Turing

The next generation of AI agents won't succeed because they know more languages. They'll succeed because they can reason, plan, use tools, and stay contextually correct wherever users are. That's the training signal frontier models need. Read the full case study below:

@jeffreyweichsel Gotta love the blue sprinkles!

Today was amazing! 1. Now I can walk to work at our new Palo Alto office! 2. I got to eat @turingcom cupcakes in celebration! AGI never tasted so good.

@jeffreyweichsel Nom!🧁

@Shreyaan18 @deedydas 👀

@deedydas Shout out to @turingcom

Every single startup selling AI Training Data (July 2026) >50 cos sell data and RL environments to big AI labs and drive AI progress behind the scenes. They total ~$8.5B in rev and ~$100B in valuation, >75% of which are just 4 players: Scale, Surge, Mercor and Handshake.

Turing works both sides of AGI. Research: frontier labs use us to push model capability. Coding first, now all knowledge work. Deployment: enterprises use us to ship that capability. Each side compounds the other. The Secret Turing Master Plan: x.com/jonsid/status/…

Let’s talk data, RL environments and what you’re working on. Booth #B406 at #ICML2026!

More highlights from #ICML2026! Appreciate the researchers and everyone building in the AI space who attended our HH event. Conversations were had and connections were made!

Come find us in booth B460 at #ICML2026!

Look for this crew at @ICMLconf in Seoul. -OTS data packs -Coding Benchmarks -RL environments for post-training -Evaluation -Reward modeling -Production Just a few of the things they want to discuss. Book your meeting below: #ICML2026

The next generation of AI agents won't succeed because they know more languages. They'll succeed because they can reason, plan, use tools, and stay contextually correct wherever users are. That's the training signal frontier models need. Read the full case study below:

The project also scaled 300+ multilingual evaluators with software engineering, ML, and data science backgrounds. Every conversation received 100% human review coverage before delivery.

Quality was built into every stage. Every task was evaluated across 10+ dimensions, including: - Tool accuracy - Hallucinations - System prompt adherence - Datetime reasoning - Dialogue naturalness - Grammar and language quality Then reviewed by both automated systems and human experts.

Each conversation ran 10 to 15 turns and required agents to: - Choose the right tools - Chain multiple tool calls together - Handle corrections - Follow detailed system prompts - Maintain natural dialogue throughout That is closer to production than single-turn benchmarks.

The challenge wasn't language. It was consistency. Every response, tool argument, currency, location, date format, and conversational reference had to remain aligned with the user's locale across an entire interaction.

Most AI agents can speak multiple languages. Far fewer can work across them. Real-world multilingual agents need to reason across cultural context, local conventions, tool calls, and multi-step workflows, not just translate text. Here's how Turing built training data that gets models closer to that reality.

The dataset includes: - 3500+ multi-turn agentic conversations - 15+ locales - Conversations spanning sequential reasoning, tool calls, corrected responses, and locale-specific instruction following This is the kind of data production agents actually need.

@rbordoli @icmlconf Next time!!

@turingcom @icmlconf That’s a good looking crew. Sorry I couldn’t be there with you team.