Corporate AI Fails Basic Office Pop Quiz.
Also Boeing Smoke, Tiny Datasets, and Blog Existential Dread.

The New Intern Is Charmingly Full of It

The general sentiment that Large Language Models, or LLMs, are nothing more than extremely sophisticated "bullshit machines" is finally being treated as a serious academic topic. University of Washington professors Carl Bergstrom, a data scientist, and Jevin West, a biologist, have taken the lead by creating an entire course dedicated to navigating the new reality. The key takeaway: LLMs like ChatGPT are best described as "autocomplete in overdrive," which means they are simply generating plausible text resembling their training data without any genuine concern for truth or logical coherence.

It is important to note the difference between a malicious system and a merely incompetent one. The authors argue that AI is not technically "lying," which requires intent to deceive. Instead, LLMs are simply practicing bullshitting, which is the act of communication designed to appear authoritative or persuasive, yet entirely without regard for the actual facts. The problem is that the energy required to refute AI misinformation far outweighs the effort needed to create it, a dynamic that will flood the internet with subtle inaccuracies. As one Hacker News commenter put it, you are headed into a time where a lot of people are going to implicitly trust the output from these devices and the world is going to be swamped with a huge quantity of subtly inaccurate content.

The LIMO Hypothesis: Training Is Just Good Flashcards

For years, we have been told that the path to true AI reasoning is through simply pouring quadrillions of tokens onto the datacenter floor. Now, a new paper challenges this brute force approach with the LIMO model, short for "Less Is More for Reasoning." The researchers demonstrated that complex mathematical reasoning can be elicited from a model using a surprisingly minimal set of examples; only 817 carefully curated training samples were needed.

This is a catastrophic embarrassment for the Big Data teams. LIMO, which achieved state-of-the-art performance on difficult math benchmarks, used only 1 percent of the training data required by previous methods, proving the expensive, petabyte-scale training dataset was mostly just digital noise and cat GIFs. The finding suggests that if the foundation model already has the domain knowledge encoded, all you need is a high-quality "cognitive template" to show it how to think, instead of forcing it to re-read the entire library. It turns out the true innovation was just paying someone to make decent flashcards.

Boeing System Tries To Save Engine, Accidentally Smokes Out The Cabin

Boeing has done it again. It seems a "safety feature" called the Load Reduction Device, or LRD, on the 737 MAX is designed to activate after a major engine incident like a bird strike. The device works to reduce vibrations transmitted to the airframe, which is commendable. Unfortunately, its activation can cause tubes supplying oil to the engine sump to become dislodged, allowing burning oil fumes to be fed directly into the cockpit and cabin through the pneumatic bleed ports. The video analysis confirms this occurred in two separate incidents involving Southwest flights.

The NTSB, who were clearly unimpressed, noted that Boeing and engine manufacturer CFM International are developing a software update to close a pressure regulating shutoff valve more quickly. One internal FAA memo cited a Boeing presentation that reportedly stated that the release of oil from the sump is an expected result of LRD activation, yet nothing was in the pilot manuals about the LRD's existence. It is a classic move of an overworked engineering department fixing a high-level problem and creating a low-level, high-visibility ops-center catastrophe. The good news is the fix is a software update, which should install without issue, given the company's track record.

Briefs

• Digital Cremation: An article asks what happens to your online accounts when you die. The answer is mostly "nothing, they just sit there, adding to the general entropy of the universe" and proving that even in death, your data is still a dependency problem.
• Hard Drive Time Theft: Reports of fraud involving long runtimes on new Seagate hard disks indicate that the new drives might not actually be new. They have been running the equivalent of a server farm in Shenzhen before they even reach your desk.
• The Insurance of Boredom: UnitedHealth is reportedly sick of everyone complaining about its claim denials. The insurance company feels misunderstood, which is precisely how every helpdesk ticketing system feels after auto-closing a ticket that was clearly still valid.