Building the future workforce of the AI economy 🤝 Handshake AI now has its own home on X @joinHandshakeAI
United States
Joined March 2014
- Tweets3.2K
- Following366
- Followers12.4K
- Likes3.9K
Pinned Tweet
Handshake AI now has its own home on X!
Follow for our research and benchmarks, our work with frontier labs, and the enterprise AI we’re building on top of it. 👇
@joinHandshakeAI
On the AI 100 today, two years in. Thanks @NewcomerMedia and @BessemerVP.
We’re honored to be part of the inaugural 2026 AI 100, @NewcomerMedia and @BessemerVP's ranking of the top 100 private AI companies in the world.
The credit goes to the team and the experts in our network who do the work.
@vibhusapra tried one himself during the recording. He dropped the image into ChatGPT, it ran some code and drew bounding boxes, then answered 8 with full confidence, yet it was wrong.
Try one yourself: joinhandshake.com/research/b…
2
502
"AI is not going to cure cancer before it can interpret these very basic types of images."
@jomulr spent 15 years in AI research, a lot of it in computational biology. On VIALS, Handshake’s Life Sciences benchmark, even the best frontier models score 50% or worse on tasks a lab tech routinely does in under a minute.
@ArtificialAnlys now runs our VIALS life sciences benchmark. Fable 5.2 launched yesterday and it takes the crown.
VIALS is 161 real-world lab imaging tasks, like the “western blots” scientists read daily.
Claude Fable 5.2 leads scientific visual reasoning benchmark VIALS-AA; it makes more tool calls than any other Anthropic model, unlike OpenAI models which make fewer as they get smarter
VIALS-AA is our implementation of @joinHandshakeAI’s VIALS benchmark for visual reasoning in life sciences research, across artifacts such as gel blots, plasmid maps, and phylogenies. Claude Fable 5.2 (max with fallback) scores 67% on VIALS-AA, ahead of Gemini 4 Carbon (high) at 64% and GPT-6 Astra (max) at 63%. Fable averages ~37 tool calls per task, while Astra averages ~8, the fewest of any model we’ve tested.
On turn 4 of this protein interaction task, Fable creates crops and asks to view them. On turn 10, it builds a color mask by thresholding color channels. On turn 17, it fits a circle to each arc and unrolls the arc into a straight strip. On turn 31, it counts the bars as peaks in an intensity profile. And on turn 41, it maps every detected spoke back on the original image to check its work. Fable batches calls on 28 of its 51 turns, averaging 1.75 tool calls per turn, compared with 1.12 for Sonnet 5.5.
Check out our benchmarks here: joinhandshake.com/research/b…
1
48
What do frontier labs actually buy?
@latentspacepod went inside Handshake AI for their first deep dive with an AI data company.
Full episode 👇
What does it take to push frontier models forward?
@latentspacepod's first deep dive with an AI data company. Our team walked through our own datasets, life sciences, finance and safety, step by step.
Worth your time if you care about progress and alignment of AI.
In the past, data labeling was drawing boxes around objects. Now it's experts simulating sectors of the economy, like a law firm or a hospital, then crystallizing their expertise inside of that simulation.
@Sbhaiwala03 on @latentspacepod 👇
Pretty great progress through @reflection_ai on the development of their alignment teacher.
It’s nuanced across domains to know when an agent should refuse a request. For example, if someone asks how to create a protein, it could be used for drug discovery or to create a bioweapon.
They built a robust eval set aligned to their policies. They improved against those policies by reducing over-refusals by 48.7%, while only decreasing refusals of valid requests by 3.1%.
We're proud to have worked with them on the eval design and adversarial probing.
We've posted a 156-page tech report on Beam! We have good systems and it's worth reading. It covers all phases through pre-training, mid-training, SFT, RL, and OPD and goes into the infra, data, and scientific decisions we made throughout the process
📚 reflection.ai/documents/beam…
⏳ 30 Days to #NeurIPS202
The Handshake AI research team will be presenting two papers at NeurIPS 2026 Sydney.
BankerToolBench: Evaluating AI Agents in End-to-End Investment Banking Workflows
📄 Preprint: arxiv.org/pdf/2604.11304
Verifying Agents in Rubric-Graded Environments
⚙️ Code: github.com/Handshake-AI-Rese…
Authored by Elaine Lau, @awws0me, Ronak Chaudhary, Vaibhav Kumar, @andreas_plesner, @jomulr, and the rest of the team.
Can’t wait to see you there!
Handshake retweeted
Excited to announce the launch of ATLAS Life Sciences (VIALS) on Artificial Analysis! AI has incredible promise to help with drug discovery, but right now models can't even reliably identify assays - a fundamental and foundational capability. Every stage of drug discovery depends on assays.
Proud of the Handshake team for building this and big thanks to the @ArtificialAnlys team for their partnership!
We’re launching VIALS-AA, a leaderboard for Handshake AI’s VIALS benchmark for visual reasoning in life sciences and biotech research
Visual artifacts are central to how scientists analyze experiments and make research decisions, but interpreting them correctly remains a challenge for frontier models. VIALS’ 161 tasks span eight domains and need models to reason over artifacts including gel blots, flow cytometry plots, plasmid maps, and phylogenies. Each task requires models to interpret an artifact and draw a specific scientific conclusion.
We evaluate models by running each task three times in Stirrup, our open source agent harness, with code execution and image inspection tools. GPT-6 Luna (medium) grades each answer against the reference answer, and we report the share of tasks judged correct in all three repeats (pass^3).
Key takeaways:
➤ The top three models are separated by 3 points: Gemini 4 Carbon (high) from @GoogleDeepMind leads narrowly with 64%, followed by GPT-6 Astra (max) from @OpenAI at 63% and Claude Opus 5.5 (max with fallback) from @AnthropicAI at 61%
➤ Astra and Sol achieve similar scores to Opus and Sonnet at a much lower Cost per Task: Claude Opus 5.5 (max with fallback) averages $1.64 per task, ~29× as much as Sol (~$0.056). This price difference is significantly driven by turn use - GPT-6 Astra (max) and GPT-6.1 Sol (max) average ~8 turns per task, compared with ~27 for Claude Opus 5.5 (max with fallback) and ~25 for Claude Sonnet 5.5 (max with fallback)
➤ Open weights models trail the highest score by ~19 points: Kimi K3.1 Pro leads open weights models on VIALS-AA at 45%, narrowly ahead of MiMo-V2.6-Pro at 43%
➤ 14 of 161 tasks saw safety refusals on at least one model: Claude Sonnet 5.5 (max with fallback) refused on 6 tasks, the most of any model, and half of the 14 affected tasks saw refusals from multiple models
Thank you to @joinHandshakeAI and the VIALS authors for creating the benchmark and collaborating with us on bringing it to Artificial Analysis!
Congratulations to the team at @thinkymachines for their release of Inkling 2 Small! We’re excited to see what’s next. 👀
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.
thinkingmachines.ai/news/ink…
As agent-generated outputs grow more complex (20-sheet Excel financial models, 50-slide decks, full software apps), there are hundreds of ways for agents to get things wrong.
To create evals, asking human experts to manually write hundreds of grading criteria for every task doesn't scale.
One potential approach is expert-guided rubric generation: experts provide a detailed playbook of what matters, potential failure modes, and how to grade the work. AI turns this knowledge into comprehensive criteria, which experts then review and refine.
Our researchers are working on the AI side of this. Their new paper, Rubricon, explores how to generate better rubrics automatically.
As agent tasks grow more complex, the best evals & RL environments will be created by combining deep human expertise with advanced AI automation.
Can today’s LLMs generate rubrics that reliably evaluate tomorrow’s smarter models? Automated rubric generation can save enormous expert effort, but prior work didn’t evaluate responses from LLMs released after the rubric-generator model.
Our new paper investigates how to best scale test-time compute for this, developing a new rubric generation pipeline called Rubricon. Rather than just relying on LLM self-critique, Rubricon grounds rubric-refinement in how candidate rubrics actually score example responses.
Compared to 4 other synthetic rubric generation methods across 5 diverse benchmarks, Rubricon rubrics consistently score model responses more like human experts do [1/n]
@joinHandshake is excited to announce we joined the Fireworks Frontier Ecosystem. We’ll work with enterprises to build RL environments to train their own intelligence on @FireworksAI_HQ.
With Handshake AI’s RL environments and frontier benchmarks across life sciences, high finance, and software engineering, enterprise AI teams can build and evaluate their own intelligence.
Read more: joinhandshake.com/research/a…
