THE AI LOOP
This week reinforced a view I’ve been forming for a while: software engineering is no longer the main constraint.
If you have an idea, it can be turned into a working product quickly. Tools like Claude Code, Codex, and Cursor have changed how building happens.
If you’re an exec with 10 ideas to test, you can now prototype, validate, and drop ideas faster and at scale. Execution is less scarce. Judgment about what to build is now the limiting factor.
There is still a gap between people who use AI tools daily and people who make decisions about AI strategy. Sitting with someone who knows these tools and watching an idea turn into code in a few prompts changes how problems are approached.
ARTIFICAL INTELLIGENCE
AI Is Reshaping Entry-Level Tech Jobs
Hiring for traditional entry-level developer roles is down about 25% at large tech firms. Software demand has not dropped. Routine work is being automated.
Roles seeing growth include:
Information security analysts
AI engineers
Systems roles around automation
Entry-level roles often existed to handle repetitive tasks. That layer is shrinking.
Why it matters:
If agents handle execution, humans need to contribute through system design, evaluation, and judgment. Education and apprenticeship paths will need to adjust, or skill gaps will grow.
SOCIAL MEDIA
Instagram Says “Polished” Is Dead
Adam Mosseri (CEO Instagram) says AI-generated content has undermined the curated aesthetic the platform was built on. What is being emphasised instead:
Private messages
Candid photos
Greater reliance on who posted content rather than how it looks
Instagram is now:
Labeling AI-generated content
Showing more account-level context
Exploring cryptographic signing at capture time
Why it matters:
Large volumes of generated content reduce signal quality. Platforms are adding verification and context layers to compensate. Similar issues apply to other information systems.
LLM RANKING
LMArena 2025 Results: Models Are Converging
LMArena, a community platform for comparing and evaluating AI models, released its 2025 rankings:
Gemini 3 Pro by Google leads in text, vision, and search
Veo 3.1 by Google ranks highest for video generation
Performance differences between top models are narrowing.
Why it matters:
Model selection is becoming less important than how models are combined, tested, and deployed.
AI FUTURE
IBM: What Actually Shapes AI in 2026
IBM’s 2026 outlook focuses on:
Agent-based workflows inside enterprises
Stronger governance and evaluation
AI built into business processes rather than added on
Less emphasis on demos. More focus on automation that reduces cost or time.
AI TIPS
A Practical Way to Make AI More Creative
New research describes a prompting method called CreativeDC that increased creativity scores by 51–63%. The method separates idea generation into two distinct steps.
Step 1: Diverge first
Ask the model to generate ideas without constraints.
Prompt: Generate as many different approaches as possible to solve this problem.
Do not evaluate quality, feasibility, or realism.
Focus on variety.
Step 2: Converge later
Apply constraints only after ideas exist.
Prompt: From the ideas above, select the top 3 using these constraints:
– Feasible with current resources
– Directly addresses the problem
– Can be implemented within a fixed timeframe
Briefly justify each selection.
The study reports three reasons this works:
Models default toward average outputs
Early constraints reduce novelty
Exploration before filtering changes results
📄 Paper: https://arxiv.org/abs/2512.23601
One Thought to Sit With
AI has reduced the effort needed to make things.
It has not reduced the difficulty of choosing what to make.
The next advantage will come from people who:
Choose problems carefully
Evaluate results honestly
Design systems where humans and agents work together
Until next week,
Asim - The AI Loop