The hottest AI Systems Substack posts right now

Hardware drives modern deep learning: algorithms explain maybe 40% of progress and the rest comes from the compute, memory, and system-level engineering that makes training and inference practical.
GPUs were a lucky fit for neural nets because their high arithmetic density matched the workload, but custom AI chips are needed to close remaining gaps by optimizing dataflow, precision, and memory access.
Designing an AI chip is a layered engineering craft from architecture to physics and tape‑out, involving RTL/Verilog work, hardware–software co‑design, and careful trade‑offs across performance, power, and manufacturability.

The current LLM trend is to scale models huge and use sparsity tricks like Mixture-of-Experts so only a small part of the model activates per token, reducing FLOPs.
Reusing an old technique — storing large, static lookup-like memories on CPU RAM and conditionally accessing them — can let models hold around 100B parameters off-GPU and avoid expensive dense computation.
The key insight is that many LLM costs come from simulating static lookup tables with neural computation, so replacing that simulation with real conditional lookups makes models much more efficient.

Autonomous weapons systems are rapidly developing, especially after the Russia-Ukraine war, with countries learning from real battlefield experiences. Bigger nations like the US and China may soon engage in a 'drone wars' cold war using these technologies.
There are phases of evolution for these systems. It starts with semi-autonomous units, progresses to more independent operations, and eventually leads to fully integrated battle networks where AI makes most tactical decisions.
By 2030, the use of autonomous weapons will be widespread, making human combatants less effective on the battlefield. New strategies will focus on mass deploying these systems and using advanced AI for decision making.

One approach is to ensure alignment research stays ahead of AI capabilities to prevent issues, which could involve slowing down capabilities research or dedicating compute to alignment research.
Finding a comprehensive once-and-for-all solution to the alignment problem is crucial for ensuring all future AI systems are aligned, but it remains uncertain if this is possible.
Developing formal theories for alignment, creating processes to elicit values inclusively and fairly, and training AI systems to be fully aligned are key components that require significant effort and progress in the field.

The US and China are racing to develop their own advanced AI systems. This competition is seen as crucial for future global power dynamics.
The banning of TikTok in the US reflects a growing belief that it poses a threat as a Chinese intelligence tool, highlighting rising tensions between the two nations.
There's a shift happening towards two separate technological worlds, with each side training their AIs to align with their own cultural values and ideologies.

Get a weekly roundup of the best Substack posts, by hacker news affinity:

A perfect model can create high-quality data to build strong AI, like AlphaZero - AIZero
Without a perfect model, gathering high-quality data is essential for competent AI - AI∞ or AIx
It is important to start AI systems with ground truth data and work towards bridging the gap between simulation and reality

AI can really help with organizing and prioritizing tasks in many areas like customer support and fraud detection. This means faster and more efficient decision-making for businesses.
Cloud software companies like Amazon, Microsoft, and Google are seeing some slower growth lately. It's important to keep an eye on how they perform in future reports.
The value of a software company is often based on its revenue, especially when it's not profitable yet. Understanding these valuation methods can help investors make smarter choices.

RoboNet introduces a new Internet protocol aimed at regulating AI content and systems for transparency and accountability.
RoboNet enables a clear distinction between AI-generated and human-created content on the Internet, empowering users to make informed choices.
The protocol shifts the burden of content classification from Internet Service Providers to a common technical environment, promoting fairness, transparency, and accountability online.

Perceived possibilities shape perceived options, interests, and goals, and recognizing new possibilities can lead to better outcomes.
AI has the potential to create and destroy options, impacting the interests of various entities, and understanding this could align goals.
To benefit from AI while reducing risks, there needs to be a better understanding of safe, highly capable AI systems and their potential impact on society.

Automating processes requires trust and relinquishing some direct control.
Machine learning verification involves creation-time and post-deployment steps for ensuring model performance and reliability.
Artificial intelligence must undergo fact verification, coherency, consistency, and quality assessment to ensure reliable outputs.