Posts

🚦 When GPT-5 Felt “Dumber” — The Hidden Risks of Model Routing

A single “autoswitcher” glitch recently made GPT-5 behave differently.
It wasn’t just a tech hiccup — it exposed something deeper about how AI systems decide which model answers you.

In the latest blog, I unpack:

• What model routing actually is — and why it matters.
• The risks for everyday users, developers, and enterprise leaders.
• How routing extensions can reflect compliance, cost, and performance goals.
• Why routing is not just backend plumbing — but a lever for control, trust, and competitive advantage.
• Why model routing is a strategic capability every organization should own — not outsource.

If you’re building, buying, or scaling AI — this one’s worth your time.

Big models don’t win. Builders do.

Every week, a new model shows up—GPT-5, Claude Opus 4.1, and more.
Leaderboards light up, each release competing to outperform the last. They’re fast, powerful, and impressive.

But if history is a guide, that’s not what decides the winners.
When Java came out, it wasn’t the language alone that mattered.
It was the apps, the platforms, the ecosystem, the tools people built on top of it.

The race isn’t to have the biggest model. It’s to turn technology into lasting impact.

Models are just raw power.
The real value comes when someone shapes that power into something useful— something people use every day without even thinking about the model behind it.

The multi-platform architecture support in the latest release of OpenAI GPT-OSS enables it to run seamlessly across MacBooks (Apple Silicon) and NVIDIA GPU clusters—making it adaptable to both local and cloud environments.

I’ve started exploring the codebase, and one of the standout components is the native Metal backend for Mac—which enables efficient, on-device LLM inference using Apple’s low-level compute framework.

💡 The repository includes custom Metal kernels for core transformer operations:
✅ RMS Normalization – Normalizes using root mean square for better speed and stability
✅ RoPE with YaRN Scaling – Positional encoding that scales well for long sequences
✅ Scaled Dot-Product Attention – Core attention computation, optimized for performance
✅ MoE MatMul with SwiGLU – Efficient execution for sparse expert layers
✅ Expert Output Accumulation – Combines outputs from selected experts per token

🛠 These kernels are compiled via CMake into .metallib and exposed to Python through native bindings. Memory is managed efficiently using mmap and Metal buffers for weights and activations.

🌐 Beyond Metal:
🔹 Triton backends for NVIDIA GPUs
🔹 vLLM and Transformers compatibility
🔹 Developer tools via browser UI and Dockerized Python runtime

The native utilization of Mac hardware through Metal opens up real possibilities for running performant LLMs directly on consumer devices.

I was able to run this locally on a MacBook using the reference implementation for Metal for Apple Silicon.

Explore the repo 👉 https://lnkd.in/dj6hezFF

🚀 Just ran OpenAI’s new GPT‑OSS 20B model locally on my Mac M-series — and it worked seamlessly.

No GPUs. No cloud setup. No latency.
Just one line to get started:
ollama run gpt-oss:20b

🧠 For context — GPT‑OSS is OpenAI’s first open-weight model release since GPT‑2. It’s a major step forward for developers and teams looking to run large language models completely offline.

📦 First-time setup downloads a ~4.6GB quantized version:
ollama pull gpt-oss:20b

⚡ After that, it loads quickly and runs smoothly — making it one of the fastest ways to experiment with OpenAI-quality LLMs directly on-device.

🔍 Why this matters:

• Open-weight release from OpenAI (Apache 2.0 license)
• Optimized for laptops (≥16GB memory)
• No data leaves your device
• Great for prototyping, prompt design, and local copilots
• Works with agent frameworks and RAG pipelines

This opens new possibilities for building secure, low-latency, and cost-effective GenAI experiences — right from your own machine.

Next up: connecting GPT‑OSS to local agents and intelligent workflows.

Every system has a simpler version buried beneath complexity. Constraints help you find it.

Can we achieve the same outcome, faster, with reduced cost, minimal complexity, and lower environmental impact?
That single question reshaped how I approach building systems over the years.

When you constrain your mindset, you stop defaulting to brute-force solutions and start designing with intent. You begin solving problems in novel, intentional ways, with greater responsibility and purpose.This applies to pretty much everything — from software development to optimization strategies, and even how we make decisions in everyday life.

Solving with constraints isn’t just a technical principle. It’s a mindset shift that sharpens clarity and drives more responsible outcomes.

For instance, let me give an example with AI co-pilots. From my own experience, they often generate complex or inefficient code, even when a much simpler solution exists. They tend to default to common patterns, not necessarily the most effective ones. Only when I guide the AI with clear constraints do I consistently arrive at the most elegant and efficient outcome.

Constraint is not a roadblock.
It is a lens. A mindset. One that sharpens clarity.

Next time you’re solving a problem, whether with an AI model or a system design, try nudging your AI (or yourself) with questions like:
✅ What is the most efficient way to achieve this outcome?
✅ Can I reduce the number of calls, memory usage, or components?
✅ What would this look like if I had limited energy, time, or compute?
✅ Is this complexity necessary, or is there a simpler path?
✅ What hidden cost such as carbon, latency, or effort am I ignoring?

When was the last time a constraint led you to a better solution?
Share your story in the comments.

In the age of infinite scale, constraint is a competitive advantage.
Not because it limits what we can do.
But because it reveals what we should do.

Sam Altman’s recent interview gives a thoughtful and honest look at where AI is taking us. Two things really stand out: fear and regulations

⚠️ Fear is real—and personal:
🔹 Many people worry about losing their jobs, but this time it goes deeper. What happens when AI starts doing creative work too?
🔹 Sam talks about feeling strangely useless when a model answered something better than he could.
🔹 There are concerns about mental health, with people forming emotional bonds with AI, and how that might affect relationships.
🔹 The speed of change is making it hard to keep up. And no one really knows what comes next.

📜 Regulation needs to catch up—fast:
🔹 People share very personal things with AI, like they would with a therapist. But there’s no legal protection for those conversations.
🔹 Governments are starting to talk about rules, but things are moving quickly, and laws are slow.
🔹 There’s also a real worry about surveillance—that AI could be used to track people or invade privacy.
🔹 Sam suggests ideas like universal basic wealth, where everyone benefits from what AI creates—not just a few big companies.

💬 His words say it best: “You have to be both excited and scared.

👉 Watch out for this excellent video—one of the clearest views on where we’re heading with AI:
📺 https://lnkd.in/dmuf555e

This conversation closely aligns with the thinking I’ve shared in my book The New AI Engineering Mindset: Navigating Uncertainty and Opportunity in the Age of Intelligent Machines. AI is here to stay—and it will continue to evolve. It’s not just about tools or technologies—it’s about the mindset needed to navigate uncertainty, face fear, and still build, adapt, and move forward with clarity and purpose. (📘 https://amzn.to/4ocIOXj)

When advancing Agentic AI systems from prototype to production, these three operational realities define the boundary between experimentation and enterprise-scale impact:
❗ COST – Every prompt, tool invocation, and memory update consumes tokens—and dollars.
🌍 CARBON – Emissions don’t just come from training. Inference, planning loops, and ambient orchestration contribute significantly.
⚠️ COMPLEXITY – Adding more agents doesn’t equate to added intelligence. Without guardrails, orchestration grows fragile, memory inflates, and troubleshooting delays compound.

Making Agentic AI work at scale demands deliberate design:
✅ Define outcomes before deploying agents
✅ Optimize planning depth, memory scope, and tool calls
✅ Choose the right-sized models for each task
✅ Track cost, emissions, and operational metrics end-to-end

Agentic AI is not just a technical evolution—it’s a systems shift. Mastering cost, carbon, and complexity is what transforms potential into production.
For more insights, visit - https://leanagenticai.com/

Mistral just published one of the most detailed environmental lifecycle assessments for a large language model.

Their report on Mistral Large 2 transparently covers emissions, water usage, and resource depletion across training, inference, and hardware lifecycle.
📊 Key figures:
🌍 20,4 ktCO₂e for training over 18 months
💧 281 000 m³ of water consumed for training
⚡ 1,14 gCO₂e per 400-token inference

This level of transparency sets a new benchmark for what responsible AI should look like.

One of the key calls in their report:
🗣️ “AI companies ought to publish the environmental impacts of their models using standardized, internationally recognized frameworks.
And that’s exactly where the Green Software Foundation SCI for AI initiative steps in.
🌱 SCI for AI is working to define a consistent, neutral, and extensible framework for measuring the environmental impact of AI systems across the full lifecycle—from training to inference.

📉 But measurement is only one part of the equation.
✅ The real goal is reduction.
That means embedding carbon awareness into how we design, build, deploy—and even use—AI systems.

🔁 Everyone has a role to play:
🏗️ Model builders: Report full lifecycle impact and optimize model architectures
🛠️ MLOps & Infra teams: Track emissions, right-size compute, and enable carbon-aware scheduling
👨‍💻 Developers: Reduce prompt length, choose efficient models, and avoid over-inference
🎯 Product & AI owners: Consider environmental cost per feature or outcome—not just latency or accuracy
♻️ Sustainability leaders: Integrate standards like SCI for AI into reporting and governance, and drive reduction initiatives
📦 Procurement teams: Demand transparent impact data for models and APIs
🙋‍♀️ End users: Awareness matters. Help them understand how prompt length, frequency, and unnecessary queries increase emissions—especially with large models running in the background

As AI becomes more deeply embedded in life and business, it’s time to align innovation with impact.
👏 Kudos to Mistral AI for leading with action.

Now the ecosystem must follow—with transparency, standardization, and responsibility.
🔗 Read Mistral’s announcement - https://lnkd.in/duFcyRaw

Tokens are the new compute currency.
They seem cheap—fractions of a cent, a flicker of GPU time. But scale them across agents, retries, and escalations, and they explode into massive financial and environmental burdens.
💰 Pricing Reality (OpenAI API):
GPT-4o mini: $0.15/M input, $0.60/M output
GPT-4o: $2.50/M input, $10/M output
GPT-4.5: $75/M input, $150/M output

📊 For a 1,500-input + 500-output interaction:
Mini: ~$0.0005
Standard: ~$0.009
Premium: ~$0.19

At 1M daily users, that’s $15K/month (mini) to $5.6M/month (premium).
Now multiply that by agent chains, multi-model routing, and retries—costs rise fast.

🌍 The Carbon Shadow:
100M tokens/day = 30–1,000 kWh, enough to power 10–30 homes/day.
These emissions stay hidden in cloud dashboards—but they’re real, driven by compute-hungry infrastructure.

Agentic AI must be lean by design—efficient, intentional, and accountable.
🔗 Learn more at leanagenticai.com

Vibe coding needs a VIBE mindset—Validated, Integrated, Balanced, Effective.
In an era of fast AI-powered development, vibe coding feels great: fast starts, fluid ideation, and instant results.

But here’s what I’ve learned after evaluating top AI models for performance, security, and real-world readiness:
⚠️ Security Gaps – Generated code often misses critical checks, opening up avenues for misuse.
⚠️ Performance Inconsistencies – Benchmark scores rarely predict real-world stability or scalability.
⚠️ Complexity for Simplicity – Tasks that should be simple end up tangled in verbose or inefficient logic.
⚠️ No Trade-off Awareness – High benchmark performance doesn’t mean optimal cost, speed, or carbon efficiency.
⚠️ Lack of Holistic Design – A change in one part of the system breaks others—clear signs of siloed optimization.

If you’re vibe coding, expect an easy start—but be ready for unexpected detours unless you apply a VIBE mindset:
✅ Validated: Test everything. Never assume the output is safe or correct.
🔄 Integrated: Connect outputs to existing systems and policies, not silos.
⚖️ Balanced: Consider trade-offs across performance, cost, carbon, and maintainability.
🎯 Effective: Ensure your VIBE code is modular, extensible, and purpose-driven—built to evolve, adapt, and integrate with future workflows, not just to compile and run once.

Vibe coding is here. The term might evolve—but the need for a mindful approach is here to stay. 🙂

✨ From single prompt to Mini AI Apps - Google Labs has just launched Opal, an experimental new platform that empowers creators and developers to build powerful AI mini-apps using nothing more than natural language and visual editing.

With Opal, you can:
✅ Chain prompts, AI models, and tools into workflows – no code required
✅ Visually edit flow logic or tweak steps on the fly
✅ Share your mini-apps instantly using just a Google account

🔍 What Can You Build with It? Whether you’re:

• Prototyping AI ideas
• Building internal productivity tools
• Crafting proof-of-concepts for clients

🚧 Opal is currently in public beta – not yet available globally. Opal offers a glimpse into a future where AI development is accessible to everyone — not just engineers. This is how AI will scale: not through code, but through creativity.

🔗 Learn more: https://lnkd.in/dfFUwcxT

We consume more content than ever. But much of it has been summarized, adapted, or reshaped with the help of AI across multiple stages.

🎥 This short video explores the AI content loop.
A subtle cycle that influences what we read, share, and come to believe.
It’s worth asking:

Are we building real understanding, or just engaging with compressed versions of thought?

In a world moving fast, the real edge is clarity, depth, and intent.
That’s how you stay ahead.

We’re caught in an AI content loop. And most of us don’t even realize it. Much of what we read today is no longer original thought.

• It’s a reflection of a reflection.
• A summary of a summary.
• We write. Machines compress.
• We consume. Then rephrase again. The loop continues—until the nuance disappears.

You read a single answer and feel informed.
But are you?
Or have you simply accepted something that sounds like knowledge?

This isn’t just about AI tools. It’s about how our entire content ecosystem is shaped by shortcuts—and how quickly we move from source to summary, often without noticing what’s been lost.
📍 News: A detailed report becomes a short article, which becomes a social media caption, which becomes an AI-generated response.
📍 Research: A peer-reviewed study is reduced to an abstract, then to a headline, then to a single sentence in an AI-generated insight.
📍 Blogs & Thought Pieces: One idea is rewritten, summarized, paraphrased—until the output feels familiar, but disconnected from the original thought.
📍 Conversations & Podcasts: Rich discussions are clipped, quoted, and stitched into synthetic takes that sound right—but may no longer carry the original intent.

It’s efficient. It’s fast.
But when we rely only on compressed content, we risk losing depth, context, and perspective.

AI is here to stay.
And in a world driven by speed and summaries, the real edge comes from something simple:
Thinking clearly. Asking precisely. Creating with intent.
That’s how you stay relevant.
That’s how you move ahead.

How do you learn in the Age of AI?
Not just by reading or watching tutorials — but by engaging, questioning, validating, and refining your understanding.

Here’s how to use tools like ChatGPT, Gemini, or Claude to actively learn and grow — across any topic.
🧠 1. Set a Learning Path
🗣️ “I want to learn [topic]. Create a 3-week plan with key concepts, milestones, and practice tasks.
🗣️ “Now adjust this plan for someone with no prior experience.

🧠 2. Curate Smart Resources
🗣️ “For Week 1, suggest three free resources — a video, an article, and an interactive tool — to build foundational understanding.
🗣️ “Add one hands-on activity or project to apply what I’ve learned.

🧠 3. Understand Through Clarity
🗣️ “Explain [complex concept] using a real-world analogy.
🗣️ “Simplify it in under 100 words for a beginner.

🧠 4. Learn from What You See
📸 Upload a page or diagram from a book
🗣️ “Summarize this visually and explain the key insights in simple terms.

🧠 5. Practice and Apply
🗣️ “Create a scenario where I can apply this concept. Let me solve it and review my reasoning.

🧠 6. Review and Improve
🗣️ “Here’s my code/work. Review it for logic, quality, and performance. Suggest specific improvements.
🗣️ “What could be done differently or better?

🧠 7. Evaluate and Reflect
🗣️ “Test my knowledge with 10 questions. Score my answers and suggest areas to revisit.
🗣️ “What should I learn next to build on this?

⚠️ Note:
AI can speed up your learning journey, but it cannot replace critical thinking. Validate insights, question assumptions, and use your judgment — especially when outcomes matter.

Just remember, there are two ways to learn with AI.

1. One is to use it as a shortcut — to get quick answers, skip the hard thinking, and move on.
2. The other is to use it as a thinking partner — to ask why, explore how, and grow through curiosity and reflection.
   Choose wisely. One path upgrades your knowledge. The other just replaces it.

LangChain recently published a helpful step-by-step guide on building AI agents.
🔗 How to Build an Agent –https://lnkd.in/dKKjw6Ju

It covers key phases:

1. Defining realistic tasks
2. Documenting a standard operating procedure
3. Building an MVP with prompt engineering
4. Connect & Orchestrate
5. Test & Iterate
6. Deploy, Scale, and Refine

While the structure is solid, one important dimension that’s often overlooked in agent design is: efficiency at scale.
This is where Lean Agentic AI becomes critical—focusing on managing cost, carbon, and complexity from the very beginning.

Let’s take a few examples from the blog and view them through a lean lens:

🔍 Task Definition
➡️ If the goal is to extract structured data from invoices, a lightweight OCR + regex or deterministic parser may outperform a full LLM agent in both speed and emissions.
Lean principle: Use agents only when dynamic reasoning is truly required—avoid using LLMs for tasks better handled by existing rule-based or heuristic methods

📋 Operating Procedures
➡️ For a customer support agent, identify which inquiries require LLM reasoning (e.g., nuanced refund requests) and which can be resolved using static knowledge bases or templates.
Lean principle: Separate deterministic steps from open-ended reasoning early to reduce unnecessary model calls.

🤖 Prompt MVP
➡️ For a lead qualification agent, use a smaller model to classify lead intent before escalating to a larger model for personalized messaging.
Lean principle: Choose the best-fit model for each subtask. Optimize prompt structure and token length to reduce waste.

🔗 Tool & Data Integration
➡️ If your agent fetches the same documentation repeatedly, cache results or embed references instead of hitting APIs each time.
Lean principle: Reduce external tool calls through caching, and design retry logic with strict limits and fallbacks to avoid silent loops.

🧪 Testing & Iteration
➡️ A multi-step agent performing web search, summarization, and response generation can silently grow in cost.
Lean principle: Measure more than output accuracy—track retry count, token usage, latency, and API calls to uncover hidden inefficiencies.

🚀 Deployment
➡️ In a production agent, passing the entire conversation history or full documents into the model for every turn increases token usage and latency—often with diminishing returns.
Lean principle: Use summarization, context distillation, or selective memory to trim inputs. Only pass what’s essential for the model to reason, respond, or act..

Lean Agentic AI is a design philosophy that brings sustainability, efficiency, and control to agent development—by treating cost, carbon, and complexity as first-class concerns.
For more details, visit 👉 https://leanagenticai.com/

#AgenticAI #LeanAI #LangChain #SustainableAI #LLMOps #FinOpsAI #AIEngineering #ModelEfficiency #ToolCaching #CarbonAwareAI LangChain

Google Gemini now lets you generate short, dynamic videos from a single static image, using Veo 3 technology. 🎨 What you can do with it:

• Bring travel photos to life with ambient sound and motion 🌊
• Animate illustrations or designs with matching audio 🎧
• Add motion and narration to social media posts or presentations ✨

📌 How it works:

1. Open Gemini (web or mobile app)
2. Select “Video” from the prompt bar
3. Upload an image
4. Enter your prompt
5. Receive an AI-generated 8-second video

ℹ️ Note: Google Gemini will automatically add a visible watermark to indicate the video was AI-generated, along with an invisible SynthID digital watermark for traceability.

📍 Currently available for Gemini Advanced (Pro) subscribers in select regions.

The video attached here was generated using two static images. There’s a message in it — hope you all get it. 🙂

#Gemini Google Cloud #Veo3 #PhotoToVideo #AI #GenerativeAI #ContentCreation #FutureOfWork #Efficiency #CreativeTech #SynthID

When software begins to think, plan, and act, how do humans stay ahead?

Today, we mark 50 episodes of Agentic AI: The Future of Intelligent Systems, reflecting deeply on our relationship with software that no longer just computes but genuinely thinks.

From autonomy and orchestration to sustainability and alignment, each conversation brings us closer to answering one critical question:

How do we build intelligent systems that amplify human intention, rather than override it?
🎧 Episode 50 flips the challenge:
How do we out-think Agentic AI?
📍 Spotify → https://lnkd.in/dpE3BqKu
📍 Apple Podcasts → https://lnkd.in/dpfiyWDK

Thank you for listening and thinking alongside us.

Around four years ago, the Green Software Foundation Standards Working Group started as a small group of us meeting every week. Through small, consistent actions, regular check-ins, GitHub updates, and collective thinking grounded in inclusivity and consensus, we shaped what would become the world’s first ISO specification for measuring the carbon emissions of software: the Software Carbon Intensity (SCI) standard.

What began as a modest effort steadily grew as more organizations came on board to contribute. Participation expanded, ideas multiplied, and today the Green Software Foundation has evolved into a thriving community of over 60 member organizations, all united by the goal of making software part of the climate solution.

Building on this foundation, we are now extending the specification to Artificial Intelligence through SCI for AI. We recently hosted the first SCI for AI workshop, bringing together over 20 contributors from across industries, all driven by the shared goal of creating consistent, transparent, and actionable metrics to measure the environmental impact of AI systems. Here is the link to the workshop report, which captures the key discussions, outcomes, and the path ahead: https://lnkd.in/di_8ebY5

To explain to my daughter what I do on these weekly calls, I tell her it’s like coming up with the formula for calories on food labels, which she related to instantly. Once something becomes visible, people become more aware and can make better choices.

A big thank you to all the members who have been participating, contributing their time and expertise every week, and to the organizations supporting them to help take this important work forward.

I am fortunate to chair the Standards Working Group and help steer the SCI for AI initiative. Together, we are making the invisible visible by exposing emissions that are otherwise hidden and unaccounted for in the digital systems we use every day, and by providing ways to reduce them. Making emissions measurable is the first step toward managing their impact and driving reduction.

Let’s make Sustainable AI the new norm.

In the Agentic era, choosing how to think is as important as what to think.
Agentic AI systems don’t just act—they reason.
But not all reasoning is created equal—and getting it wrong can waste both cost and carbon.

Here’s a simple framework for reasoning optimization in Agentic AI:
🔹 Stage 1: Close-Ended Reasoning
✅ Deterministic answers (yes/no, select one)
✅ Lowest cost
✅ Lowest emissions
✅ Ideal when the workflow is clear, the outcome set is limited, but AI is still useful for interpreting unstructured inputs.

🔹 Stage 2: Pre-Compiled Thought
✅ Curated, finite responses from precomputed knowledge
✅ Moderate cost
✅ Low emissions
✅ Use when the space of possible answers is finite and repetitive—and full creative generation is unnecessary.

🔹 Stage 3: Open-Ended Reasoning
✅ Generative, exploratory, creative thinking
❌ Highest cost
❌ Highest emissions
✅ Reserve for tasks that demand creativity, ideation, or complex synthesis.

👉 The path to lean, responsible AI is not just about selecting the right model—
It’s about selecting the right type of reasoning at the right time.
Smarter reasoning choices drive:
✔ Lower compute
✔ Lower cost
✔ Lower carbon
✔ Better outcomes

Most Agentic AI conversations are missing a key dimension: cost, carbon, and complexity.

While the spotlight is often on autonomy, orchestration, and innovation, the reality is that Agentic AI systems—if not designed intentionally—carry hidden risks that quietly erode value:
❗ Vague goals that trigger unnecessary actions, retries, and compute waste
❗ Over-planning and decision loops that burn resources without meaningful benefit
❗ Overuse of large models when smaller models would suffice
❗ Redundant tool calls and uncontrolled memory growth
❗ Silent system inefficiencies that drive up cloud costs and emissions without notice

As most organizations are experimenting with or just getting started on Agentic AI, this is the right time to embed efficiency, sustainability, and cost-awareness at the foundation—not as an afterthought.

That’s why I wrote Lean and Green Agentic AI—a white paper with a practical framework for building AI that is not only intelligent but also efficient, scalable, and economically viable.

The paper introduces:
✅ The six-stage Agentic AI lifecycle
✅ Lean principles to minimize cost, carbon, and complexity
✅ Practical techniques for energy-efficient models, inference, and carbon-aware execution
✅ A standardized approach to measurement using the Green Software Foundation Software Carbon Intensity (SCI) framework and its AI extension
📄 Access the full white paper here:
👉 https://lnkd.in/dU_zHHXg

The agentic future is coming fast. Let’s ensure it’s smarter, greener, and built to scale responsibly.