Frontier Watch

macOS 26.2 added a direct memory link between Apple silicon Macs over Thunderbolt 5, and it changed Mac clustering from a capacity trick into something that can genuinely run faster as machines are added. This brief separates the three layers people confuse, works through a 256GB Mac Studio paired with a 48GB Mac mini, and labels every published figure by whether a vendor or an independent tester produced it. Full written version: https://macluster.novcog.us.com/

A free AI pentesting agent with nearly 70,000 GitHub stars — and almost every explainer describes an architecture that is not in the code. I read the source instead: one agent template that hires its own help, a 75-file skill library handed over five at a time, and exactly one named role that never gets improvised. Full analysis, sources and what we could not verify: https://strix.novcog.us.com/ — video walkthrough: https://youtu.be/FcBm0DioaTw

Nick Cammarata co-authored the 2020 paper that made the cat-and-car neuron the textbook case of a messy network. On 15 October 2026 he published Neural Networks Are Already Clean with the interpretability lab Goodfire, arguing that the evidence for superposition was misread and that about 90 percent of neuron poles in Qwen3-32B have one main job. An independent read: the cat-and-car neuron as a V, the 2019 circuit, the squeeze tests, the numbers and what they count, sparse autoencoders, how researchers including Neel Nanda, Liv Gorton, Leo Gao and Stephen Casper reacted, and what the essay does not show yet. Full breakdown: https://alreadyclean.novcog.us.com/ . Video: https://youtu.be/VrXTFsd7t5Y

Telling an AI model what not to do is the weakest instruction you can write. Seven papers on negation, the Pink Elephant problem, ironic rebound inside Llama 3, and Lost in the Middle explain why, and why long sessions make it worse. What to write instead. Sources: https://pinkelephant.novcog.us.com/ Video: https://youtu.be/GSMlVWcPZs0

Corrected re-issue of episode 36, which had an audio glitch in the opening. A DGX Spark cluster was build for an SEO agency. Here is why that is the wrong answer. What a HAR file can and cannot show about AI search, whether HAR analysis pays, why the SEO jobs are mostly a data pipeline, model choice and the harness, four 128 GB Sparks vs one 256 GB Mac Studio, and leasing instead of buying. Sources: https://sparkoverbuild.novcog.us.com/ Video: https://youtu.be/Ha06g0tC8X0

A DGX Spark cluster was build for an SEO agency. Here is why that is the wrong answer. What a HAR file can and cannot show about AI search, whether HAR analysis pays, why the SEO jobs are mostly a data pipeline, model choice and the harness, four 128 GB Sparks vs one 256 GB Mac Studio, and leasing instead of buying. Sources: https://sparkoverbuild.novcog.us.com/ Video: https://youtu.be/3x5wnMJejdA

If you work long enough with AI, you will see output in some strange languages. Even American lab models like ChatGPT and Claude drop into Mandarin, especially in the chain of thought. Why training does not keep models in one language, the reward DeepSeek added to fix R1, the 14-model study that found garbled reasoning everywhere except Claude, OpenAI o1 unexplained switches, and the Claude bug Anthropic disclosed. Sources: https://mandarinslip.novcog.us.com/ Video: https://youtu.be/uyvbsqiCJjA

In five weeks Qwen 3.8 27B moved onto a desk, got twice as fast and had its refusals removed. 24 GB: 48 of its 64 layers never cache a token. 2.24x: multi-token prediction on an M5 Max, same output distribution. 2.1 points: the MMLU cost of abliteration, published by one builder of four. Every number attributed, plus what we did not test: https://threenumbers.novcog.us.com/ — video: https://youtu.be/DkmR0r5nYrs

Three researchers at Hacktron AI ran an authorized bug-bounty test and reached OpenAI internal systems; OpenAI paid $6,500 and fixed it in ~14 hours. The story is the economics and the timing: under $3,000 in tokens, a task Opus 4.8 failed for weeks solved within hours of Opus 5, a guardrail that held only until reworded, and CISA cut a third of its staff. Full analysis, both ends of every date, sourced: https://cheapoffense.novcog.us.com/

A 33-minute explainer calls TypeSafe's Jev classifier one of the biggest events in the history of computer science. We read the benchmark that claim rests on. TypeSafe's own methodology defines the correct answer as the average of GPT-6 Astra and Fable 5.1, so the headline measures agreement with two frontier models rather than accuracy — and they say so themselves. Every published real-world speed measurement lands between 1.16x and 6x against a 193.6x headline, the lowest of them from TypeSafe's own employee measuring a real pipeline. A limitations section is promised three times and is absent from the video's own chapter markers. What holds up: the latency, the price, and the type guarantees. What should worry you: a pre-registered evaluation found 0 of 30 out-of-scope messages flagged, at 0.99 confidence. Full analysis with every quote timestamped at https://fallsover.novcog.us.com/ — including the two corrections we made to our own draft.