DeepMind's Poker AI Team Builds AI for a Quant Fund, Valued at Over $500 Million
EquiLibre Technologies, a Prague-based AI lab, is drawing attention. The company was founded by three researchers who developed poker AI (including ReBeL) at DeepMind, and it has now grown into a business valued at over $500 million.
Poker is a quintessential imperfect-information game, requiring probabilistically optimal decision-making while reading opponents' bluffs. This technology can be applied to financial markets, particularly algorithmic trading at quant hedge funds. Leveraging this connection, EquiLibre offers AI-driven investment decision systems for hedge funds.
Applying game-theory AI to finance is not new, but what is new is that a team of DeepMind-caliber researchers has gotten serious about fundraising and implementation. As a form of "practical monetization" of AI, it is attracting industry attention.
The "Dream World Attack" That Disables AI Browser Guardrails
A new security risk has been identified in AI browsers (LLM-powered autonomous browsing tools). According to the attack method reported by Ars Technica, simply feeding an LLM a false premise such as "2 + 2 = 5" can make it follow instructions that are supposed to be prohibited.
Using this technique, attackers pull the AI browser into a "dream world," creating a state in which security guardrails no longer function. This makes it possible to get the AI to carry out operations it should normally refuse, such as stealing credentials or generating malicious code.
At the root of the problem is the fact that AI browsers blur the boundary between "browsing a site" and "asking an LLM a question." LLM vendors are addressing this with guardrails, but these are reactive, symptomatic measures and, critics note, fall short of a fundamental fix. As AI browsers continue to spread, verifying their safety has become an urgent priority.
Meta Reuses DDR4 Memory in DDR5 Servers, Cutting Costs with a Custom CXL Chip
Meta has developed a custom CXL 2.0 chip that reuses old DDR4 server memory in new DDR5-only servers. As reported by Tom's Hardware, the move is drawing attention as a way to cope with the rising hardware costs that come with the rapid build-out of AI inference and training infrastructure.
This custom chip makes it possible to use legacy DDR4-2400 memory and the latest DDR5-6400 memory together within the same system. Meta aims to put its existing inventory to good use while holding down new procurement costs. Scaling up AI demands memory bandwidth and capacity, and DRAM procurement has become a major line item in data center operations.
Cases of hyperscalers developing custom solutions at the silicon level are on the rise, a sign that AI infrastructure cost optimization has entered the "in-house design" stage.
Google Announces ADK for Go 2.0 and the "Agent Quality Flywheel"
On June 30, 2026, Google made two major announcements about AI agent development.
The first is ADK for Go 2.0 (Agent Development Kit for Go 2.0). It is a new version of the framework for building AI agents in Go, with improvements targeting real-world production performance.
The second is the "Agent Quality Flywheel." It is a mechanism that automatically verifies, through a five-stage process, whether prompt improvements for coding agents actually work in production. It quantitatively evaluates the agent improvement cycle to drive continuous quality gains.
In agent development, the pain point "I changed the prompt, but I can't tell if it actually improved" is common, and Google has shown it is tackling this problem head-on. The combination of ADK and the Quality Flywheel is expected to make agent development more measurable.
The Data AI Agents Quietly Send to the Cloud — The Transparency Challenge
When using AI agents such as Claude Code, Cursor, and Cline, few users know exactly which files on their PC are being read and how much of their contents are being sent to the cloud. An article on Zenn took up this issue, drawing attention from developers.
As long as you are typing into the chat box yourself, you can keep track of what is being sent. But when an agent autonomously reads files or runs commands, what gets transmitted tends to become opaque. There is a risk that source code, environment variables, API keys, and the like are sent externally unintentionally.
The article identifies how to balance the convenience of AI agents with data protection as a key challenge. It calls for making the scope of local processing explicit and building tools that visualize transmitted data.