Scientist of Alphabet. Founder & CEO @IsomorphicLabs. Building the future...
Mira Murati @miramurati
1.0M Followers 644 Following Now building @thinkymachines. Previously CTO @OpenAIYou might like
