🔥RTN #13: a new VC meme, AI for architects + Wayve world model
The weeks tech news in physical industries
A meme is taking shape around the AI funding environment - VCs are deploying hail mary-esque sums of capital into nascent AI startups. Particularly ones, like Mistral in Paris that just raised a $113m seed round, that are building foundational language models.
The reason for this is a) many believe there is an oligopoly-like market taking shape for foundational models (similar to cloud) b) training these models costs >$200m due to GPU costs
The models need to be trained on large GPU clusters, on trillions of words. As we’ve previously covered there are only two companies that provide GPU chipsets powerful enough to do this inference, Nvidia and AMD (who this week released a number of new products and partnerships - more below).
Now, I’m not here to talk about AI. However, it’s clear that as we reach the limits of the Dennard scaling and Moore’s Law, we can only increase compute by using more GPUs and thus much more energy. GPUs are power hungry beasts. Dominion Energy in Virginia in the US is already delaying new data centre projects, as they can’t service the required power through current grid infra:
“North American utility Dominion Energy says it may not be able to meet demands for power in Ashburn, Northern Virginia, delaying building projects in the world's fastest-growing data center hub by many years.”
This problem is fairly common, though likely to worsen if the current AI trajectory remains. Energy and hardware availability are currently the boundaries for progress in much of AI.
🚖 Moving things
Wayve just teased their new Gen AI-based world model - it’s 🤯
Why it matters: we’re starting to see the development of Gen AI + LLMs being applied to real-world scenarios. In this case, the model is “imagining” a training environment based on Wayve’s data, leading to massively reduced costs for data collection and traing. Hardware can now progress as fast as software
Wayve just launched its own GenAI model, GAIA-1. The model represents a significant technical breakthrough as a generative AI model designed to serve as a world model for autonomous systems.
A world model is an AI model that predicts the future state of the world based on the actions taken. It acts as a simulated environment or a "what if" thought experiment for model-based reinforcement learning or planning
A well-developed world model functions as a learned simulator, allowing for model-based reinforcement learning and planning through comprehensive "what if" thought experiments. The potential of world models has been discussed previously, and now GAIA-1, the first Generative AI for Autonomy model, has been successfully built. Similar to ChatGPT, GAIA-1 is trained in a self-supervised manner, leveraging a vast corpus of UK urban driving data provided by Wayve AI.
The remarkable aspect of GAIA-1 is that it surpasses the capabilities of a conventional generative AI video model, as it embodies a genuine world model that learns to comprehend and disentangle critical driving concepts such as cars, trucks, buses, pedestrians, cyclists, road layouts, buildings, and traffic lights.
🌍 Policy and Geopolitics
Oof - Bit harsh. Whilst AI is getting the attention. The UK has a serious policy and investment deficit around a number of key technology areas. Such as solar, batteries, advanced manufacturing, semiconductors
🦾 Manufacturing and Robotics
Why it matters: AMD is claiming that their GPUs have lower training costs v Nvidia. Whilst they didn’t share costs for a GPU, they’re rumoured to be at $20k v Nvidia's $30k. If this stands up it might erode Nvidia’s market share
AMD released the MI300X GPU chipset and their third-gen GPUs codenamed “Antares”
The MI300X GPU accelerator consists of eight Antares GPU chiplets and boasts a high memory capacity of up to 256 GB of HBM3 memory.
The MI300A APU variant integrates Genoa CPU cores with the GPU accelerator, sacrificing some GPU compute and memory capacity but potentially achieving full bandwidth
Critically AMD is claiming their GPU have a sizeable inference advantage against Nvidia H100’s
They announced a partnership with Hugging Face, placing them closer to PyTorch and the open-source LLM community
The real winner in the AI wars is TSMC, who Nvidia and AMD rely on to manufacture chips
🌇 Built Environment
How Architects can embrace AI by Stanislas Chaillou at Rayon
Why it matters: Generative design could be one of the most impactful areas of AI, across different domains. Though will require a shift in thinking, design workflows and tooling
Architects have long used projection and semantics in the description and translation of 2D to 3D, features and analysis
NeRF, for instance, can generate complete 3D models from a few 2D views, inferring the geometry of hidden parts of an object
Other AI models operate on various spatial abstractions, such as graphs and point clouds, and contribute to the rapid advancement of spatial projective techniques
AI, with its statistical learning capabilities, offers a robust computational method for generating forms. This represents a significant epistemic shift in architecture, as AI replaces explicit rule-based approaches with machines that learn from repeated observations of architectural examples
As architects continue to adopt AI, we’re going to see scope for a whole new generation of tools, like Rayon
⚡️Energy, Materials and Climate
The IRA’s tax credit breaks by CTVC
Why it matters: The IRA is unlocking huge amounts of capital in the US to fund clean energy and infrastructure spending. This guide helps founders better understand the funding structure
Of the much anticipated $1.7T public and private IRA climate spend provisions, tax credits alone are predicted to make up over 2/3rds of all spending
Tax equity financing involves using tax credits as incentives for (often non-financial) investors to offset their tax liabilities. Third-party tax equity investors can partner with developers who don't have enough taxable income to take advantage of the credits
There are two main types of tax credits: production tax credits (PTC) and investment tax credits (ITC)
Tax equity investors cover a significant portion of solar and wind project costs in exchange for tax credit incentives and cash returns once the project is operational.
The tax credits are:
Production tax credits (PTC): Revenue for investors earned as a per-kWh tax credit based on the production of energy generated and sold, paid over a 10-year period
Investment tax credits (ITC): Covers costs to get projects up and running by allowing taxpayers to deduct a percentage of the project’s total cost from their tax liability, provided as a one-time credit once operational
Depreciation benefits: Investors receive a tax deduction each year for any normal wear-and-tear or decrease in the value of assets related to a project. Some renewable projects qualify for 100% depreciation the year they begin operating
Recapture risk: There is a risk that the tax equity investor must repay a portion of the tax benefits if projects that elected ITCs are taken out of service or sold during the five-year recapture period.