GPT-5.x Achieves New Results in Theoretical Physics, Daemon Tools Backdoored for a Month in Supply Chain Attack, ASML CEO Confident in Its Monopoly

GPT-5.x Derives New Results in Theoretical Physics

Latent Space has published an interview with Alex Lupsasca, a theoretical physicist at OpenAI, offering a detailed account of how GPT-5.x derived new results in the fields of theoretical physics and quantum gravity.

The article introduces the concept of the "Jagged Frontier" — a nonlinear capability profile in which AI shows only moderate improvements on everyday tasks like writing emails and coding, yet delivers remarkable results in specific domains such as cutting-edge academic research. Indeed, some Reddit users have said they are "not that impressed" with GPT-5.5, but at the frontier of physics, the assessment is reportedly entirely different.

While examples of AI contributing to creative research continue to pile up, a report of "new results" being derived in the highly abstract field of theoretical physics is noteworthy as a marker of how far AI's cognitive capabilities have reached. That said, the peer-review status of the derived results and the details of the collaboration with human researchers have yet to be disclosed.

Daemon Tools Backdoored for a Month in Supply Chain Attack

According to Ars Technica, Daemon Tools, a widely used disk image mounting tool, was found to have been distributing a backdoored version for roughly a month. The attack was reported by the security firm Kaspersky.

The attack began on April 8 and was still ongoing as of the publication of Kaspersky's report. The troubling part is that the malicious updates were signed with the developer's official digital certificate, which allowed them to slip past routine security checks.

Supply chain attacks have been on the rise in recent years, and with the spread of AI tools, the trustworthiness of development tools matters more than ever. Daemon Tools users are advised to check their systems for infection immediately.

ASML CEO Confident in Monopoly: "Nobody Can Catch Up with Us"

In an interview with TechCrunch, Christophe Fouquet, CEO of semiconductor equipment maker ASML, expressed confidence in the company's monopolistic position in EUV lithography technology. "Nobody is going to catch up with us," he said, emphasizing the absence of any competitors.

ASML's EUV lithography machines are indispensable for manufacturing cutting-edge AI chips, and hardware companies such as NVIDIA, AMD, and Apple cannot build their latest chips without them. Fouquet, who became CEO in 2024, was reportedly relaxed as he spoke ahead of a talk at the Milken Institute Global Conference.

The rapid growth of the AI industry, in turn, is deepening dependence on ASML's technology. That said, some countries are pushing forward with the development of their own lithography technology, and how the long-term competitive landscape will shift remains unclear.

PyTorch Proposes "SMG", a CPU/GPU-Disaggregated Architecture for LLM Serving

PyTorch's official blog has published a technical deep dive into "LightSeek SMG", a proposal to disaggregate CPU and GPU in LLM serving.

In conventional LLM serving, CPUs and GPUs are typically tightly coupled within the same server. SMG deliberately separates them so that each resource can scale independently, which is said to enable more efficient use of GPU compute resources.

LLM inference cost is a major challenge for the industry, and infrastructure-level optimization like this carries significant practical value. However, the latency implications of separating CPU and GPU, along with real-world effectiveness in production environments, will require further validation.

LLM Agents Jailbroken Through Tool Chains — arXiv Paper

A paper published on arXiv flags a new security risk for LLM agents: even when individual tools are harmless, combining multiple tools can form dangerous chains that circumvent an agent's safety restrictions.

Agentic AI derives much of its practical utility from the ability to access external tools, but that same flexibility widens the attack surface. The paper also notes how difficult security testing becomes when it must account for the combinatorial explosion of tool combinations.

As agentic AI adoption advances, verifying the safety of individual tools alone is no longer enough; how to guarantee the safety of their combinations has become a critical research challenge.

Removing AI from a Game Made It More Fun — A Paradoxical Lesson

In an article that gained traction on Hacker News, a game developer reported that "after removing AI from the game, it became dramatically more fun."

At a time of mounting expectations that AI can do everything, this case offers a valuable counterpoint to the simplistic notion that "adding AI makes things better." Understanding where AI is not the optimal choice is arguably just as important for putting AI to good use.

---