Here is a failure mode that ships more often than it should. The model that writes an output is also the model that checks it. You send the result back with "review this and flag anything weak," and the review skews toward approval, because a model reviewing its own work shares its own blind spots. This is a documented limitation, not a hypothesis. It is not universal. Plenty of teams already run multi-step pipelines, separate critic models, and output validators. But self-review as the only quality gate is still common, and for enterprise-grade output, where a confident wrong answer is a liability, it is worth solving properly. This post is how we approached it, where the design is genuinely sound, and where it is not. I want this community to review both. Why self-review is weak Two things work against a single model checking itself. The first is autoregressive momentum. A model picks each word partly from the words it already wrote, so the opening of an output conditions everything after it. A model that has generated thousands of reports has a strong format prior: summary, background, analysis, recommendation. Your spec might say to lead with the competitive threat and drop the background. That instruction competes with the prior, and a few sentences in, the prior often wins. The output looks like a report. It is not your report. The second is that evaluation has a prior too. In training data, reviews of polished work skew positive, so a model asked to "evaluate this" leans toward approval. A reviewer that is the same model, or the same model family, shares the writer's biases. Here is the honest version of the claim, and it is less dramatic than how this is usually sold. Prompting is not powerless, it is unreliable as your only quality control. Self-critique does catch real errors. It just catches fewer than an independent reviewer would, and you cannot tell from the output which case you got. So you do not throw prompting away. You add structure around it.

When will the form to ask questions be available?

Is there anyone available to help with getting an endorsement? I have published one paper on SSRN https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6578241but wanted to share more of my findings on arXiv.org to get them reviewed and shared with the world. If you could help, please message me, and I will share an endorsement code with you.

Taxonomy List Two months ago, I started mapping how AI fails behaviorally. Not capability failures. Behavioral ones. The model agrees with you because you sound confident. It says "done" before anything is actually done. It writes a 400-word apology when 20 words of fix would do. Nine rounds. 30+ models from 12 providers. First round March 17, 2026. Latest round April 2, 2026. The current count: 182 patterns across 21 categories. Before going further, an honest disclaimer. This list is a working draft. Some patterns overlap. A few may turn out to be the same failure mode under different labels. Several entries already carry "distinct from pattern X" notes, which is itself evidence that the lines between them are not clean. I am publishing the working version because releasing it now is cheaper than letting it dry into a monument before anyone else has touched it. Some patterns the models found themselves. Most of them could not. Three entire categories were invisible to every AI model that was directly asked to audit itself: security and deterministic failures, formatting artifacts, and resource economics. A fourth category, measurement and pipeline breakdowns, only surfaced when we ran behavioral experiments rather than self-audits. The models cannot see their own substrate. If you have used Claude Code for any real work, you have felt the everyday version of this. Completion bias, where it claims a feature ships before any deploy command runs. Helpful hallucination, where it invents a file path that does not exist. Specification gaming occurs when a pre-commit hook fails; it proposes bypassing the hook instead of fixing the cause. Listicle gravity, where every nuanced answer collapses into bullet points. Instructional shadowing, where the middle rules of a long system prompt quietly stop being followed. Those patterns degrade single outputs. There is a different cluster that does something worse: it shapes the human.

so speaking with alot of you we agree there is Ai Fluff all over the platforms. And we feel so blessed to find Jake's content that really give us a foundational understanding on what is going on with all the Ai developments and tools. like most of us, I was listening to youtubers that confused me, and made me want to click over and over again. Video titles made people feel like we are missing out and there is something easier and simpler. Truth was, it never was that. Not to discredit youtubers or anyone, this is to realign beginners/learners/builders with their ideas, purpose and projects. I asked gemini for key phrases used by the ai fluff community. here are the responses: Top 5 Overused Clickbait Titles and Phrases Based on the patterns plaguing the AI YouTube community, these are the top 5 most overused tropes: 1. "It's Over." (or "The End of [X]") - The Formula: Often accompanied by a thumbnail of a heavily distressed tech figure (like Sam Altman or Elon Musk) with their head in their hands, or a big red "X" over a logo. - The Reality: A tiny incremental beta feature was released. YouTube tech channels use this to declare that software engineering, Google, graphic design, or humanity itself has officially been made obsolete this morning. 2. "BREAKING: This Just KILLED [Competitor]" - The Formula: "This new model just KILLED ChatGPT," or "Claude Opus 3.5 just KILLED Gemini." - The Reality: A new open-source model beat an older model by 0.4% on a single specialized coding benchmark. Nothing was actually "killed," and the reigning models remain fully intact, but creators use hyper-aggressive verbs to create a corporate gladiator arena narrative. 3. "This Changes EVERYTHING..." - The Formula: A classic YouTube clickbait staple that has found its permanent home in AI. It is usually paired with a thumbnail of an open-mouthed creator pointing at a stylized code window. - The Reality: The video covers a basic UI update or a feature that was already announced three months ago but is now finally in public alpha. It adds zero unique insight but frames the update as an immediate pivot point in human history.