“AI #114: Liars, Sycophants and Cheaters” by Zvi
02:07:20 Share

Gemini 2.5 Pro is sitting in the corner, sulking. It's not a liar, a sycophant or a cheater. It does excellent deep research reports. So why does it have so few friends? The answer, of course, is partly because o3 is still more directly useful more often, but mostly because Google Fails Marketing Forever.

Whereas o3 is a Lying Liar, GPT-4o is an absurd sycophant (although that got rolled back somewhat), and Sonnet 3.7 is a savage cheater that will do whatever it takes to make the tests technically pass and the errors go away.

There's real harm here, at least in the sense that o3 and Sonnet 3.7 (and GPT-4o) are a lot less useful than they would be if you could trust them as much as Gemini 2.5 Pro. It's super annoying.

It's also indicative of much bigger problems down the line. As capabilities increase and more RL [...]

Outline:

(01:39) Language Models Offer Mundane Utility

(04:29) Language Models Don't Offer Mundane Utility

(06:57) We're Out of Deep Research

(12:26) o3 Is a Lying Liar

(17:27) GPT-4o was an Absurd Sycophant

(20:54) Sonnet 3.7 is a Savage Cheater

(22:27) Unprompted Suggestions

(31:27) Huh, Upgrades

(32:14) On Your Marks

(32:55) Change My Mind

(42:52) Man in the Arena

(45:05) Choose Your Fighter

(45:45) Deepfaketown and Botpocalypse Soon

(49:43) Lol We're Meta

(52:48) They Took Our Jobs

(59:15) Fun With Media Generation

(59:53) Get Involved

(01:03:21) Introducing

(01:03:50) In Other AI News

(01:08:10) The Mask Comes Off

(01:24:25) Show Me the Money

(01:27:32) Quiet Speculations

(01:29:55) The Quest for Sane Regulations

(01:37:04) The Week in Audio

(01:38:08) Rhetorical Innovation

(01:44:59) You Can Just Do Things Math

(01:45:34) Taking AI Welfare Seriously

(01:47:54) Gemini 2.5 Pro System Card Watch

(01:52:29) Aligning a Smarter Than Human Intelligence is Difficult

(01:58:49) People Are Worried About AI Killing Everyone

(01:59:46) Other People Are Not As Worried About AI Killing Everyone

(02:04:55) The Lighter Side

First published: May 1st, 2025

Source: https://www.lesswrong.com/posts/pazFKtkp7T8qaRzva/ai-114-liars-sycophants-and-cheaters)

    ---
    

Narrated by TYPE III AUDIO).

Images from the article: )))))))))![Text excerpt discussing OpenAI's nonprofit and profit entities in California.

The passage mentions charitable initiatives, control relationships, and mission-related concerns between OpenAI's nonprofit and profit segments.](https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/pazFKtkp7T8qaRzva/r74z3zbf8randshpi4il)))))![Table comparing OpenAI's governance safeguards between current state and proposed restructuring.

The table shows six key governance safeguards, with ](https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/pazFKtkp7T8qaRzva/qdkxdqlic1lh1xz1ytaf)))![Text excerpt discussing OpenAI's governance, safety practices, and contradictory actions regarding AI development.

The highlighted sections emphasize concerns about OpenAI abandoning safeguards, lobbying against regulations, and using restrictive employee agreements.](https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/pazFKtkp7T8qaRzva/w3cf68wyjike5ma7kaqt))![This is a Winnie the Pooh meme format commenting on model reliability.

The meme contrasts two statements: a simple claim ](https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/pazFKtkp7T8qaRzva/l6bdg16itewxel76wwlf)))![Hugging Face Spaces directory showing AI app categories and trending applications.

The interface displays various AI applications organized in a grid layout, with categories including Image Generation, Video Generation, Text Generation, and others. Each application card shows its title, description, and engagement metrics. The cards feature gradient backgrounds in purple, blue, red, and green colors, with icons and running status indicators.

The top section includes a search bar and navigation menu, while the main content area showcases ](https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/pazFKtkp7T8qaRzva/ghop30rvjtuhubyu7ysr))))![Table showing capability assessment results for Gemini 2.5 Pro Preview across security areas.

The table summarizes test results in CBRN, Cybersecurity, Machine Learning R&D, and Deceptive Alignment, indicating CCL (Critical Capability Level) statuses.](https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/pazFKtkp7T8qaRzva/ila8zztj7bspk8xcydrr))![Diagram showing model validation methods for baking with frozen butter.

The image illustrates different evaluation methods (MCQ Knowledge, MCQ Distinguish, Open-Ended Belief, and Generative Distinguish) for testing a model's understanding of using frozen butter in cake baking, with example questions and responses shown for each method.](https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/pazFKtkp7T8qaRzva/rmx8ih4gojb9wndxrjtb))![Screenshot of OpenAI governance requirements listing six key areas with highlighted text.

The text outlines requirements for board restructuring, director removal, independence, expertise, resources, and oversight measures for OpenAI's operations.](https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/pazFKtkp7T8qaRzva/sn7mmwc7o47tw2gx7fga))))![Benchmark comparison table showing performance scores of different AI language models.

The table compares Gemini 2.5 Pro, OpenAI models, Claude 3.7, Grok 3, and DeepSeek R1 across various capabilities including reasoning, science, mathematics, coding, and language tasks. Each row shows percentage scores for different benchmark tests.](https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/pazFKtkp7T8qaRzva/whjklcqi493rj0vzbypl))))))) Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts), or another podcast app.