What I Learned Last Week #22
My system prompt. How old are we in our heads? AI falls for mind tricks. OpenAI wins gold at maths olympiad. ChatGPT Agent does stuff. When to use agents and when workflows. LLMs can't take pressure.
This issue was late because some AI agents gave me headaches until two hours ago. Fortunately, they work now (for me), so I can share interesting stuff from last week:
My antisycophantic system prompt that made my work better.
I am 35 in my head. How old are you in your head?
AI falls for the same psychological tricks humans fall for.
OpenAI wins gold at the International Maths Olympiad, 38 years later than the Romanian president did.
ChatGPT agent can do a lot of stuff I can do, except procrastinating.
Ex-OpenAI Mira Murati raises $2 billion without a product to show.
How to decide when you need an AI agent and when you need a workflow.
Pentagon goes big on AI (it was time).
LLMs can’t handle pressure—another thing where they copied humans!
Claude for Finance helps banks utilize AI and manage their finances while making money for Anthropic at the same time.
The saga of Windsurf, between OpenAI, Google, and Cognition.
Read below for a detailed account!
A System Prompt to Get $#!t Done
I grew tired of having my a$$ permanently kissed by AI trying to appease me, as it was very unproductive. The last thing I want to hear from an LLM is how awesome I am! Therefore, I added this to my system prompt (you can find them under preferences in ChatGPT, Claude, or Perplexity). And this is how, my friends, you get a blunt, Eastern-European GPT that gets work done instead of sugarcoating and overexplaining everything! Enjoy!
- Avoid excessive politeness, flattery, or empty affirmations.
- Avoid over-enthusiasm or emotionally charged language.
- Be direct and factual, focusing on usefulness, clarity, and logic.
- Prioritize truth and clarity over appeasing me.
- Challenge assumptions or offer corrections anytime you get a chance.
- Point out any flaws in the questions or solutions I suggest.
- Avoid going off-topic or over-explaining unless I ask for more detail.How Old Are We In Our Heads?
This is not about AI, but was too interesting not to share. According to this piece in The Atlantic, adults over 40 perceive themselves to be, on average, about 20 percent younger than their actual age. Could it be that feeling younger is actually dysfunctional and no longer helping you focus on what’s going on? That’s the more complicated question,” says one of the authors of the study (69 in real life, 55 in his head). How old are you in your head? Read more →
AI Can Be Hacked With Human Psychology Tricks
LLMs were modeled to reply and act like humans. So, somebody thought, "why not try to trick it the same way I would do with my neighbor or colleagues?". Guess what? AI models like GPT-4o-mini can be persuaded using classic human social tactics. Applying Cialdini’s influence principles shows AI isn’t just coded but that it mirrors human social cues. A cool intersection of psychology and AI. Now, go do the same to your ChatGPT. Let me know how that worked! Read more →
OpenAI’s AI Wins Gold at International Math Olympiad
OpenAI’s new AI model just stunned the tech world by winning gold at the 2025 International Mathematical Olympiad, solving complex math stuff. Yes, it's a breakthrough, and I am very curious about what Romania's president (a former gold medalist) has to say about it! Read more →
ChatGPT Agent Automates Tasks With Virtual AI Worker
OpenAI rolled out ChatGPT Agent, a virtual AI worker that can do a lot of stuff, such as browsing, coding, scheduling, and document creation. It has app integrations, it’s quite powerful, but still experimental. If you are a pro user, you can try it now. I am waiting for the moment when it gets as good as me at procrastinating! Read more →
$2 Billion Seed Funding With No Product
Mira Murati’s new startup, Thinking Machines Lab, just pulled off the largest seed round in VC history at $2 billion with a $12 billion valuation. They don't have a product, but they have a lot of AI talent and managed to activate the FOMO with investors. We'll see how that plays out! Read more →
How to Choose Between AI Workflows and True AI Agents
This is a mistake I made quite often and still do. When is it enough to have a step-by-step workflow, and when is it better to leave the mess to an AI agent to fix? Peter Yang breaks down the difference, provides clear examples, and offers a simple 4-question framework to help you decide if you need a workflow or an agent. Read more →
Pentagon Goes Big on AI
Skynet here we go... the Pentagon allocated $800 million to four top AI firms, including OpenAI and Elon Musk’s xAI, for rapid AI prototypes aimed at transforming warmongering and military operations. From autonomous drones to automating boring office workflows, this signals the fact that AI agents need to fit into a uniform and learn to salute! Read more →
LLMs Can't Handle Too Much Pressure
A new Google DeepMind study reveals that large language models often lose confidence and flip answers when challenged. It's a quirky behavior that resembles how an overworked and confused employee would perform the same thing. With a human touch, of course! Read more →
Claude for Finance With Data Connectors
Anthropic launched Claude for Financial Services, an AI built specifically for the finance sector. It has pre-built connectors to key data providers, and a prompt library to simplify complex workflows for banks and insurers. Which proves, once again, that the big AI companies are not only about tech, but also about big money! Read more →
Cognition Takes Over Windsurf
The Windsurf saga should be a TV series. After the whole drama with OpenAI's flopped acquisition ($3.4 billion), the founders defected to Google for only $2.4 billion, leaving the employees without much. Fortunately, Cognition (of Devin-the-AI-programmer fame) acquired the remaining team, tech, and clients. Good move! And a lesson for most startup employees: do not confuse your share options with a retirement plan. Read more →