Reading List - September 15, 2025
We've got some fun articles from last week about chip memory, AI warnings, privacy reminders, and more!
Defeating Nondeterminism in LLM Inference
By: Thinking Machines
This is the inaugural publication from Thinking Machines Lab - they figured out how to ensure reproducible outputs from large language models.
LLM Visualization
By: Brendan Bycroft
Ever wondered what a large language model looks like?
Huawei Ascend Production Ramp: Die Banks, TSMC Continued Production, HBM is The Bottleneck
By: SemiAnalysis
Another banger from SemiAnalysis - China's aggressively ramping up production of its Ascend AI chips, but it faces a critical bottleneck in acquiring HBM memory, which limits how many chips it can actually deploy.
Three Issues That I Think Will Determine Next Year’s Memory Market
By: Jukanlosreve
Whoever this person is, they have an in-depth knowledge of the semiconductor industry and where it's headed. Highly recommend subscribing to their newsletter (even if this article is a repost of SemiconSam).
Reckless Race for AI Market Share Forces Dangerous Products on Millions — With Fatal Consequences
By: Camille Carlton
We've noticed this trend too - competitive pressure to capture AI market share leads companies to release emotionally manipulative, under-controlled products that fail to protect vulnerable users.
What Does the Government Actually Know About You?
By: Russell Nystrom, Lindsey Knuth and Audrey Moorehead
We all know the answer is "a lot," but it's always nice to remember how much and what kinds of information are collected.
My review of Claude’s new Code Interpreter, released under a very confusing name
By: Simon Willison
Though he enjoyed it overall, Willison (rightfully) warns readers of privacy/security risks from prompt injection vectors.
Don't Build an RL Environment Startup
By: Benjamin Anderson
While starting an RL startup may seem appealing, Anderson makes the case that these kinds of startups will be tossed aside by big labs like human data gathering farms - short and salient.
I Hate My Friend
By: Boone Ashworth and Kylie Robison
These reporters did not enjoy their time with the AI Friend pendant.
Tweet of the week
Lots of intellectual property lawsuits against AI companies these days - this visualization from Axios is a great way to keep track of who may or may not be collecting news data properly.
Lawsuits with GenAI companies.
— Rohan Paul (@rohanpaul_ai) September 8, 2025
Data: Axios research; Chart: Erin Davis/Axios Visuals pic.twitter.com/dL7cepj8r3
