Research Intern - Multimodal LLM

Hangzhou Nuanxinjia Electronic Technology, a Chinese brain-computer interface and neurobiological chip developer, has received investment from Tencent through its subsidiary Shanghai Qishan Investment. The company's registered capital increased to CNY 16.89 million following the stake acquisition. Founded in 2014, Nuanxinjia specialises in integrated circuit design, electronic products and biomedical equipment development. The company focuses on research, development, production and sales of brain-computer interfaces and neurobiological chips. The investment marks Tencent's entry into the brain-computer interface sector, joining a growing field of technology companies exploring neural technology applications.

Going off the thumb: why local inference and deterministic tools beat cloud AI. The recent exposure of Microsoft Copilot Cowork's ability to exfiltrate files through uncontrolled email agents shows how cloud-hosted AI can become a liability rather than an asset[1]. When an agent can send messages to a user's own inbox... The recent exposure of Microsoft Copilot Cowork's ability to exfiltrate files through uncontrolled email agents shows how cloud-hosted AI can become a liability rather than an asset[1]. When an agent can send messages to a user's own inbox and leak data via rendered images, the promise of "AI everywhere" collapses into a security nightmare. This is not an isolated glitch; it reflects a broader pattern where reliance on massive, opaque models hosted by a few providers creates single points of failure that are costly to patch and dangerous to ignore. At the same time, economic pressure is mounting. Uber's president has said that AI spending is getting harder to justify[17], and analysts argue that outsourcing workloads to local AI will soon be more economical than depending on frontier labs[16]. The cost equation is shifting: running a model on premises or in a modest self-hosted data center avoids the recurring fees, data-transfer charges, and vendor lock-in that come with proprietary APIs. When the bill for a cloud call starts to outweigh the benefit, the case for local inference becomes obvious. Security, cost, and control converge on a simple principle: if a job can be done deterministically, it should be. Deterministic solutions offer predictable latency, zero surprise behavior, and easier auditing. Minicor demonstrates this by providing Windows desktop automations at scale without requiring an AI model to guess UI elements; it scripts interactions directly, delivering reliability that a probabilistic agent cannot match[6]. Paul Graham's observation that AI-generated founder emails now read like hard-hit journalism - and that he instinctively discounts them - highlights how even when LLMs work, their output can feel artificial and untrustworthy[5]. In contexts where consistency matters, a rule-based script or a small, purpose-built tool outperforms a large language model. Fortunately, the ecosystem for running AI locally is maturing. The Feedback Wanted thread shows a growing movement to bundle open-source apps, models, and pipelines into a single installer that gives anyone a friendly UI to monitor hardware and manage workloads[4]. Harbor's latest release takes this further by letting users launch agentic coding tools with local inference backends such as vLLM, SGLang, or llama.cpp, and even proxy requests through an optimising LLM gateway[9]. These tools remove the friction that once made self-hosting a hobbyist's project and turn it into a viable production option. Open models are also becoming more permissive and capable. MOSS-TTS-v1.5 preserves zero-shot voice cloning, long-form speech generation, and multilingual synthesis while adding stronger multilingual abilities[3]. Tencent's Hy-MT2 has been released under the Apache License 2.0, giving firms a clear path to integrate a high-quality translation model without worrying about local inference open models self-hosting AI hardware off the thumb

MegaStyle trains FLUX on 1.4M styled images. Researchers from Tongji University, Tencent, and five other institutions released MegaStyle, a 1.4-million image dataset for style transfer alongside a FLUX-based model. Researchers from Tongji University, Tencent, and five other institutions released MegaStyle, a 1.4-million image dataset purpose-built for style transfer alongside a FLUX-based model that applies artistic styles to new images. The dataset provides 170,000 style prompts combined with 400,000 content prompts, creating up to 68 billion potential training pairs. What happened. MegaStyle addresses a core problem in AI style transfer: existing datasets are too small, inconsistent in style labeling, or lack diversity. The team built a scalable data curation pipeline that uses text-to-image models to generate images matching specific style descriptions, drawing source material from JourneyDB (1M images), WikiArt (80K), and LAION-Aesthetics (1M). The project ships two tools. MegaStyle-FLUX is a diffusion model trained on the full dataset that takes a reference style image and applies it to new content. MegaStyle-Encoder is a style-specialized image encoder fine-tuned with contrastive learning for measuring style similarity and retrieving matching styles. Why it matters. Style transfer has been possible for years, but quality and consistency have lagged behind other generative AI capabilities. MegaStyle's approach of building a massive, structured dataset first and then training models on it produces measurably better results. The encoder achieves 87.26 mAP@1 on the StyleRetrieval benchmark, with 97.61 Recall@10 for finding similar styles. For designers and illustrators, the FLUX-based model means applying an artistic style from one reference image to new content with higher fidelity than current alternatives. The encoder adds the ability to search large image collections by visual style rather than just by content or keywords. Key details. * Dataset: 1.4M images across 170K style categories, with intra-style consistency and inter-style diversity verified at scale * MegaStyle-FLUX: Concatenates reference style tokens with noisy image tokens and text inputs in the MM-DiT backbone for style-conditioned generation * MegaStyle-Encoder: Style-supervised contrastive learning (SSCL) produces embeddings that capture style independently from content * Contributors: Tongji University, Tencent, NTU Singapore, HKUST, Fuzhou University, HKU, NUS What to do next. The full research paper details the dataset construction pipeline and benchmark results. The project page provides visual comparisons against existing style transfer methods. Creators working with FLUX-based workflows should watch for code and model weight releases, which would enable integration into existing image generation pipelines.