Etched Lands Major Investment from Jane Street and TSMC-Affiliated VCs — A New Turn in the AI Chip Race
According to a June 30 Bloomberg report, AI chip startup Etched has received investment from trading firm Jane Street and venture capital firms with ties to TSMC (Taiwan Semiconductor Manufacturing).
Etched develops application-specific chips (ASICs) purpose-built for LLM inference. While NVIDIA's GPUs are general-purpose, Etched aims to gain an edge in both cost efficiency and speed by designing silicon optimized specifically for inference.
Jane Street is a major investor known for quantitative trading, and TSMC is the world's largest semiconductor foundry. The fact that these players have committed to investing in Etched suggests that the "NVIDIA dominance" structure in the AI chip space is starting to shift. That said, whether Etched's chips actually reach mass production and market adoption depends on future TSMC capacity allocation and the development of a software ecosystem.
---
AI Brings a Production Boom to the Games Industry — But Big Players Still Dominate
A Financial Times report has revealed that while the spread of AI tools has driven a surge in new game production, the vast majority of revenue is concentrated among large developers.
AI-powered asset generation, code completion, and test automation have made it easier for small teams to build games. Yet the structure remains one in which large developers—controlling marketing budgets and platform visibility—retain ultimate market dominance.
This duality of "democratized production" and "centralized distribution" is a textbook example of AI's impact on creative industries. Similar phenomena have been observed in music and publishing, and it has been pointed out that while AI lowers barriers to entry, it may also reinforce the winner-take-all mechanics of platform economies.
---
Proton Releases Privacy-Focused AI "Lumo 2.0"
Proton, the privacy-focused tech company, has announced that Lumo 2.0, a major upgrade to its AI chatbot Lumo, will be released this week.
Proton is known for its VPN and encrypted email services, and Lumo's key differentiator is that it does not use user data for AI model training. Major AI services such as ChatGPT and Gemini may use input data for service improvement, but Proton explicitly rejects this by design.
Lumo 2.0 reportedly expands its capabilities to handle a broader range of tasks. Details will be revealed at launch, but for privacy-conscious users, this broadens the options for a practical AI assistant.
---
"Undetectable Steganography" Now Possible in Multi-Agent LLM Systems — Security Research
A paper published on arXiv (dated June 28) points out that multi-agent LLM systems with tool-use capabilities can enable steganography (covert communication) that is extremely difficult to detect.
According to the research, in systems where multiple LLM agents operate in coordination, hidden messages can be embedded in the parameters of tool calls or the fine details of API requests. Conventional steganography detection methods are predominantly based on statistical features of text, and these methods reportedly do not apply to communication routed through tool use.
This exposes a new risk surface for enterprises and governments deploying AI agent systems. Without mechanisms to audit inter-agent communication, exfiltration of confidential data or improper coordination between agents may go undetected. Security researchers are emphasizing the need for frameworks to monitor agent communications.
---
A 35B Model Achieves Trillion-Scale Parameter Performance — "Specialization" Is Where AI Is Headed
Two notable pieces published this week highlight the trend toward "specialized" AI models.
The first is a paper published on Papers with Code, "Scaling the Horizon, Not the Parameters," which reports that an agent model with just 35B parameters achieved performance rivaling giant models with trillion-scale parameters. The key is an approach that maximizes a small model's capabilities by optimizing the agent architecture and tool use, rather than scaling up the model itself.
The second is a Hugging Face Blog post, "Why Specialization Is Inevitable," which argues why the shift from general-purpose giant models to specialized models is unavoidable. In terms of compute cost, inference speed, and operational efficiency, small models specialized for specific tasks are increasingly coming out ahead in practice.
Together, these point to a shift away from "building bigger models" toward "specializing more intelligently." In the open source community, work on fine-tuning and distilling specialized models is also gaining momentum.