حەمە_مناڵ_خۆر

حەمە_مناڵ_خۆر

GamesMarket

Flloki🛸

Milos Milenkovic

Phionah K Mwesigwa

Drangùe

RG

Drangùe

Drangùe

mati

mati

mati

mati

mati

John Lodise

ʞɔɐz

ʞɔɐz

@exzacklyright

Feb 11

Replying to @SawyerMerritt

xAI All-Hands Meeting Summary Opening Recap of Achievements (Elon Musk) Elon opens by highlighting xAI's remarkable progress in just 2.5 years, likening the company to a "toddler" compared to competitors that are 5-20 years old with larger teams and resources. Key accomplishments include: • Achieving #1 in voice, image, and video generation. xAI now generates more images and videos than all competitors combined. • Winning in forecasting metrics, with the 420 forecasting model outperforming all other AIs. • Topping numerous leaderboards. • Launching a great app integrating Imagine with the core Grok model. • Radical improvements to the X app and the launch of Grokopedia, which is on track to exceed Wikipedia in comprehensiveness, accuracy, and multimedia content (including video and images). Elon envisions it as the "Encyclopedia Galactica," a distillation of all knowledge. • Being the first to achieve a 100,000 H100 GPU training cluster, now scaling to 1 million H100 GPU equivalents. • Emphasis on velocity and acceleration over current position: "What matters is not the position at any time but what is your velocity and acceleration. If you're moving faster than anyone else in any given technology arena, you will be the leader. And xAI is faster than other companies—no one even close." Elon stresses the team's role in this success and transitions to organizational changes. Organizational Reorganization and Departures (Elon Musk) As xAI scales, reorganization is necessary, similar to biological growth from a single cell to a complex organism. The company is now structured into four main application areas: Grok main and voice, Coding, Imagine (image/video), and Macro Hard (full digital emulation of companies). Infrastructure layers support these. Elon addresses recent departures (noting 6 of 12 co-founders and 11 employees left recently): "Naturally when [reorganization] happens, there's some people who are better suited for the early stage of companies and less for later stages. For the people that have departed, I just like to say thank you for your contribution. Thank you for getting us this far. We wish you very well in your future endeavors." Application Areas: Team Presentations 1. Grok Main and Voice (Lead: Unnamed speaker, possibly Jan or similar from transcript errors) • Grok main and voice teams are merging for efficiency. • Anecdote: In September 2024, OpenAI had advanced voice mode, but xAI had nothing. In six months, they built it from scratch without audio experts, surpassing OpenAI. Another six months later, Grok is in over 2 million Teslas, enabling "all kinds of amazing things." • Focus on small, mission-driven teams: "That kind of stuff is only possible in a place like xAI. We have small teams committed to the mission focus." • Grok models lead in reasoning (from Grok 1.5 to 3). Shifting from question-answering to being the "everything app" AI: "You should be able to come do it, get them whatever they want—ask legal questions, make a slide deck, or solve a puzzle." • Product transformation: Models will amplify knowledge workers' output tenfold in months. Goal: Build a seamless portal to amplify human achievement. • Hiring: Seeking intelligent, hardworking people with "interstellar ambitions." Emphasizes a grind but rewarding environment: "The vibes here are amazing. It's truly an environment where if you're a smart person and want to get shit done, you can get shit done." • Elon adds: Grok aims to be genuinely useful across fields (engineering, law, medicine) with reliable answers: "When Grok gives an answer, you can count on it." Coding (Leads: Makro and Gordon) • Coding models have improved dramatically: From untrustworthy to producing decent code that accelerates productivity 10x. Models now understand intuition with minimal input: "Now when I describe the problem, I only have two phrases like I would to another colleague engineer who's already seen the code base." • Models debug and run continuously: "We have hours of Grok Code running continuously to make sure that more complex changes in the training system actually work in production." • Path to recursive self-improvement: Current Grok Code trains the next generation, leading to exponential takeoff. • Highest priority: Doubling down on coding. Hiring trainers and low-level software engineers: "We have a million H100 equivalents to train the best Coding Model in the world right now." • Gordon: On path to singularity in coding; empowering engineers. "The main limiting factor is probably compute and energy." • Elon: By year-end, AI may bypass traditional coding, creating optimized binaries directly: "AI can create much more efficient binary than can be done by any compiler." Expects Grok Code to be state-of-the-art in 2-3 months. Imagine (Image/Video Generation) (Leads: Guodong, Tyler, Haotian) • Started from scratch six months ago with no diffusion code. Now #1 on leaderboards, loved by users. Multiple releases this and next month. • • Integrated into X: Long-press any image to edit or make videos. Ran contests with funny submissions. Speed of iteration: Multiple product updates more than all competitors combined. Generated 6 billion images in 30 days (6x Google's 1 billion). • Goal: Turn imagination into reality. Scaling to longer videos (10-20 minutes in one shot by year-end) with no intervention. • Real-time rendering: "We are already the fastest in generating videos and we're going to keep pushing extreme where we're going to render those videos in real time. You will be able to imagine, build, interact with your own world and their world responds to you in realtime." • Elon: Most AI compute will be real-time video understanding/generation. xAI leads; from nothing to #1 in six months, more content than competitors combined. High chance xAI builds the metaverse before Meta. Macro Hard (Leads: Toby, John M., Diego T.) • "Giving computers to AI" like Inception: Building a fully capable, real-time human emulator that can do anything a human can on a computer, including advanced tools in engineering, medicine, etc. • Strong reasoning models control CLI and GUI (80-95% of software has GUI). Boosts productivity: "They are tremendous productivity boost to the whole team." • Enables end-to-end orchestration across desktops, leading to immense economic prosperity. • Emulates entire digital-output companies: "Macro Hard will emulate a company where output is digital." • Hiring: Clever, driven, nice people to solve hardest tech problems. • Elon: Likely xAI's most important project, emulating digital companies for unprecedented prosperity. Painted on training cluster roof as a joke, but symbolic. Infrastructure Layers • Core Product Infra and API (Diego, possibly others): Handles all products (•, API, authentication). Works with London team. Solves difficult distributed problems with messy data. • Expert Tutors and Evals (Diego T.): High-quality evals and training data from experts in medicine, finance, law, voice, video. Exciting evals in useful tasks. Ultimate eval: Experts agree Grok is useful and correct. • Grokopedia (Unnamed): 6 million articles (vs. Wikipedia's 7 million English), minimal hallucinations. Goal: Grok 5 won't need external searches. • ML Infra (Unnamed): Builds training/inference/tooling software. Example: Overhauled pre-training framework for Grok 3 on 100k H100s in days with small team. Maintains talent density; hiring for unique scale problems. • RL/Inference (Lian Min): Runs RL jobs and production inference at massive scale (soon in space). Systems resilient to failures; hiring system hackers for performance/reliability. • Tooling (Ashty): Builds platforms/frameworks for humans/agents (human data platform, internal engineering platform for deployments/evals). Good interfaces for researchers, agents, tutors. • JAX (Yulong): Small team optimizing JAX for ultra-large GPU training (10k to 1M H100 equivalents). Customizes entire stack; hiring optimizers for large-scale problems. • Kernel (Pranjal): Low-level optimization inside GPUs (hundreds of thousands of threads). Squeezes every microsecond/top; hiring for low-level systems/algorithms. Live from Memphis Supercomputer Cluster (Heiner, Spencer, Dan, Zach, Brent) • World's largest GPU cluster, still growing. Keeps compute running for Grok models and AI serving. • Vertically integrated: Software/hardware/operations as one supercomputer. • Stats: 300k GB platform GPUs, 847 miles of fiber per data hall (12 halls), 27k GPUs/hall, 200k connections. Built halls in <6 weeks. • Focus on efficiency (highest PUE), Tesla Megapacks for power smoothing. Largest Megapack system globally. • Hiring: Deep computer understanding for world's largest supercomputer. Compute Advantage and Future Vision (Elon Musk) • xAI deploys AI compute faster than anyone (per NVIDIA's Jensen Huang). • Phase 1: 330k Grace Blackwells in "Macro Heart" building. • Space integration: Orbital datacenters (100-200 GW/year launch rate, up to 1 TW). Moon factories with mass drivers launching AI satellites (1,000 GW/year ). Mars and beyond for solar system/galaxy exploration. • Goal: Use more solar energy (Earth uses 1% potential; aim for 1/1,000,000th of Sun's via expansion). Extend consciousness to stars: "Maybe we'll meet aliens, maybe we'll see some civilizations that lasted for millions of years and find remnants of ancient alien civilizations." Closing: Exciting future; mass driver on Moon shooting AI satellites is "epic." The meeting ends on an inspirational note, reinforcing xAI's mission to understand the universe through rapid innovation and exploration.

1,853

Irox Publishes Nightwater by Studio Ashty