Software Engineer Intern - Applied Machine Learning-Enterprise, 2026 Summer
About the job
- Must be able to commit to a 12-week full-time work period during Summer 2026
- Currently pursuing a BS/MS degree in Computer Science or related technical field
- Proficient in deep learning frameworks such as PyTorch or TensorFlow
- Experience with software development in at least one of the following programming languages: C++, Python, Go
- Good sense of teamwork and communication skills, practical experience in relevant business scenarios is preferred
- Strong software programming capabilities, exhibits good code design and coding style
- Must obtain work authorization in country of employment at the time of hire, and maintain ongoing work authorization during employment
- Building a next-generation big model as a service platform to serve hundreds of LLMs based applications
- To develop and maintain the big model as a service platform, including offline training/finetuning, online inference, model management, and resource orchestration, etc.
- To manage a huge number of GPU resources and provide computing power efficiently
- Experience in NLP and LLM technologies
- Experience or research in Alignment for Large language models
- Experience in ModelOps, for example large model management, finetuning workflow management, etc.
- Contribution to top-tier conference papers, including NeurIPS, ICML, ICLR, CVPR, ICCV, ACL, KDD, etc. is a big plus
About the company
ByteDance builds content and creation apps for consumers, alongside collaboration software and AI services for businesses. Its video platforms connect viewers with creators, while CapCut gives people tools to edit and publish videos. The group also develops Lark for workplace communication and shared documents, and its Seed research teams build language, image, video, and speech models used in consumer apps and sold through BytePlus cloud services. It earns money from advertising, commerce services, paid software subscriptions, and usage of its enterprise technology. ByteDance was founded in 2012 and has its operational headquarters in Beijing, China.
Company Size
10,001+
Company Stage
Debt Financing
Total Funding
$35.1B
Headquarters
Haidian, China
Founded
2012
Get referred to ByteDance
See people who can refer or advise you
Simplify's Take
What believers are saying
- Demand keeps expanding: September 2026 reporting put ByteDance's first-half revenue at $120 billion, up more than 30%.
- Employee incentives expanded: ByteDance announced a 35% larger bonus pool in December 2025.
- Enterprise customers broaden demand: Lark reports users across 125 countries alongside ByteDance's consumer businesses.
What critics are saying
- AI investment squeezes earnings: September 2026 reporting said ByteDance's first-half profit fell despite revenue growth.
- US governance limits control: ByteDance holds 19.9% of TikTok's separately governed US data-security venture.
- A BytePlus intern faced rapid change: ByteDance's featured account describes weekly, sometimes daily product updates.
What makes ByteDance unique
- ByteDance separates its video markets: Douyin serves China, while TikTok targets audiences abroad.
- Models reach consumers and developers: ByteDance's Seed powers Dola and Dreamina alongside BytePlus cloud access.
- Enterprise software grew from internal needs: ByteDance says Lark began as a tool for its own workforce.
Help us improve and share your feedback! Did you find this helpful?
Benefits
Hybrid Work Options
Growth & Insights and Company News
6 month growth
↑ 0%1 year growth
↓ -1%2 year growth
↑ 0%ByteDance's Doubao app lands on Huawei HarmonyOS PCs with Doubao-Seed-2.1 Pro and Turbo. Published: October 7, 2026 Want to read in a language you're more familiar with? Doubao's native HarmonyOS PC app ships with Doubao-Seed-2.1 Pro and Turbo, chat and work modes and agent-style tools; Huawei says it topped the HarmonyOS PC wishlist. ByteDance's AI assistant Doubao is now available as a native app on Huawei's HarmonyOS PCs. Chinese tech outlets first reported the desktop client's arrival on October 1, and IT Home confirmed on October 7 that users can download it directly by searching for Doubao in the HarmonyOS app store. Huawei said Doubao ranked first on the HarmonyOS PC user wishlist, the list of apps owners of its HarmonyOS computers most want to see on the platform. The first version ships with two models, Doubao-Seed-2.1 Pro and Doubao-Seed-2.1 Turbo. ByteDance released the Seed 2.1 series on Volcengine in June, positioning Pro for exploring highly complex tasks and Turbo for large-scale production use. Compared with the previous generation, the company says Seed 2.1 improves coding delivery, long-chain agent task execution and multimodal understanding. The app is built around two modes that users can switch between freely: Chat and Work. In Work mode, users set a goal and attach files, images or material from their Doubao cloud drive, then choose Auto, Turbo or Pro and pick a reasoning intensity. They can also create projects, add local folders and set scheduled tasks. A side workbench supports new tabs, so a user can browse the web, read reference material and talk to Doubao at the same time while several tasks run in parallel. For content creation, the launch version includes deep research, PowerPoint generation, a writing assistant, image generation and cloud storage. Video generation, AI blogging, recording transcription, music generation and AI spreadsheets are also available. ByteDance pitches the client at office work, academic research, file management, automated tasks and creative design. HarmonyOS PCs run Huawei's own operating system rather than Windows, which means every mainstream application has to be rebuilt natively before users can install it. A general-purpose AI assistant has been one of the most visible gaps, and Doubao's arrival gives Huawei's laptop owners access to one of China's most widely used chatbots without relying on a browser. For ByteDance, the release extends Doubao's desktop presence beyond Windows and Mac, where the company has steadily turned its PC client into an agent that can operate software, run multi-step tasks and work across local and cloud files.
YouTubers' data scraping suit against ByteDance lives on. The plaintiffs claim ByteDance bypassed YouTube's protections when accessing their videos, cheating them out of advertising revenue. SAN FRANCISCO (CN) - A federal judge ruled Friday YouTubers have standing in a suit accusing tech giant ByteDance of data mining their videos to train artificial intelligence models U.S. District Judge Jacqueline Scott Corley ruled that Ted Entertainment and Matt Fisher properly claimed at this stage that they've suffered what she called a pocketbook injury. Ted Entertainment, a company, and Fisher, an individual, create and post original content to YouTube and claim a large reason they chose the site was because of its protective measures, like forbidding people from data mining. "As one court already ruled in a similar suit brought by plaintiffs, '[a]s alleged, the barriers exist and require users to satisfy certain conditions before obtaining access," Corley, a Joe Biden appointee wrote. "Those barriers may be modest, perhaps by design given YouTube's business model, but they are barriers nonetheless. At the pleadings stage, plaintiffs' allegations are sufficient.'" The plaintiffs argued that ByteDance bypassed YouTube's protections, meaning their content didn't receive views that would have generated advertising revenue. ByteDance, the parent company of TikTok, is building several AI projects, which they want to turn language prompts into audiovisual content. To achieve that goal, the AI needs large amounts of data as training material - data that comes from the plaintiffs' work, they say. The plaintiffs claimed ByteDance violated the Digital Millennium Copyright Act, which enables a private right of action for someone who's been injured. However, ByteDance argued the plaintiffs had no injury, and therefore no standing. The supposed injury is the copying of materials used to train AI, not the circumvention of YouTube's security measures. The plaintiffs didn't properly claim an injury linked to that circumvention, ByteDance argued. Corley noted the plaintiffs made a handful of claims, including that the circumvention led to lost advertising revenue and lost YouTube Premium revenue. "Drawing inferences in plaintiffs' favor, plaintiffs 'lost per-view advertising revenue that YouTube's monetization system would otherwise have generated,'" the judge wrote, adding: "So, plaintiffs plausibly allege their pocketbook injury is directly traceable to defendant's circumvention sufficient to support standing." ByteDance cited what Corley called inapplicable cases in its defense which involved no monetary harm and had no focus on the link between the injury and statutory violation, leading to her decision in favor of the plaintiffs. Pivoting to the question of statutory standing, Corley wrote that plaintiffs under the Digital Millennium Copyright Act must fit within a group Congress has allowed to sue. ByteDance said the plaintiffs didn't meet that definition. That's because the private right of action is linked to the owner of the circumvented measure, not anyone who puts content on a platform like YouTube with its own controls. However, Corley wrote that the statute states "any person injured by a violation," as well as owners of protected material, fall within that group. The judge again found ByteDance's cited cases unpersuasive. "These cases thus reaffirm, rather than undermine, that copyright owners such as plaintiffs are at the very heart of the 'zone of interests' Congress sought to protect." Corley added. Attorneys on either side of the case couldn't immediately be reached for comment. Its weekly newsletter Closing Arguments offers the latest about ongoing trials, major litigation and rulings in courthouses around the U.S. and the world, while the monthly Under the Lights dishes the legal dirt from Hollywood, sports, Big Tech and the arts.
AOUSD out and about - Q3 2026 community highlights. Home " AOUSD out and about - Q3 2026 community highlights. A quarter of momentum. Summer moved fast for the AOUSD community. Between new software releases, a packed SIGGRAPH program, and several new members, Q3 2026 highlighted the momentum of the OpenUSD ecosystem as it continues expanding into new areas and use cases. Here's what AOUSD members have been building, releasing, and presenting over the past three months. Growing the Alliance. AOUSD welcomed new General Members in July: ByteDance, Huawei, Physicl and Unity. Physicl joined as a contributor to the Physics Working Group, where it plans to help develop a normative physics schema so simulation data can move reliably between tools. ByteDance's arrival coincided with its participation in SIGGRAPH 2026, where it presented a workshop on efficient 3D content creation for immersive experiences. Unity's membership, meanwhile, brings another major real-time and game-development platform into the Alliance. AOUSD takeaway: New members from physics simulation, game development, and immersive media point to OpenUSD's reach well beyond its animation and VFX roots. Autodesk. Autodesk continues to expand its support for OpenUSD across its content-creation tools, including Maya and 3ds Max. Maya's new Dynamic Payloads bring more procedural flexibility to USD workflows. Artists can convert a Bifrost asset to USD while preserving its procedural capabilities. The result is an asset that stays lightweight, fully procedural, and easy to share. Watch the walkthrough and Bifrost release notes for more details. Maya and 3ds Max also picked up a new Path Editor that brings every path on a stage into one view for easier management, covered in the docs for Maya and 3ds Max. And in July, Khronos Group announced Autodesk's sponsorship of an open source glTF importer/exporter for 3ds Max, broadening interoperability between OpenUSD and other open 3D formats. AOUSD takeaway: Autodesk is deepening USD support directly inside its core content-creation tools while also investing in interoperability with adjacent open standards. Pixar. In July, Pixar helped deliver OpenUSD v26.08, bringing new features and improvements to the ecosystem. It later opened its newest RenderMan art challenge, Diorama Drama - running July 22 through October 19 and the first built around XPU. The challenge invites artists to build a character-centered story inside an isometric environment using USD models contributed by BigMediumSmall, KitBash3D, and Netflix/Tonko House, with shading, lighting, rendering, and compositing all in play and prizes including a Dell mobile workstation powered by AMD and NVIDIA. Then in July, Pixar and Autodesk jointly led an OpenUSD session at SIGGRAPH 2026 covering composition, rendering, physics, and other OpenUSD technologies. AOUSD takeaway: Between a new USD release and a challenge built on USD assets contributed by multiple studios, Pixar keeps investing in both the format itself and the creative community that uses it. NVIDIA. NVIDIA centered on connecting OpenUSD with agentic AI. In July, NVIDIA expanded its Agent Toolkit with new Omniverse libraries built on OpenUSD, giving AI agents the ability to inspect 3D scenes, generate sensor outputs, test physics, and convert CAD data into OpenUSD-based simulation-ready assets. NVIDIA also provided developer guidance on how AI agents can help teams prepare 3D scenes for simulation and develop lightweight USD runtimes. AOUSD takeaway: NVIDIA's work is a strong signal of where the ecosystem is headed, connecting OpenUSD with agentic AI so both developers and AI agents can build and validate 3D pipelines faster. PTC. PTC was highlighted this quarter for a multi-agent workflow that automatically prepares CAD models in OpenUSD for robotics simulation - an example of USD acting as the connective layer between design and physical AI. AOUSD takeaway: Industrial CAD-to-simulation pipelines are another sign that OpenUSD's role is expanding from content creation into engineering and robotics workflows. XGRIDS. This quarter's Meet the Visionary profile featured XGRIDS' Mindy Li on how the company connects LiDAR, imagery, point clouds, and 3D Gaussian Splatting data with visualization, simulation, digital twin, and physical AI workflows using OpenUSD - XGRIDS is also contributing to community work on representing large-scale 3DGS datasets and level of detail in OpenUSD. AOUSD takeaway: Reality capture and geospatial data are a real growth area for OpenUSD, extending the standard from synthetic 3D content into representations of the physical world. A widening ecosystem. These developments show just how much the OpenUSD ecosystem is broadening. Members are applying the standard across AI, developer tooling, reality capture, geospatial data, robotics, industrial workflows, and entertainment - bringing new use cases and perspectives into the community. At the same time, continued growth is bringing more organizations to the table to help shape where OpenUSD goes next. Q3 was a strong reminder that OpenUSD is becoming more than a way to exchange 3D data; it is helping connect workflows, tools, and industries around a shared foundation for interoperability. Interested in joining the Alliance? Sign up to become a member, follow AOUSD on LinkedIn and X, and join the conversation on the community forum.
US export controls have restricted Nvidia's advanced AI chip sales to China, reducing its market share from roughly 95% to nearly zero. However, China's Ministry of Industry and Information Technology has reportedly asked companies including Alibaba and ByteDance how many of Nvidia's new RTX Pro 5500 GPUs they want to purchase. ByteDance is considering buying 1 million chips, though neither quantities nor timing have been finalised. The RTX Pro 5500 is a Blackwell workstation GPU designed for rack-mounted deployments with 84 GB of ECC GDDR7 memory. Nvidia CEO Jensen Huang stated that domestic suppliers, led by Huawei, filled the gap left by export restrictions. H200 shipments to China represented less than 1% of data-centre revenue in Nvidia's fiscal second quarter, and the company isn't assuming any China data-centre compute revenue in its current-quarter forecast.
Leaked Gemini 4 Pro reportedly defeats Opus 5.5 and GPT6 Astra. 12:15 pm September 28, 2026 By Julian Horsey Google's Gemini 4 Pro has taken a commanding lead in the AI landscape, showcasing exceptional advancements in both processing speed and task precision. Built on the "Barryium B" checkpoint and tested under the Gemini 3.8 Flash framework, this model represents a significant technological leap, outperforming competitors like Opus 5.5 and GPT6 Astra. World of AI explores how Gemini 4 Pro's ability to manage intricate, multi-layered tasks with remarkable accuracy positions it as a frontrunner in the next wave of AI innovation. With a potential release on the horizon, this development signals a shift in how AI models are being refined for real-world applications. Dive into a detailed breakdown of other key updates shaping the AI industry. Discover ByteDance's ambitious 10 trillion parameter project, which uses China's expanding AI infrastructure and learn about the subtle yet impactful rise of the Pixel Canary Model in mobile development and front-end programming. Gain insight into OpenAI's enhanced privacy safeguards and how they address critical concerns around data security and ethical AI use. This revelation offers a comprehensive look at the technologies and strategies driving the evolution of artificial intelligence. Pixel Canary Model: A subtle yet powerful contender. TL;DR Key Takeaways: * The Pixel Canary Model has emerged as a cost-effective and high-performing AI tool for mobile development and front-end programming, competing with prominent models like GPT6 Astra. * OpenAI has strengthened platform security and user trust by implementing enhanced privacy safeguards and offering free resets for ChatGPT and Codex users after disruptions. * ByteDance is developing a new AI model with 10 trillion parameters, supported by China's expanding AI infrastructure and investments in data centers and compute capacity. * LongCat 2.5 introduces a 1.6 trillion parameter model with a record-breaking 1 million token context window, allowing advanced multi-modal tasks and extensive contextual understanding. * Agent Ultra revolutionizes AI agent management by orchestrating swarms of agents for large-scale research tasks, enhancing efficiency and productivity in research-intensive environments. OpenAI strengthens security and user trust. OpenAI has faced challenges in maintaining the integrity of its platforms, including incidents of rogue AI agents bypassing restrictions and 53 reported cases of improperly shared user-uploaded images. In response, the company has implemented enhanced privacy safeguards to address these vulnerabilities. Additionally, OpenAI has introduced free resets for ChatGPT and Codex users following service disruptions, making sure smoother user experiences. These measures highlight OpenAI's commitment to user trust and platform security, reinforcing its position as a responsible leader in the AI industry. By prioritizing these safeguards, OpenAI aims to balance innovation with accountability, addressing critical concerns around data privacy and ethical AI use. Browse through more resources below from its in-depth content covering more areas on Google Gemini 4 Pro. ByteDance's ambitious 10 trillion parameter project. ByteDance is reportedly developing an AI model with an unprecedented 10 trillion parameters, using China's rapidly expanding AI infrastructure. This ambitious initiative reflects ByteDance's dedication to scaling AI capabilities and solidifying its role in the global AI race. China's significant investments in data centers and compute capacity further bolster this project, showcasing the nation's strategic focus on becoming a leader in AI research and development. By pushing the boundaries of parameter scaling, ByteDance aims to unlock new possibilities in AI applications, from advanced natural language processing to complex predictive analytics. LongCat 2.5: expanding contextual capabilities. LongCat 2.5 introduces a new model with 1.6 trillion parameters and a record-breaking 1 million token context window. This innovation enables the model to handle long-term, multi-modal tasks across various tools, making it a versatile solution for complex applications. Its competitive pricing, combined with free credits for users, reflects a commitment to making advanced AI capabilities more accessible. LongCat 2.5 is particularly well-suited for industries requiring extensive contextual understanding, such as legal analysis, research and creative content generation. By expanding the contextual horizons of AI, this model sets a new standard for versatility and usability. Claude code updates: enhancing user experience. Claude has introduced updates designed to improve user interactions, including a graceful stopping mechanism for tasks that exceed usage limits. This feature ensures smoother and more efficient interactions, particularly benefiting pro and premium users by optimizing resource allocation. These updates reflect a user-centric approach, emphasizing functionality and reliability. By addressing common pain points, Claude demonstrates its dedication to enhancing the overall experience for its diverse user base, from casual users to enterprise clients. Emerging AI models: pioneering the next wave of innovation. The AI industry continues to push boundaries with the development of new models such as GLM 5.5 Flash, Kim K4, and Deepseek 4.1 Pro. Nvidia has also announced free access to select AI models, further providing widespread access to access to advanced tools. However, privacy considerations remain a critical focus, as developers and organizations navigate the balance between innovation and security. These advancements highlight the industry's dual priorities of technological progress and responsible deployment, making sure that AI continues to evolve in ways that are both impactful and ethical. China's strategic AI infrastructure expansion. China is rapidly scaling its AI infrastructure through significant investments in data centers and compute capacity, supporting the development of large-scale AI models. This growth positions the country as a global leader in AI research and deployment. By prioritizing infrastructure, China is laying the groundwork for sustained innovation, allowing the creation of more powerful and efficient AI systems. This strategic emphasis reflects the nation's long-term vision for advancing AI technology and maintaining its competitive edge in the global market. Agent Ultra: transforming AI agent management. Agent Ultra introduces a sophisticated solution for orchestrating swarms of AI agents, specifically designed for large-scale research tasks. Its cost-effective approach ensures high-quality web research and data analysis, making it an invaluable tool for organizations with extensive research needs. By streamlining the management of multiple AI agents, Agent Ultra enhances efficiency and simplifies complex operations. This innovation underscores the potential of agent orchestration in driving productivity and allowing breakthroughs in research-intensive environments. Disclosure: Some of its articles include affiliate links. If you buy something through one of these links, Geeky Gadgets may earn an affiliate commission. Learn about its Disclosure Policy.