Rick Rosner On, 33: AI and LLM Capabilities Change

Author(s): Scott Douglas Jacobsen

Publication (Outlet/Website): Personal SubStack

Publication Date (yyyy/mm/dd): 2026/08/25

In this interview, Scott Douglas Jacobsen and Rick Rosner discuss how AI systems have narrowed gaps in breadth, retrieval, and conversational usefulness while still struggling with calibration, personalization, reliability, and obscure information. Rosner reflects on search-enabled AI, credit-card fraud research, traffic-camera questions, and adaptive responses. Jacobsen emphasizes domain-specific expertise calibration and memory. They also compare high-IQ testing histories, public reputations, and the risks of misinformation and extremist rhetoric in public discourse.

Closing Gaps in AI Output

Scott Douglas Jacobsen: What gaps in AI output, particularly LLMs, are you noticing now compared with three years ago? What gaps have you seen close?

Rick Rosner: The expectation now is that, for almost any question you can come up with, AI can give you a fairly satisfying answer. That’s different from two or three years ago.

Credit-Card Fraud and AI-Assisted Information Retrieval

I had trouble with a credit card that was repeatedly hit by fraudulent charges, so we closed the account. With a U.S. credit card, your liability for genuinely unauthorized use is limited by federal law. If somebody steals only your account number and uses it without authorization, you generally have no liability. If the physical card itself is lost or stolen, federal liability can be as much as $50 for unauthorized charges made before you notify the issuer, although many issuers provide zero-liability protection.

Our card was locked up in a lockbox in a locked room. Somebody apparently obtained the card information and repeatedly used the account fraudulently, to the point where we cancelled the fucking card.

But the bank that issued the card was giving us shitty customer service. After an hour on the phone, I asked AI, “This bank has shitty customer service about credit-card fraud, right?” And it said, “Yes, there are complaints about that, although other customers report that the bank handled their cases well.”

The fact that I could ask that question and reasonably expect an answer is different from a few years ago. That’s not even an especially obscure question. The idea is that you can ask about an enormous range of publicly documented events, and an AI system with web access may be able to find relevant information.

That’s different from saying AI automatically knows about every event that ever became public. Models can have incomplete training data, outdated knowledge, weak retrieval, missing sources, inaccessible databases, or get things wrong.

Speed Cameras and Current Public Information

One of Carole’s friends got a red-light-camera ticket, so I asked all sorts of questions about it. Then Los Angeles started preparing to deploy speed cameras on certain streets.

I tend to go a little fast in a lot of places, and I don’t want to get a ticket. So I asked, “Where are the speed cameras going to be? When do they start working? When do citations become enforceable?”

Los Angeles is launching a Speed Safety Systems pilot in late summer or fall 2026. LADOT has published information about the program and its camera locations, so an AI with current web access can retrieve that.

It knew the publicly available stuff. I expected it to know the publicly available stuff.

The Breadth and Limits of AI Knowledge

I’m trying to come up with something sufficiently obscure to be surprising. You can ask about a huge number of actors in productions over the past 50 years and often get a ton of information.

But “AI knows every knowable thing” is too strong. Current AI systems do not contain every fact, and they don’t have reliable access to every database, private record, obscure publication, or newly occurring event. Even when the information is somewhere in the training data or available online, the model may fail to retrieve or reproduce it accurately.

Sometimes an AI will also refuse to answer certain questions because of safety rules or other restrictions. Politically loaded questions aren’t necessarily refused, but systems may be designed to avoid partisan persuasion or unsupported political claims. They also won’t provide dangerous instructions, such as detailed guidance for building an explosive.

So my answer remains that what’s impressive is how broad the accessible information space has become. An AI connected to search can reach a huge fraction of what has been publicly documented, even if “just about anything that could be known” overstates it.

If I asked it to give me 25,000 words on you, it might generate 25,000 words. But length isn’t evidence of knowledge. Some of that could be repetitive, speculative, or completely fabricated unless it had enough reliable source material.

Is that surprising? I don’t fucking know.

What gaps do you notice? I don’t work extensively with LLMs; I mainly ask Google AI questions. What gaps do you see?

Calibrating AI to User Expertise

Jacobsen: One gap is automated calibration to the user’s prior knowledge, especially their domain-specific expertise and level of engagement.

The ideal system would infer, from a user’s history and the immediate context, roughly how sophisticated an explanation should be and automatically adjust the language, assumptions, technical detail, and background information.

So you wouldn’t constantly have to say, “Dial it up because I’m already an expert in this,” or, “Dial it down because I’m a beginner.”

Current systems can personalize when they have access to conversational context, saved preferences, or other user-specific information. But finely calibrated, reliable adaptation to a person’s expertise across different domains is still incomplete. For example, someone can be an expert in psychometrics and a novice in particle physics. A good system has to distinguish those rather than assigning one global level of sophistication to the person.

Personalization as an Unsolved Problem

Rosner: So you’re saying that’s still a gap? Are you not seeing it fully solved yet?

Jacobsen: I’m not seeing that consistently yet, but these systems clearly have some capacity for it because they can retain or reference user-specific context when memory features are enabled. They can use saved memories, chat history, custom instructions, and other context to personalize responses. That doesn’t mean every system keeps a complete record of everything a user has ever done, or that it automatically builds a precise expertise profile for every domain.

But if a system has enough of your interaction history and can analyze it, it could tailor the experience in a way that improves usability. It’s small, but even a modest usability improvement, multiplied across hundreds of millions of users, would be substantial. I wouldn’t put a precise “one billion monthly users” figure on it without tying that to a specific product and date.

So that could be.

Associative Personalization

Rosner: Sure. I’ve seen AI tailor itself to me, not necessarily in the sense that I might use longer words than somebody with a fourth-grade education, but it definitely reflects to me what I give it. With a lot of prompts, it comes back with reasonable associations or related details that it thinks I might appreciate.

If I write to it about John Brunner’s 1968 novel Stand on Zanzibar, for instance, it might bring up another of Brunner’s best-known novels, The Sheep Look Up, from 1972.

Jacobsen: Yes, right.

Malcolm Gladwell, Christopher Langan, and the Other Bouncer

Rosner: And I assume I embarrassed myself slightly in front of Malcolm Gladwell. He wrote about Christopher Langan in Outliers, including Langan’s unusual intellectual history and his work as a bouncer.

Gladwell was on Kimmel, and I came up to him and said, “I’m the other bouncer with the high IQ.”

He said, “What?”

Because I assumed, how could you know about Langan and not know about me, since we’re basically the same fucking person? And I’m like, forget it. I didn’t want to go into the whole thing. It would have looked weird.

High-Range IQ Testing

Jacobsen: It’s funny because Langan and you both became prominent in the high-range IQ-test subculture, but you have to state the exact score history carefully.

You scored 44 out of 48 on Ronald Hoeflin’s Mega Test in 1985, later scored 47 on another attempt, and received the only reported perfect 48 out of 48 on Hoeflin’s Titan Test before the test was compromised. Hoeflin historically assigned those scores extremely high IQ estimates. Although later psychometric analysis has argued that the official IQ conversions were inflated.

Langan’s score history is messier. He is widely reported as having scored 42 out of 48 on an initial Mega Test attempt and 47 out of 48 later.

So on Hoeflin’s tests, you had exceptionally high scores, including the perfect Titan score. But saying you were definitively better then tied in overall intelligence would go beyond what those tests can establish.

Conspiracy Rhetoric, Fraud Claims, and Community Responsibility

There’s another major difference now. Langan has made increasingly grandiose, resentful, conspiratorial, demonological, and white genocide theory-based claims against “Big Jew Inc.,” “Ziobanksters,” etc. It’s a real achievement to do that unabashedly and to inspire further copycats in well-known and thoroughly investigated frauds like YoungHoon Kim who worked for him as his Senior Membership Officer with overlaps in highest IQ claims, simulation theory, Christian theology, Jesus exaltation, self-aggrandizement, Intelligent Design community creationism/rejection of naturalistic evolution alone, and messianic statements. Langan became the template for Kim. He witnessed the inner circle and things from the inside.

Although, many in the Christian communities in the high-IQ communities plus Christian media are the primary culprits in bringing them unquestioned modest prominence, then backing away from the mess created. So, if one inspires hate rhetoric, conspiracy theories, and prejudice while the other continues to lie, defame, and scam people, then they have primary responsibility in propping them up – not the specific rhetoric or actions, but I do not expect any or much accountability – which is typical.

Analogously, I’ve seen the reportage and spoken with experts on Christian community cross-denominational reactions to clergy-perpetrated abuse: The broad, consistent finding is defaming truth-tellers, blaming victims, denying the reality, and delaying justice. I suspect the same long-term pattern here, as we have already seen for years now. It’s a similar unfortunate pattern running from the far more prominent cases like Jim Bakker, Ravi Zacharias, Bill Hybels, Jean Vanier, Marcial Maciel, Peter Popoff, and others.

The Little Jewish Comedy Writer Who Could

You, by contrast, are the little Jewish comedy writer who could. Thank you very much for the opportunity and your time, Rick.

Last updated May 3, 2025. These terms govern all In-Sight Publishing content—past, present, and future—and supersede any prior notices.  In-Sight Publishing by Scott  Douglas  Jacobsen is licensed under a Creative Commons BY‑NC‑ND 4.0; © In-Sight Publishing by Scott  Douglas  Jacobsen 2012–Present. All trademarks, performances, databases & branding are owned by their rights holders; no use without permission. Unauthorized copying, modification, framing or public communication is prohibited. External links are not endorsed. Cookies & tracking require consent, and data processing complies with PIPEDA & GDPR; no data from children < 13 (COPPA). Content meets WCAG 2.1 AA under the Accessible Canada Act & is preserved in open archival formats with backups. Excerpts & links require full credit & hyperlink; limited quoting under fair-dealing & fair-use. All content is informational; no liability for errors or omissions: Feedback welcome, and verified errors corrected promptly. For permissions or DMCA notices, email: scott.jacobsen2025@gmail.com. Site use is governed by BC laws; content is “as‑is,” liability limited, users indemnify us; moral, performers’ & database sui generis rights reserved.