Absolutely, noticeable even via the API in other tools — for months now imho! Best time to code with agents is in the mornings CET.
PROMPT ENGINEERING
-
Building Efficient AI Agents: Memory, Tools, and Planning
By
–
Towards Efficient Agents. This is a great read if you are an AI dev building agents. Great survey on how to build efficient agents and leverage memory, tool learning, and planning. There is also a great discussion on evaluation and costs, which are important aspects of
-
Cold prompting floor, context engineering compounds, ChatGPT vs Claude
By
–
exactly, this study tested cold prompting.. no memory, no history. you're describing context engineering. that 49% prompting effect is the floor. stack persistent context on top and it compounds. chatgpt optimizes for "user feels good." claude with proper context optimizes
-
Prompting skills determine half of model’s value potential
By
–
What this means for you: → Model upgrades alone deliver only ~50% of potential value → Your prompting skill determines the other half → Learn each model's specific capabilities → Communication > technical background
-
Prompting evolves as key to unlocking AI potential
By
–
The researchers' conclusion:
"As models continually improve, people will continually respond by adapting their prompts to take advantage of new capabilities." Prompting isn't dying. It's becoming the mechanism by which humans unlock AI potential. The future isn't just better -
Exploration vs Exploitation: Pivot or Iterate in Prompting
By
–
The exploration/exploitation pattern: When previous output was bad? Exploration helped (trying new approaches) When previous output was good? Exploitation helped (refining what worked) The best performers knew when to pivot and when to iterate. Prompting isn't just
-
Crossover test reveals prompt adaptation is model-specific
By
–
The crossover test proved adaptation was model-specific. When DALL-E 3 prompts were played on DALL-E 2? Performance difference nearly vanished (p=0.56). Users had unknowingly learned to write prompts that exploited DALL-E 3's unique capabilities. The prompts were tuned to
-
DALL-E 3 users learned to exploit capabilities without instruction
By
–
The performance gap widened over time. As participants made more attempts, DALL-E 3 users pulled further ahead (β=0.0010, p=0.023). They were learning to exploit its specific capabilities. Without anyone telling them.
-
Users adapted to better model with longer, descriptive prompts
By
–
How did users adapt without knowing they had a better model? → 24% longer prompts (6.9 more words on average) → Same ratio of nouns/adjectives (48% vs 49%) → More semantically consistent across attempts Translation: they added more descriptive information, not filler.
-
AI-assisted prompting reduces DALL-E 3 performance by 58%
By
–
The third group had GPT-4 automatically rewrite their prompts "to help." Performance dropped 58% compared to baseline DALL-E 3. Read that again. AI-assisted prompting made results WORSE than no assistance at all.