Log inSign up
user avatar
Lennart Heim
@ohlennart
managing the flop | prev @RANDcorporation @GovAIOrg @EpochAIResearch
Washington, DC
heim.xyz
Joined February 2013
754
Following
15.3K
Followers
  • Pinned
    user avatar
    Lennart Heim
    @ohlennart
    Apr 29, 2025
    China's AI models are closing the gap—and will continue to improve. However, this misses America's strategic compute advantage. In my new commentary, I argue that the TOTAL compute advantage is what export controls preserve and—if leveraged correctly—provides the real edge. 1/
    108K
  • user avatar
    Lennart Heim
    @ohlennart
    Mar 11, 2025
    Huawei's next AI accelerator—the Ascend 910C—is entering production. It's China's best AI chip. Thanks to backdoor sourcing, we could easily see 1M H100-equiv this year. Here’s what we know about its performance and strategic implications. Spoiler: selectively competitive. 1/
    386K
  • user avatar
    Lennart Heim
    @ohlennart
    Feb 15, 2022
    **ML training compute has been doubling every 6 months since 2010!** Our preprint "Compute Trends Across Three Eras of Machine Learning" is out. arxiv.org/abs/2202.05924 🧵 Thread below ↓ 1/
  • user avatar
    Lennart Heim
    @ohlennart
    Jan 29, 2025
    Here, now you hear it from a US company itself: it turns out they also had compute efficiency improvements but did not tout it to the world.
    user avatar
    Dario Amodei
    Anthropic
    @DarioAmodei
    Jan 29, 2025
    My thoughts on China, export controls and two possible futures darioamodei.com/on-deepseek-an…
    146K
  • user avatar
    Lennart Heim
    @ohlennart
    May 12, 2022
    🐈Gato (@DeepMind) is impressive. Single transformer with the same weights for multiple tasks. It'd only cost you around $50K to train it in the GCloud. That's nothing against the $11M+ for PaLM. We could scale Gato 100x with a PaLM compute budget and scaling probably works. 🧵⬇️
  • user avatar
    Lennart Heim
    @ohlennart
    May 21, 2025
    Yes, we do. It's ~21GW. You count all the AI chips produced, factor in that they're running most of the time, add some overhead—and you got your answer. It's a lot. And will only get more. But you know what? Probably worth it.
    29K
  • user avatar
    Lennart Heim
    @ohlennart
    Dec 23, 2024
    To get o3's strong performance on ARC-AGI for a single task in 10min, you'd need 10,000 H100s. Even the high-efficiency version would need >100 H100s for a 5min response per task. That's quite a wait and not a small cluster. Welcome to industrial-scale thinking!
    84K
  • user avatar
    Lennart Heim
    @ohlennart
    May 15, 2025
    To put the new 5GW AI campus in Abu Dhabi (UAE) into perspective. It would support up to 2.5 million NVIDIA B200s. That's bigger than all other major AI infrastructure announcements we've seen so far.
    109K
  • user avatar
    Lennart Heim
    @ohlennart
    Mar 1, 2023
    Here's a visualization of the US export controls on chips. Explicitly designed to target chips with high performance and interconnect (chip to chip) bandwidth - as commonly used in high-performance clusters. Purposely designed to not target Gaming GPUs and consumer hardware.
    79K
  • user avatar
    Lennart Heim
    @ohlennart
    May 27, 2025
    My team at RAND is hiring! Technical analysis for AI policy is desperately needed. Particularly keen on ML engineers and semiconductor experts eager to shape AI policy. Also seeking excellent generalists excited to join our fast-paced, impact-oriented team. Links below.
    70K
  • user avatar
    Lennart Heim
    @ohlennart
    Oct 17, 2023
    The US just published its revised export controls on AI chips, moving away from the 'chip-to-chip' interconnect bandwidth threshold to a threshold on computational performance (OP/s), including its derived performance density (OP/s per mm²). 1/
    68K
  • user avatar
    Lennart Heim
    @ohlennart
    Jul 21, 2024
    Looks familiar? The new GPT-4o mini uses "instruction hierarchy" to combat prompt injection attacks, applying lessons we've learned from operating system architectures to LLMs—but instead of well-defined boundaries, we're training these principles into the systems. 1/
    24K
  • user avatar
    Lennart Heim
    @ohlennart
    Aug 19, 2025
    The speculated B30A would be a really good chip. “50% off” is false reassurance. -½ B300 performance, ½ price = same value (just buy 2x) -Well above (12x!) export control thresholds -Outperforms all Chinese chips -Delivers 12.6x the training perf of the H20 -Better than H100 1/
    71K
  • user avatar
    Lennart Heim
    @ohlennart
    Jan 13, 2025
    The key principle of the diffusion framework: build your AI infrastructure in the US and partnered nations. This reflects reality—most AI compute is already here, and if deployed globally it's mostly US companies. Access follows clear pathways based on company HQ and location 1/
    72K