Infrastructure Accounting Lead

MIT phd Li Muyang launches Nunchux AI to accelerate multimodal inference. Nunchux AI has officially launched as a new player in the generative AI infrastructure space, founded by MIT doctoral graduate Li Muyang, Carnegie Mellon University associate professor Zhu Junyan, and researchers Yujun Lin and Zhekai Zhang. The company targets the growing industry demand for efficient, low-cost, and reliable multimodal inference as deployment scales rapidly. Backed by early-stage investments from Emergence Capital and E14 Fund, MIT Media Lab, Nunchux AI builds upon a decade of research into neural network compression and system optimization. Li's academic work, particularly the SVDQuant framework for 4-bit weight and activation quantization, forms the technical foundation of the venture. This research materialized in Nunchaku, an open-source inference engine that drastically reduces memory footprint and latency, enabling high-resolution image generation on consumer hardware. The engine has secured significant developer adoption, recording nearly 4,000 GitHub stars and native compatibility with major open-source models and development frameworks. Commercially, the startup will debut Modelverse, a unified API platform integrating over thirty image, video, and virtual avatar models. The service offers two operational tiers: Radical Speed for ultra-low latency and Radical Value for cost optimization. Unlike existing inference aggregators that primarily manage GPU orchestration, Nunchux AI pursues vertical integration by combining custom quantization algorithms with optimized inference kernels. This approach delivers superior performance and lower overhead on identical hardware. The Nunchaku core will remain open-source, while commercial revenue will be driven by premium API access, enterprise-grade customization, and dedicated developer support. The venture extends the entrepreneurial legacy of MIT professor Han Song's HAN Lab, which previously spawned companies such as DeePhi Tech, OmniML, Eigen AI, and Inco AI. Academic production continues in parallel, with recent publications including the FourTune paper explicitly listing the company as an affiliation. In a crowded inference market featuring competitors like fal.ai and Replicate, Nunchux AI differentiates itself through algorithmic efficiency rather than mere aggregation. The founding team maintains that reducing latency and deployment costs is critical for preserving creative velocity and democratizing access to advanced generative models, while strictly preserving output fidelity. This news is intelligently aggregated by AI to deliver industry updates efficiently. It does not constitute opinions or advice. MIT Technology Review

Fal, a generative media platform, has launched H3 Max, a new video generation model that ranks first on independent benchmarks from Artificial Analysis and Design Arena. The model generates five-second videos in approximately three seconds, achieving faster-than-real-time generation. Built on the open-weights MiniMax H3 model, H3 Max was developed through post-training for improved prompt adherence and visual quality. Fal reports the system delivers roughly 35 times the throughput of the official MiniMax H3 endpoint and averages 15 times faster than comparable quality models. H3 Max is available through the fal API and Playground. During a promotional launch period ending 7 September, pricing starts at $0.04 per second at 768p resolution, rising to $0.08 per second thereafter. Fal developed the model by co-designing training and inference optimisation, treating them as a single problem rather than separate layers.

Lore issue #200: OpenAI's first AI chip beats Nvidia's GB300 on efficiency. PLUS: Nvidia nears a $13B Hugging Face deal, Sam Altman says AGI could arrive this year, and Hugging Face unveils a $399 open-source robot Aug 28, 2026 Good morning, welcome to this week's Lore Brief, your 3-minute brief of the most important moves in AI and tech. This issue is brought to you by Factory, the fastest way to ship software with autonomous engineering agents. * OpenAI's Jalapeño chip beats Nvidia GB300 on efficiency | First published results show the custom inference chip delivering 1.5 to 1.9 times more work per watt than Nvidia's GB300 while cutting end-to-end latency by as much as 3.6 times. The 700-watt part goes up against a 1,400-watt flagship and still returns answers faster on open models like DeepSeek R1 and Kimi K2.5. Deployment inside OpenAI's own infrastructure is planned by year-end. Read more here | * fal's H3 Max claims the top spot in video generation | The post-trained model ranked first for quality prompt understanding and aesthetics against a dozen leading systems. It can produce a 5-second 720p clip in under three seconds and is 50 percent off this week. fal built it on MiniMax's open H3 base then tuned it for speed without giving up looks. Read more here | * Nvidia closes in on a Hugging Face acquisition | Talks point to a deal around $13 billion that would give the chipmaker a major foothold in open-source AI. Hugging Face last raised at $4.5 billion in 2023 and now does roughly $150 million in annual revenue. The move would help Nvidia stay central as closed labs build their own chips. Read more here Lip Sync tutorial for AI videos: The same style ad like the one of the new Mac mini but made with AI: * OpenAI leaders say AGI is getting close | Sam Altman told TIME the company could have an internal system that qualifies as AGI by the end of 2026. Mark Chen put the lab at about 80 percent of the way there. Astra already acts as an automated research intern that can run week-long experiments inside OpenAI's own codebase. Read more here | * A practical field guide to living with Grok Bot | Two weeks after launch Matt Van Horn published every hack he has found from giving a bot its own inbox to letting it place phone calls in Portuguese. The write-up covers planning layers named roles cookie sync and the habit of drafting before anything gets sent. The honest caveat is that unsupervised crews can multiply their own errors fast. Read more here | * TIME drops its 2026 list of AI's most influential people | The annual TIME100 AI ranking is out again with the usual mix of lab chiefs and public figures. Online reaction quickly turned to the oddities including Paris Hilton while Jensen Huang is missing from the list. Readers treated that combination as the punchline. Read more here | * Claude memory now follows you from chat into Cowork | One shared memory means a task in Cowork can start from what you already discussed in chat including projects preferences and past clients. You can read edit or delete every saved topic in Settings. Sensitive subjects stay out unless you turn them on. Read more here | * Hugging Face unveils a $399 open-source robot | Microduck can walk pick things up get back up after falling and even roller-skate. Users teach it new tricks with reinforcement learning instead of waiting for a closed lab to ship a firmware update. Clem Delangue frames it as affordable hardware for physical AI and world models. Read more here | * Skild AI's S1 learns 10-minute robot tasks from one video | Show the foundation model a single demonstration and it can complete jobs it never saw in training from making coffee to flipping pancakes. No fine-tuning is required and it can recover from mistakes the human in the video never made. The company says matching that accuracy with older VLAs would take 50 to 100 hours of extra data. Read more here | * Google ships Gemini 3.5 Transcribe for cleaner live speech | The new model turns messy audio into formatted text while cutting filler words and handling self-corrections like "Tuesday no Wednesday." It recognizes more than 85 languages and drops word error rates well below earlier Chirp systems. Developers can use it now through the Gemini API and it is already landing in the macOS Gemini app. Read more here That's it for this week's Lore Brief. See you next week!