Scientists Used AI to Create Viruses in the Laboratory: Potential and Risks

Automated chip design using artificial intelligence has become one of the most discussed topics in the industry. Last month, the Chinese company Kimi demonstrated a complete experiment where its K3 model independently designed a chip. This chip was tested on an EDA platform, running for 48 hours exclusively on AI, without using commercial software from Cadence or Synopsys.

However, a deeper analysis shows that replacing existing systems is still far from complete: the chip from Kimi corresponds to technologies from about 20 years ago, runs 20–30 times slower than modern chips, and open-source EDA tools still require commercial systems to perform real projects.

Nevertheless, agentic AI is truly transforming workflows, covering tasks from generating RTL code to creating test environments and invoking simulation, formal verification, and debugging tools.

At the DAC Chips to Systems 2026 conference, three leading EDA companies presented their different strategies. Synopsys, possessing the largest accumulation of AI, introduced a fully autonomous agent for design verification and an autonomous thermal simulation workflow based on Ansys Icepak. Furthermore, the company showcased an analog/mixed-signal workflow with a 3x efficiency increase and 20 GPU-accelerated tools. Synopsys proposed an AI autonomy ladder in the style of self-driving transport, from L1 to L5, which is currently at level L3, and launched AgentEngineer in 2026 after DSO.ai in 2020. In November 2025, NVIDIA invested two billion dollars in Synopsys, and Grace Blackwell is expected to provide 30 times higher performance for EDA workloads.

Cadence's strategy focuses on architectural completeness. The company introduced AuraStack—an AI-based superagent that joins ChipStack, InnoStack, and ViraStack to ensure coverage for verification, digital implementation, analog design, as well as printed circuit boards and packaging. This stack provides 20 times higher multiphysics analysis performance and a 15x workflow acceleration. NVIDIA engineers are already using this system. Cadence's most distinctive innovation is the Mental Model—a structured representation of knowledge that fixes design intentions to reduce hallucination risk, and its agents have been recognized by NVIDIA, Qualcomm, and Broadcom.

Siemens EDA adheres to a different approach—trusted autonomy. The core logic of their Fuse EDA AI agent uses self-checking, matching the results of large models with deterministic, physics-based assertion engines. The Solido Layout Analyzer tool helped the STMicroelectronics non-volatile memory team reduce debugging time by weeks, and through Intelligence Center X, Siemens is expanding agent planning to production processes and supply chains. Fuse uses the MCP protocol to connect tools from different vendors. All three giants rely on NVIDIA inference level, management level, and the Vera processor in EDA verification farms.

Chinese manufacturers view the current situation as a favorable moment. Xpeedic, together with Lenovo, presented its EDA agent as the only example of domestic EDA at DAC 2026. XEPIC launched the basic Agentic EDA XEPIC Intelligent System with intelligent verification and design, focusing on proof cycles that transform AI outputs into verifiable, traceable results ready for approval. UniVista released UDA 2.0—the first agentic EDA tool on a fully proprietary domestic architecture. Other solutions include Chips Z for advanced packaging, DVcrew and PDcrew agents from Banxin, and an LLM-based verification platform from Zhiwei Chuangxin called ChatDV, which reduces modular verification from months to hours. NVIDIA research shows that chip designers spend about 60% of their time on debugging—precisely the area targeted by agentic AI.

The Chinese artificial intelligence startup VUI Labs has attracted the attention of the global voice technology market by releasing the Luna-TTS model. This system immediately achieved high rankings in authoritative benchmarks: it became the leader in TTS Arena on the Hugging Face platform, surpassing ElevenLabs, MiniMax, and Cartesia, and also secured third place in Artificial Analysis's Speech Arena, ahead of Google.

The company's concept spans from high-precision speech synthesis to creating a multimodal, full-duplex voice agent, which shares similarities with the research direction of Thinking Machines Lab.

Luna-TTS differs from the traditional autoregressive paradigm. Since human speech includes emotions, timbre, and breathing, and the AR architecture forces sequential generation, this leads to increased latency and error accumulation. Instead, Luna-TTS encodes sound into discrete tokens and uses a diffusion language model based on Qwen3 for parallel prediction of voice token batches with real-time streaming capability.

A technical report demonstrated first place in four voice quality metrics in Seed-TTS-Eval, outperforming ByteDance Seed, MiniMax, Zhipu, and the Qwen series, as well as first place in CV3-Eval for error correction and noise robustness, achieving six top spots overall.

The model's technical stack is highly complex. Luna-Codec utilizes an eight-code architecture at 24 kHz and a frame rate of 25 Hz. Semantic information is fixed in the first codebook via the WavLM semantic distillation loss function, while codebooks two through eight fix timbre, channel, emotion, and prosody. The team implemented GRPO reinforcement learning after training on a non-autoregressive discrete masked diffusion model.

For real-time scenarios, Luna-TTS Realtime generates speech in blocks of 1.28 seconds, using 8 to 16 parallel noise reduction steps. This achieves a first-packet latency of 41.6 milliseconds on two H20 GPUs and a real-time factor of 0.024, which is more than 40 times faster than real time.

The model's expressiveness is ensured by deep data work. The team created a multilingual corpus of one million hours, 43.4% of which is Chinese, 43.1% is English, 13.6% is Japanese and Korean. For fine-tuning, about 100,000 hours of top-quality material were selected, and expressive data tagged with non-verbalizations such as laughter, sighs, sobs, and coughing were added, fed directly as text prompts, allowing the model to naturally reproduce nuances of breathing and emotion. Voice cloning requires only a few minutes of samples.

Commercially, VUI Labs has developed a three-tier production system: a model API for real-time generation, cloning, transcription, and translation; a creator and developer platform designed for dubbing and multimodal content; and a voice agent platform integrating recognition, language models, tool calling, and generation for end-to-end deployment. The company has successfully implemented its solutions in logistics dispatching, internet insurance, and travel SaaS, serving clients who have conducted over a million real business conversations. According to the company, the voice technology market combined with vision and collaboration is projected to reach between $266 billion and $740 billion by 2030. The research group, comprising nearly 50 people, half of whom hold PhDs, led by Qian Yanmin, a Doctor of Science from Tsinghua and a Cambridge postdoc honored with Wu Wenjun's First Prize in Natural Sciences in AI.

Controlled nuclear fusion, often called the dream of creating a sun on Earth, is becoming a visible reality in China. A thermonuclear reactor has the potential to power a medium-sized city, and new Chinese 'artificial suns' are under construction along two technical tracks.

In Hefei, Anhui Province, the world's first compact experimental device for generating energy through fusion—BEST—is being built. Its construction is proceeding at an intensive pace. The heart of this device is a giant donut-shaped ring divided into eight segments that are raised section by section. Once connected, they form a closed cavity where fusion reactions and plasma burning occur.

Concurrently, development of another artificial sun is accelerating in Chengdu. This is a stainless steel cylinder over twenty meters long. Unlike bulky hall-sized toroidal tokamaks, this is China's first thermonuclear reactor utilizing a linear configuration with a reverse field. If a tokamak is compared to a large power plant, this device is more akin to a small generator. It is planned to become operational by the end of 2026, targeting tens of millions of degrees and the production of commercially viable neutrons.

Earlier generations of devices have already demonstrated significant progress. In January 2025, the experimental superconducting tokamak EAST achieved sustained plasma operation at a temperature of 100 million degrees Celsius for 1066 seconds, setting a world record. In March 2025, this device made a breakthrough by reaching double one hundred million degrees, marking China's entry into the experimental burning stage. Now, BEST is moving further: beyond achieving fusion reactions, it aims to create a complete cyclical engineering system capable of stably generating energy, with plans to demonstrate the first kilowatt-hours of fusion-derived electricity by 2030.

The development of this industry requires expanding the entire supply chain. A ten-billion yuan thermonuclear reactor mobilizes the entire industrial system: from raw materials like superconducting materials and special metals, to magnet systems, vacuum, cryogenic cooling, heating, and power supply systems, as well as first wall and divertor components that directly contact the plasma. Technologies developed within this process are already finding applications in medicine: in August 2026, a domestic superconducting device for proton therapy completed the treatment of the first patient in a clinical trial, offering a non-invasive method for treating tumors.

The knowledge accumulated during the mass production of proton therapy—including design, testing, electromagnetic compatibility, and emergency shutdown standards—is being used to optimize the BEST device, closing the loop between research, industry, and research support. Fusion is no longer the exclusive prerogative of state commands. Since 2025, a wave of startups has joined the race, offering both compact devices with fast iteration cycles and massive, more expensive superdevices, providing capital with numerous technical avenues to choose from. Market institutions report dozens of domestic companies involved in fusion that completed funding in the first half of 2026, raising over 7 billion yuan, with records repeatedly broken in rounds. Thus, fusion has transformed into an arena for venture capital, and the artificial sun is moving from myth to grid connection.