Fluency is easy. Reasoning is hard. For models to be effective, they need to do more than just sound human, they must also be able to reason. How do we teach them to do this? Read here: toloka.ai/blog/how-post-…

Join Toloka at #ICML2025! Come meet the Toloka team at the Vancouver Convention Centre, Silver Pavilion, and don’t miss our social event, “Agents and Safety,” on Wednesday, July 16, from 7:00 to 9:00 PM in West Ballroom D. Find more details at: icml.cc/virtual/2025/4…

Our work was featured in Shopify’s article “Augmented Commerce: Machine Learning at Shopify”! We’re proud to contribute to helping merchants succeed on the Shopify platform. Read the full article here: shopify.engineering/machine-learni…

Great week at #ICML2025 in Vancouver! Our “Agents and Safety” social covered key AI agent safety challenges such as evaluation, red-teaming, and real-world trade-offs. Thanks to all who joined and shared insights. Looking forward to continuing this important conversation!

Harmful content can be hidden deep within long texts, and detecting it is a major AI safety challenge. It happens because subtle, context-dependent risks are often scattered throughout lengthy documents or conversations, making them difficult for AI models to detect without

Toloka is heading to #ACL2025 in Vienna, July 27th – August 1st. Join us at the Generation, Evaluation & Metrics (GEM) Workshop on July 31st, where we’ll present our poster on U-MATH: A University-Level Benchmark for Evaluating Mathematical Skills in LLMs. Find more details at:

Our CEO Olga Megorskaya explores how agentic environments help enterprises safely test AI agents before deployment in her latest article for Forbes These environments uncover critical vulnerabilities that traditional testing might miss, protecting systems from costly failures.

The Toloka team is back from a productive week at #ACL2025 in Vienna. The conference provided a valuable forum for connecting with the global NLP community and observing the field's continued international growth. Our team presented our poster on U-MATH, a university-level

Great insight from Dwarkesh Patel on the Dwarkesh Podcast about one of AI’s key bottlenecks: continual learning. He’s right. To move beyond today's static models, we need to solve two core challenges: the fundamental theory of continual learning and the lack of high-quality data

Great energy at #SIGGRAPH2025 in Vancouver yesterday! Our very own Catherine F. took the stage to present Toloka's "Mainstream Movies Video Eval Toolkit," sharing deep insights into the methodologies behind evaluating the next wave of generative AI. Her presentation highlighted

Congratulations to our partners at Moonvalley on the recent launch of Marey, the world’s first model trained entirely on licensed, high-definition footage! 🎬 This is a huge leap forward for generative video. We’re proud to be partnering with the Moonvalley team on model

Want to share a bit of what I’ve been working on a lot recently with Toloka. I always wanted a service that takes its time, but then exceeds even the longest-thinking models. So we’ve been building Centaurus: people and LLMs working synergistically together, exceeding the

Was lucky to try this on a special project thx to Mikhail Parakhin. The way AI and a human work in tandem feels obvious once you see it. Elegant. Effective. Perfectly in "Tendem". Nice work guys!

Tendem's AI agent—tested in fully automated mode without any human input—performs on par with leading AI systems on standard industry benchmarks:. Dive into our benchmarks and results: toloka.ai/files/tendem_w…

My main focus is always on machines (especially now—making sure Shopify’s infrastructure is in tip-top shape for BFCM), but humans, it turns out, are still pretty useful :-) toloka.ai/tendem-benchma…