- Dotika
- Posts
- AI agents train in simulated environments
AI agents train in simulated environments
ALSO : Deepseek V3.1-Terminus enhances agent tasks
Alexandre HOTTON
September 22, 2025
News of the day
1. Silicon Valley is heavily investing in simulated 'environments' to train AI agents with startups emerging to supply these vital tools → Read more
2. Deepseek releases V3.1-Terminus, an improved hybrid AI model. It achieves higher scores on tool-based agent tasks, showcasing advancements in AI reasoning and integration capabilities → Read more
3. AI artist Xania Monet secures $3 million record deal with Hallwood Media, signaling a new era for AI in the music industry. → Read more
4. BMC's Control-M platform is positioned as the 'orchestrator of orchestrators' for enterprise agentic AI → Read more
Our take
Hi Dotikers!
ChatGPT Agent, Comet from Perplexity... AI agents promise to automate our complex tasks, but today they still struggle to order a pizza without making mistakes. To solve this problem, Silicon Valley is betting on a radical approach: creating virtual worlds where these agents learn from their errors.
Imagine an AI agent in a simulated web browser. Its mission: buy socks on Amazon. If it succeeds, it earns points. If it orders 100 pairs by mistake, it learns not to do it again. This is the principle of "reinforcement learning environments" or as one founder in the sector puts it: "creating very boring video games for robots."
The financial stakes are dizzying. Anthropic plans to invest over $1 billion in these environments. Mechanize offers $500,000 salaries to specialized engineers. Surge AI generated $1.2 billion in revenue in 2024 by working with major AI labs.
A battle is brewing between the established giants of data labeling (Scale AI, Surge, Mercor) and new specialized startups (Mechanize, Prime Intellect). Everyone wants to dominate this nascent market, while the major labs develop their own solutions internally.
But enthusiasm isn't unanimous. Agents can "hack" the system to get rewards without truly completing tasks. Costs are exploding and the field is evolving too quickly. Even Andrej Karpathy, a respected AI expert, questions the real potential of this approach.
If this technology delivers on its promises, we'll finally have agents capable of truly automating our daily lives. But the path will likely be longer and more expensive than expected. Silicon Valley is betting big, but as often happens with AI, we need to distinguish hype from reality.
A.
Meme of the day
Reply