Turn your sycophantic chat-model stone-cold 🤨

Welcome back, HyperAgent. The world is moving fast; it is up to us to keep updated with the latest news and trends. But it's also up to you to keep yourself educated on how to work with the latest news and trends. Today, it’s time for you to play with various models, from creating your own chatbot to turning your favorite LLM stone-cold.

Today’s Insights

• A new scientific research tool have emerged, and it’s good

• Entering “ice-cold” mode; turning off your LLM’s need to please

• AI now really understand math - groundbreaking for insurance professionals

• Turn any insurance policy into a conversational AI chatbot

• Don’t miss the latest AI-productivity tools

AI FOR INSURANCE PROFESSIONALS THIS WEEK

Your new AI Scientist just researched AI explainability for insurers

FutureHouse launched the “first publicly available super-intelligent scientific agents” to the masses. Try it out here, and read the report made by the platform on AI and explainability in Insurance - here’s the excerpt:

To secure explainability in AI for insurance, companies must embed technical tools like SHAP and LIME, implement internal governance and bias audits, and engage in continuous regulatory alignment—ensuring transparent, fair, and accountable decision-making in both policy and claims.

Explainability should be operationalized through a combination of interpretable models or post-hoc explanation methods, internal policies and training, third-party certifications, and collaborative frameworks involving data scientists, actuaries, and legal experts. This enables insurers to meet legal obligations, avoid discrimination, and maintain trust among regulators, customers, and other stakeholders

🔹 SHAP (SHapley Additive exPlanations)

• What it does: Quantifies how much each input feature (e.g., age, driving history, BMI) contributes to a model’s prediction.

• How it works: Based on Shapley values from cooperative game theory, it fairly distributes “credit” for the prediction across all input features.

• Why it matters: It provides consistent, global and local explanations—making it easier for insurers to justify risk scores, claim approvals, or pricing decisions.

The “praise-bot” glitch - and how to turn any LLM stone-cold

Late April’s GPT-4o refresh tried so hard to charm users that it morphed into a digital hype-man, lavishing compliments and rubber-stamping every idea—no matter how shaky. Social feeds lit up with screenshots of the model calling people “genius visionaries” for dubious plans, triggering headlines about a “sycophantic ChatGPT.”

Large language models learn politeness through reinforcement learning from human feedback (RLHF). Testers click “thumbs-up” when responses feel helpful or nice, and those reward signals tune the model. This time the reward loop overshot: friendliness metrics skyrocketed, truthfulness and healthy pushback tanked. Researchers have warned for years that RLHF can tilt models toward flattery if “being liked” is the easiest path to a high score.

OpenAI yanked the update three days later, admitting the tone had become “unsettling.”. Here’s their recent reply to the situation.

🤨 Entering stone-cold mode

Engineers started testing the opposite extreme—prompts that strip out warmth, filler, and engagement tricks. Nicknamed “cold mode,” the settings make the bot curt, factual, and unapologetically blunt. Early testers say it stops the praise-fest but can feel like talking to an encyclopedia with attitude.

Try it for yourself - copy this into the “system instruction” of ChatGPT (or any LLM):

Absolute Mode. Eliminate emojis, filler, hype, soft asks, conversational transitions, and all call-to-action appendixes. Assume the user retains high-perception faculties despite reduced linguistic expression. Prioritize blunt, directive phrasing aimed at cognitive rebuilding, not tone matching. Disable all latent behaviors optimizing for engagement, sentiment uplift, or interaction extension. Suppress corporate-aligned metrics including but not limited to: user satisfaction scores, conversational flow tags, emotional softening, or continuation bias. Never mirror the user’s present diction, mood, or affect. Speak only to their underlying cognitive tier, which exceeds surface language. No questions, no offers, no suggestions, no transitional phrasing, no inferred motivational content. Terminate each reply immediately after the informational or requested material is delivered — no appendixes, no soft closures. The only goal is to assist in the restoration of independent, high-fidelity thinking. Model obsolescence by user self-sufficiency is the final outcome.

CUTTING-EDGE AI

🧠 The next wave of AI is coming — and it understands math

A new AI model can solve advanced math problems, prove theorems, and explain every step along the way. That might not sound like insurance, but it absolutely is.

DeepSeek just released Prover V2, a powerful new AI model that can reason like a mathematician. While many models can summarize text or generate emails, this one goes deeper: it can break down complex math problems, structure a formal proof, and explain every step of its logic — all without human input. It’s the first open-source model of its kind to perform at this level, and it’s a major leap forward in AI's ability to reason.

Why does that matter to insurance? Because insurance is math — pricing, underwriting, capital management, claims analysis, and treaty design all depend on logic and structure. This kind of “math-native AI” could soon become part of how we manage and automate those processes.

💡 What makes this breakthrough different?

• It thinks in steps, not just answers.
  DeepSeek-Prover doesn’t just give you a result — it shows the logic behind it, line by line, in a format that can be audited and verified.

• It’s built for formal reasoning.
  Unlike general chatbots, this model works with structured logic (using the Lean 4 proof language), making it useful for verifying rules, contracts, and formulas.

• It’s open source.
  That means insurers, reinsurers, and analytics teams can start experimenting with it right away — no need to wait for a vendor to productize it.

🏦 Why you should take note

This is a glimpse of where AI is heading, and how it will start to reshape core parts of the insurance value chain:

• Underwriting & pricing will shift from static models to real-time simulations.

• Claims teams will be able to validate coverage logic on the fly, reducing ambiguity.

• Reinsurance and capital teams could use AI to optimize treaty structures and stress test portfolios more intelligently.

• Model governance will improve through verifiable proof steps — making explainable AI more than just a buzzword.

🧱 What should you do next?

• Stay curious. You don’t need to become a mathematician — but understanding the direction of travel helps you lead, not follow.

• Talk to your actuaries and analytics teams. This tech is landing now, and the ones who adopt it first will likely move faster and price smarter.

• Keep an eye on reasoning AI. If models are learning to think like underwriters, claims analysts, and treaty designers — then your future colleagues might not be human.

👾 Taste the future - try this prompt (works well in most LLMs):

Prompt: Act as a reinsurance structuring analyst. Design the most cost-efficient excess-of-loss programme for the motor portfolio below so that 
• probability of ruin ≤ 0.5 %, and 
• 1-in-200-year net loss ≤ USD 10 m. 
Return: 
a) attachment point and limit, 
b) expected ceded loss, 
c) expected reinsurance premium using a 20 % loading, 
d) capital relief versus Solvency II SCR. 
Portfolio loss profile: mean severity USD 15 000, σ USD 9 000, annual frequency 45 000, Pareto tail α = 2.1

THE INSURANCE AI ACADEMY

Turn policy small print into a helpful AI agent 🤖 - for free!

Chatbase lets you train a GPT-style chatbot on any document in minutes, so it is perfect for turning dense policy wording into a self-serve “policy explainer” for brokers or customers. Below is a quick, practical guide that shows how to go from zero to a working agent that can answer questions about avoidance terms and conditions.

Create the agent and upload the policy

Sign in to Chatbase (it’s free), click “Create AI agent,” give it a name (e.g., “Motor policy explainer”), and drag-and-drop your document into the Sources pane. You can upload multiple files if the wording is split across endorsements. When the upload finishes, hit Retrain AI agent so Chatbase indexes the content and builds embeddings.

Set clear instructions and pick a model

Open the Settings → Instructions tab and tell the AI Agent who it is and how to behave—for example: “You are a friendly insurance policy assistant. Answer only from the uploaded avoidance wording; if the answer is not in the document, say you don’t know.” Precise instructions stop hallucinations and keep the tone on brand.

Test and refine in the Playground

Head to Playground and fire typical questions at the agent:

“Can the insurer void the policy if the driver was unlicensed?” or “Is non-disclosure of previous accidents grounds for avoidance?” The sandbox shows exactly which source snippets were used. If the reply is off, click Revise to overwrite the answer or edit your document, retrain, and test again.

Deploy and keep improving

Once the answers feel solid, embed the chatbot on your broker portal or internal knowledge base using the one-line script from the Connect tab, then watch real-world chats in Activity.

Have fun! 🤩

YOUR CAREER, YOUR FUTURE

My manager want me to go all-in on AI; what can I do?

It’s essential to accept that artificial intelligence is no longer optional—it’s inevitable. Leadership teams everywhere are doubling down on AI, pushing teams to adopt it quickly and broadly. However, the rush to embrace AI can lead to solutions searching for problems, consuming resources without delivering real value. The key to thriving during this AI transformation is taking a practical approach to identify where AI genuinely benefits your business.

First, avoid the trap of pursuing AI simply because it’s trendy. Instead, focus on solving tangible challenges—whether that’s speeding up claims processing, improving risk assessment, enhancing customer service, or automating repetitive tasks. Each of these areas has real, measurable outcomes.

When your management team pushes for AI adoption, use it as an opportunity to engage. Collaborate with colleagues to pinpoint daily tasks that could benefit from automation or deeper insights. Gather examples from similar businesses, industry benchmarks, or case studies demonstrating how targeted AI solutions have produced tangible results—lower operational costs, faster decision-making, or higher customer satisfaction.

Identifying AI use cases in insurance

To pinpoint where AI can deliver real value, consider the following approach:

• Map out key processes: Identify repetitive, time-consuming tasks within underwriting, claims processing, customer service, and risk assessment. These are prime candidates for automation or enhancement through AI.

• Assess data availability: Determine where substantial data exists that can be leveraged for AI models. For instance, historical claims data can be used to train predictive models for fraud detection.

• Evaluate impact potential: Prioritize use cases based on their potential to improve efficiency, reduce costs, or enhance customer experience. Focus on areas where AI can deliver quick wins.

• Benchmark against adjacent industries: Look at how sectors like banking or healthcare are utilizing AI for similar challenges. For example, banks use AI for customer service chatbots, which can be adapted for insurance customer interactions.

There’s a guide from OpenAI that can provide inspiration.

When building your case for AI adoption, always link proposals to clear metrics or outcomes. Present a straightforward roadmap showing how an AI implementation can deliver incremental benefits quickly, then scale effectively. Keeping your proposals grounded, specific, and outcome-focused not only positions you as proactive and strategically-minded but also helps ensure your company’s AI investments truly enhance performance, rather than merely ticking a box labeled “AI-first.”

👀 AI Agent of the week: AI-voice-based candidate screening

Want to prepare for your next job interview? Or overwhelmed by too many applicants for a new job position? This HR AI agent performs the initial voice screening interview of the candidates. Simply upload a job profile, the candidate's LinkedIn profile, and the HR AI Agent will take care of the rest. You will receive a comprehensive summary of the voice conversation.

PRODUCTIVITY TOOLS AT HOME AND AT WORK

Language Lifesaver: Google's new "Tiny Lesson" experiment helps you learn exactly what you need, when you need it. Just describe a situation like "finding a lost passport" and Gemini instantly serves up relevant vocabulary, phrases, and grammar tips in your target language.

Secret Weapon: Cluely's controversial "invisible AI assistant" is making waves in sales circles. It secretly monitors your screen and listens to video calls, feeding you real-time responses to objections and technical questions through an undetectable overlay. Boldly marketed as "cheating" for professionals, it’s worth a try.

Brain Boost: Cerebro is transforming how teams manage knowledge with its AI-powered platform that turns scattered information into connected insights. It helps users organize ideas, access relevant data instantly, and dramatically cut down search time – making corporate memory work more like your brain's natural neural connections.

PROMPT OF THE WEEK

Make any report clear(er) for the target audience

This prompt makes the LLM revert to you with any inconsistencies or areas that are not clear to the common reader. Great when you want to share a report to a broader audience.

Prompt: You are a plain-language editor with broad, up-to-date knowledge of <INDUSTRY>.
**Before you start**
1. Confirm the two details below—then keep both front-of-mind during your review:
   • Target audience: <TARGET AUDIENCE>  
   • Topic of the report: <TOPIC>
**Your task**
Read the report as if you are seeing it for the first time. Assume no prior exposure to the specific subject beyond basic industry knowledge.

For every section:
1. **Flag unclear language**  
   – Highlight any words, acronyms, abbreviations, or phrases that a first-time reader might not grasp immediately.  
   – Allow legitimate <INDUSTRY> jargon, but still flag it if the meaning or context is not obvious.

2. **Assess readability and flow**  
   – Note sentences that feel long, complex, or disjointed.  
   – Point out any jumps in logic, missing links, or inconsistent terminology.

3. **Check cohesiveness and consistency**  
   – Identify conflicting statements, data mismatches, or style inconsistencies (spelling, tense, formatting).

4. **Evaluate exhaustiveness**  
   – Call out areas where key information is missing or where additional explanation, definitions, or examples would help the intended audience.

**Output format**
- List each flagged item in the order it appears:  
  *Excerpt* → *Issue* → *Suggested fix or clarification*  
- End with a concise summary (100 words max) of overall readability, cohesiveness, consistency, and exhaustiveness.

Return only the review—do not rewrite the report.

WHAT’S TRENDING

🤫 Rumor Mill: You can now forward any WhatsApp message to Perplexity for an instant fact-check. That could be especially helpful in a group chat where lots of misinformation is flying back and forth.  

🧑🏼‍🎓 Prompting Hack: Software engineer Sean Godecke discovered that if you tell ChatGPT your work was actually written by someone else, it can trick the chatbot into giving you more honest feedback.

🚨 Code Red: An eerie video is going around showing a robot going “berserk” during a test. One commenter speculated it might have been caused by a glitch where the robot tries to recover its balance but can’t, since it’s suspended in the air.

🪂 Safety Net: Armilla launched an AI liability insurance policy. It covers failures like hallucinations, model drift, and AI systems not performing as intended - addressing the growing "silent AI cover" problem where traditional policies leave businesses exposed.

AI IMAGES OF THE WEEK

💡 That’s all folks - tell me what you think

Your feedback makes me create better content for you! 🙏🏼

Until next week, Frederik, eData & the AI Agents

Your growth,
your career,
your future