与「FoundationModels」相关的搜索结果

Can robot foundation models scale? Xiaomi-Robotics-1 explores this question with over 100,000 hours of real-world manipulation data. Pre-trained on large-scale real-world trajectories and post-trained with cross-embodiment robot data, Xiaomi-Robotics-1 demonstrates consistent scaling across data and model size, strong generalization in unseen environments, and efficient adaptation to new tasks. 🔗 #Robotics# #EmbodiedAI# #FoundationModels# #RobotLearning# #XiaomiAI# #XiaomiRobotics#

Introducing Edge0-145B and Edge0-120B, @Edge0_AI’s first device-native 100B-class foundation models with open weights. Cloud-level intelligence on your laptop. Free, offline, and fast enough for real vibe coding. - Smart: beats Claude Opus 4.6 on Terminal-Bench 2.1 (86.5% vs. 78.2%) - Light: needs just 29G memory, 1/3 of a standard setup (100G+) - Fast: decodes 75+ toks/s, prefills 1050+ toks/s on RTX Spark and Mac. - Steady: handles 128K–256K tokens of context without slowing down, even on long tasks Frontier-class agentic coding. No cloud API. No token bill. Your code and data never leave your machine. Open weights are live on @huggingface. Run it locally or plug it into your favorite coding agent.

What Menlo has done over the last 24 months is exceptional. Yes, they did Anthropic’s round at $4BN that will make them $50BN+ as a firm. NICE! But they then took that and went f****** hard. They do Lovable, Higgsfield, Legora. They compounded the brand win of Anthropic to cement their position from respectfully a slightly ageing firm to being a dominant venture leader at the forefront with Thrive, a16z, Benchmark etc. It is incredible to see and I sat down with @venkyganesan to shoot the s*** on this. 1. Every Seed Investment Is an Option Bet Seed-stage deals function as financial options to discover true outlier potential before committing major capital. VCs must construct a broad enough portfolio to secure sufficient at-bats, only sizing up their position when quantifiable, revenue-based metrics prove an outlier has emerged. 2. What Is the First Sign of the AI Wave Cracking? Market cycles rarely collapse due to equity write-downs. They fracture when levered debt defaults occur. When market participants take on heavy leverage during an AI boom, precise timing becomes as critical as being directionally correct, making debt obligations the primary failure vector when cycles turn. 3. Why Every VC Must Now Prioritize IRR Every startup in the current landscape writes a structural tax to the Mag 7 and hyperscalers for compute and foundation models. Because LPs can buy index exposure to those same public tech giants without fees or carry, private VCs must generate outsized IRR performance, beating public markets by at least 1,000 basis points, to justify private capital allocation. 4. Why LPs Wanting Smaller Funds and Slower Deployment Is BS LPs cannot demand both smaller fund sizes and multi-year deployment paces in the middle of a generational platform shift. Because AI startups face immense capital requirements to scale compute, VC firms that refuse to deploy capital rapidly will watch competitors step in and capture the category winners. 5. The Theory of Reflexivity and How It's Causing Markups Reflexivity creates self-reinforcing loops where fast revenue growth triggers rapid valuation markups, attracting elite human and financial capital that further accelerates growth. However, when copycat investors chase markups for momentum rather than underlying fundamentals, reflexive valuation bubbles inflate until market realities force an inevitable correction. 6. Sometimes You Pay Up Because You Have Unique Insight Premium entry prices are not always the result of irrational bidding. They often stem from an investor's ability to recognize a fundamentally larger total addressable market than the consensus. When a VC possesses unique conviction around an expanding market opportunity, paying above current rates reflects calculated foresight rather than price-taking behavior. 7. There's No Limit to What a Person Can Do as Long as They Don't Care Who Gets the Credit Personal ego remains one of the most expensive liabilities in venture capital and company building. True long-term success requires setting aside individual notoriety, focusing ruthlessly on generating LP returns, and empowering teammates and founders without worrying about who receives the public credit. (links in comments)

Meta now owns the entire AI stack. 1. Chips: MTIA 2. Infrastructure: Meta Compute 3. Foundation Models: Muse Spark 4. Applications: Muse & Meta AI Every layer matters. The more layers you own, the bigger your advantage.

From Digital Twins to Data Engine: Cutting the Real-Data Burden with Sim-Powered Robot Learning Teaching a robot a new task could take hundreds of teleoperated demonstrations. For foundation models, adapting to an entirely new robot can cost orders of magnitude more — dedicated hardware, trained operators, months of engineering. On @boosterobotics' dual-arm robot, we studied this at two levels: ✱ Specialist: Can task-aligned simulation mixed with a small set of real demonstrations reduce the real-data burden? ✅ Yes. With only 10 real demos, the policy made no contact at all in physical rollouts (0/20). Adding 50 simulated trajectories brought contact to 17/20. ✱ Foundation: Can data accumulated across tasks build a reusable starting point (a Booster-specific model prior)? ✅ Yes. After full-parameter continued pretraining, a model adapted with just 30 demonstrations per task beat the original given twice as many: 14/16 vs 10/16 in simulated evaluation. Before any task-specific adaptation, in zero-shot simulation, it was already roughly 3× closer to the target (17.27 cm → 5.78 cm). This work runs on Axis Suite, our Physical AI solution across different robot embodiments. Distributed contributors generate task-aligned sim data on Axis Hub at scale, reducing real-data needs for specialist adaptation while powering cross-embodiment generalist training. Read the full blog:

XPENG’s robotics business has completed a 900M+ Series A funding round at a post-money valuation exceeding $6.3B, setting a new fundraising record for China’s embodied AI industry. The round was led by IDG Capital, joined by Gaorong Ventures, and strategically backed by Tencent and Alibaba, with XPENG retaining controlling ownership. The capital will accelerate full-stack hardware/software R&D, physical AI training, smart manufacturing infrastructure, and global market expansion. Backed by 12 years of in-house tech innovation—from physical foundation models to Turing AI chips and robust AI infrastructure—XPENG is uniquely positioned to lead the scale-up and commercial rollout of high-end humanoid robots. XPENG IRON is slated for mass production in late 2026, launching first in XPENG retail stores and campuses before rolling out globally with customer deliveries in 2027. $XPEV

Video foundation models (e.g., Seedance 1.0 → 1.5 → 2.0; V-JEPA 1.0 → 2.0 → 2.1) keep getting stronger, while public descriptions of how their training data is built keep getting shorter. ⚒️ We built VidaForge, an open-source, five-stage video data pipeline that turns raw video collections into training-ready datasets. In an academic lab, studying video data recipes begins with a lot of tedious engineering: handling broken videos and transcode failures, keeping large jobs resumable, tracking every clip, and packaging the result for training. Across ingestion, segmentation, selection, annotation, and training dataset packaging, data moves through: raw videos → standardized videos → clips → curated clips → annotated clips → training datasets At every step, we can inspect what happened to each video or clip. This lets us see which samples changed under a data recipe and trace a training dataset back through the pipeline. VidaForge currently connects processed data to two video foundation model pretraining paths: Wan video generation through NeMo-AutoModel, and self-supervised video representation learning through the official V-JEPA2 repository. We ran VidaForge end to end on 200K videos, producing over 700K clips. From these clips, we built Selected-200K, Mixed-200K, and Rejected-200K datasets for Wan2.1-1.3B and V-JEPA2.1-1B pretraining. In these early runs, the three datasets produced different training behavior: data selection appeared in Wan2.1-1.3B eval curves and V-JEPA2.1-1B training stability. Pipeline, experiments, open data, and project resources in the thread below ↓

Apple runs its own AI models, its own silicon, its own map platform. And it still needs a guy in a red polo walking Malaysian sidewalks with a $50,000 LIDAR backpack strapped to his shoulders. Five states in Malaysia. Pedestrian-only zones where the Apple Maps cars can't drive. Areas around markets, courtyards, mosques where the roads stop but the sidewalks keep going. Somebody has to physically walk it. What's actually on the backpack. A 360 spherical camera on the top pole. A LIDAR unit for depth. GPS at the base. Combined cost of the rig runs around $50,000 in sensors. Contractors get hired through outsourced recruiting sites, not the Apple careers page. Apple doesn't want its logo on this part. This is the layer at the bottom of the AI stack that nobody writes threads about. Foundation models don't wake up knowing Kuala Lumpur's alleys. Somebody feeds them the alleys. That somebody is a guy in a red polo stacking bag as an Apple contractor while everyone else fights over the newest Claude release. There's a whole layer of AI-adjacent jobs that pay quietly while the discourse stays loud somewhere else. Pedestrian mapping. RLHF review. Data annotation. Voice recording for TTS training. And one more that most of AI-Twitter would refuse to admit even exists.

Anthropic 更新了 API 文档，提供了 Foundation Models 的 SDK。 感觉中文圈因为没法用上 Siri AI ，所以对他讨论不多。 事实上， Apple 在 WWDC 2026 给出的 AI 方案，协调了 - iPhone 硬件销售需求（根基） - 模型厂商需求（上游） - 第三方 App 开发者 - 消费者体验 的利益，让大家都能在 Apple 生态中获取自己所需。 Apple 的给的答卷，很正确，很务实，也很 John Ternus。