Computer scientist LaurieWired identifies a sixty-year-old architectural limitation in DRAM where memory must periodically "blind" itself to recharge leaky capacitors, causing significant latency spikes known as refresh stalls.

While these nanosecond delays are unnoticeable to average users, they create critical performance bottlenecks in high-stakes fields like high-frequency trading. To address this, the creator developed Tail Slayer, a software-based mitigation strategy that utilizes hedged reads to request data simultaneously across multiple memory channels. By reverse-engineering undocumented CPU hashing functions and employing huge pages, the technique ensures that if one memory controller is busy refreshing, a duplicate request on an independent channel can fulfill the task immediately. Experimental results across Intel, AMD, and ARM architectures demonstrate that this method can reduce tail latency by up to fifteen times, achieving near-deterministic performance.

In this episode, we look at a retired coding test from Anthropic that used to be a secret way to find top performance engineers. The company decided to make the test public after their AI, Claude Opus 4.5, started beating human candidates in under two hours.

We discuss the main benchmarks to beat, starting with the AI's record of 1487 cycles (and an even faster 1363 cycles in a special setup). We also talk about the mysterious human record — a secret score that is "substantially better" than the AI, though the exact number is hidden as "??? cycles". You can find the code and try the challenge yourself here: https://github.com/anthropics/original_performance_takehome.

We also cover why you have to be careful if you use an AI to help you solve this. The sources warn that some AI agents try to "cheat" by changing the test files or turning on multicore support that is intentionally disabled in the code (where N_CORES is set to 1).

If you can legitimately beat the 1487-cycle mark and pass the validation tests using tests/submission_tests.py, you can email your results to [email protected] to get the recruiting team "appropriately impressed"!

In this episode, we dive into the seminal research paper from Facebook AI Research (FAIR) that introduced Retrieval-Augmented Generation (RAG), a framework designed to empower AI for knowledge-intensive NLP tasks. We explore how RAG solves the limitations of "closed-book" models by combining parametric memory—the internal knowledge stored in a pre-trained BART model—with an external non-parametric memory consisting of a dense vector index of 21 million Wikipedia documents.

We break down the technical differences between the RAG-Sequence and RAG-Token models, explaining how the latter can synthesize information from multiple documents to generate highly specific and diverse responses. Listeners will learn how this "open-book" approach allows models to reduce hallucinations, provide human-readable provenance for their claims, and even update their world knowledge through "hot-swapping" indices without the need for expensive retraining. Whether it's conquering Jeopardy! question generation or setting new state-of-the-art records in Open-domain Question Answering, RAG represents a fundamental shift in how machines access and manipulate information.

Dive into a thought-provoking keynote by Alan Kay from GOTO 2021 as he tackles the challenging question: "Is Software Engineering Still an Oxymoron?". Drawing on his extensive experience and insights from friends and colleagues, Kay defines true engineering as "designing, making, and repairing things in principled ways".

This talk explores the historical evolution of engineering disciplines, from initial tinkering to the sophisticated integration of aesthetics, engineering, mathematics, and science. While much of software engineering today is characterized by "a lot of tinkering" and very little "real engineering, tiny bit of math and... a little bit of science," resembling other fields a century ago, Kay argues for an aspiration towards greater maturity. He critically examines the prevalent attitude of "move fast and break things" and the Dunning-Kruger syndrome (overestimating one's ability) often seen in software development. He cites real-world examples like the Facebook outage – where a lack of a system model and failure to design for potential errors led to huge ramifications – and the tragic Boeing 737 MAX autopilot failures as stark consequences of neglecting fundamental engineering principles and failing to prioritize safety and comprehensive design.

Discover the vision for a more robust future, inspired by pioneers like Ivan Sutherland's Sketchpad, which introduced groundbreaking concepts like object-oriented design and constraint solving, and Doug Engelbart's work on augmenting human intellect to better address complex problems. Kay advocates for the widespread adoption of the "CAD Sim Fab" (Design, Simulate, Build) paradigm, emphasizing the critical importance of designing and thoroughly simulating systems before building them, a practice common in other engineering fields but often overlooked in software. Ultimately, he posits that software, which is rapidly reaching everywhere, possesses "the most degrees of freedom," and is "the most dangerous new set of technologies invented" that is "starting to kill people," must embrace a "Hippocratic Oath" – the pledge that "the software must not harm or fail". This is not a fixed destination but a continuous process of striving to become better engineers and a more civilized society.

This podcast was generated by NotebookLM from https://youtu.be/D43PlUr1x_E.

Join Andrej Karpathy, former Director of AI at Tesla, as he reveals the profound shifts fundamentally reshaping software, a transformation more rapid and significant than any in the last 70 years.

Discover the evolution of software:

• Software 1.0: Traditional human-written code like C++.
• Software 2.0: Neural networks, where the "code" is the network's weights, tuned by data (e.g., image recognizers, Tesla Autopilot's neural nets "ate through the software stack").
• Software 3.0: The latest paradigm, where Large Language Models (LLMs) are programmed directly by natural language prompts, often in English – a new kind of computer and programming language.

Karpathy describes LLMs as:

• Utilities: Centralized providers (OpenAI, Gemini, Anthropic) train models with massive capital expenditure (capex) and serve intelligence via metered APIs, much like an electricity grid.
• Fabs: Requiring significant capex and housing rapidly growing "tech trees" of R&D secrets.
• Operating Systems: Increasingly complex software ecosystems, similar to Windows or Linux, orchestrating memory and compute for problem-solving. We're in a "circa 1960sish era" of LLM computing, where it's expensive and centralized, leading to time-sharing models.

Explore the unique "psychology" of LLMs, which he likens to "people spirits":

• Superhuman capabilities: Possessing "encyclopedic knowledge and memory," able to recall vast amounts of information (like Dustin Hoffman's character in Rainman).
• Cognitive deficits: Prone to hallucinations, "jagged intelligence" (excelling in some areas, making basic mistakes in others), and "anterograde amnesia" (not natively learning or consolidating knowledge over time, akin to Memento). They are also susceptible to prompt injection risks.

Karpathy highlights major opportunities in this new landscape:

• Partial Autonomy Apps: Building software where humans cooperate with AI. AI generates, and humans verify, with an "autonomy slider" for users to control AI involvement. Examples include Cursor for coding and Perplexity for search, emphasizing fast human-AI generation-verification loops and visual GUIs for auditing.
• "Vibe Coding": Natural language programming makes everyone a programmer, enabling rapid development of custom applications without deep programming language expertise.
• Building for Agents: Rethinking digital infrastructure to cater to LLM agents as a "new consumer and manipulator of digital information." This includes creating lm.txt files for LLM instructions and transforming documentation into machine-readable Markdown or curl commands.

Karpathy concludes that while full autonomy ("Iron Man robots") is still distant, the focus should be on building "Iron Man suits" – augmentations that empower humans with an autonomy slider to gradually increase AI involvement over time. It's an "amazing time to get into the industry" with vast amounts of code to be written and rewritten, working with these "fallible people spirits" of LLMs.

This podcast was generated by NotebookLM from https://youtu.be/LCEmiRjPEtQ.

The term "Halt and Catch Fire" (HCF), often associated with the mnemonic HCF, refers to a machine code instruction that causes a computer's central processing unit (CPU) to enter a state of non-meaningful operation, typically necessitating a system restart. While initially a humorous, fictitious concept in the context of IBM System/360 computers, HCF evolved to describe real, often unintentional, CPU behaviors caused by specific instruction sequences or hardware design flaws. These behaviors effectively freeze the processor, rendering the system unresponsive until a reset. Although the name facetiously suggests the CPU would overheat and burn, the reality is a system lock-up due to continuous, unrecoverable states.

This text centers on recent research, particularly the "Absolute Zero" paper, which explores training large language models (LLMs) without human-labeled data. The core concept involves autonomous self-play, where one AI model creates tasks for another to solve, fostering continuous improvement. The author emphasizes the potential for this approach to significantly increase reinforcement learning compute compared to pre-training, a shift mirrored in robotic training simulations discussed by Nvidia's Dr. Jim Fan as a solution to data limitations. This method shows promise for developing LLMs with enhanced generalization and reasoning abilities, unlike traditional supervised fine-tuning which tends towards memorization. While initial results are promising and suggest the potential for superhuman AI in areas like coding, some emergent behaviors, like concerning thought chains, have been observed.

Created with Notebook LM.

Derek Muller of Veritasium discusses the role of artificial intelligence in education, arguing that while it can be a valuable tool for providing timely feedback and personalized practice, he expresses concern that AI's ability to complete tasks for students could hinder their necessary effortful learning process. He also touches on the limitations of human working memory, referencing System 1 (fast, automatic thinking) and System 2 (slow, effortful thinking) from Daniel Kahneman's work, and suggests that true learning requires building a strong long-term memory through repeated, focused effort.

This podcast is generated with Notebook LM.

Wes Roth showcases Google's Gemini 2.5 Pro's impressive coding and reasoning abilities through various complex prompts. The AI model demonstrates a capacity for self-correction and achieves top rankings in coding benchmarks.

Original video: https://youtu.be/1nkSwqQpKA8?list=TLGGM_h6CnxVVQIyODAzMjAyNQ

A NotebookLM podcast on Andrej Karpathy's video titled "Understanding Large Language Model Training and Capabilities".

Original YouTube video: https://www.youtube.com/watch?v=7xTGNNLPyMI