To think big about AI's potential, start small

Could small language models plus slow and steady actions move you faster?

Bigger isn't always better in AI. Colossal AI benefits are greatly overstated, leading to eye-wateringly big investments in data centres. As I’ve said before:

If you're not working at the bleeding edge of tech, you don’t need to bleed.

Despite talking about GenAI with many organisations, I rarely meet any that think they’ve hit their stride. For the bigger (or rather, noisier) ones, tangible results are hard to separate from hype.

The problem: We're told AI's possibilities are endless. So, where to start?

Daily, the noisier AI evangelists shout "Think big, people!", like Shopify's CEO, who wants to crush it with AI. Those pesky dissenters who want more resources to deliver more can wait. AI is the answer! Now what's the question?

This big-AI-at-any-cost play might earn 'hoorahs' from the C-suite. But for your people, it can spark fear. How can we begin to imagine all the possibilities when we're just learning to use LLMs for basic summarising, alongside half-functioning pilots delivering bits and pieces? For most knowledge workers, the gains aren’t there to "unleash productivity" beyond squeezing in a full instead of a half sandwich between back-to-back Zooms.

"Do more with less" or "be productive with AI" isn't empowering workers forced into offices (asking themselves, “if AI is my co-worker, why am I here?”). The plan and message need to be more specific to succeed.

Start with the end in mind by setting pragmatic goals

To engage people, encourage them to think big by acting small with determination.

Ad legend David Ogilvy said:
"Give me the freedom of a tight brief".

Guardrails create the best work. LLMs are infinite monkeys with typewriters who've consumed humanity's knowledge for training. That's quite some working lunch.

Recently, I've shifted from a limiting mindset ("the world is doomed, take what you can get") to abundance thinking ("reap what you sow, gold in them there hills"). "Wealth influencers” sometimes use these Jedi mind tricks to sell you get-rich-quick myths. But for me, it helps envision the ideal, or "begin with the end in mind," then plan practical steps forward. My artist friend calls it "planting a flag in a day" when you've done *something* that moves you an inch further. Abundance thinking works best with narrow and actionable steps towards the goal.

For your AI adoption, what's a realistic end goal for this and next year? Is it "AI-first" like Shopify and Duolingo, digital-native consumer brands? Perhaps your B2B medical manufacturing firm needn't go all-in now. Instead, start optimising the most common task flows.

Could your goal be getting everyone to base-level knowledge with an AI literacy plan? Only 1 in 4 organisations are planning any AI training this year, so doing *something* puts you in the top 25%. As AI literacy is now mandatory under the EU AI Act, it's about compliance too.

Rather than saying "everyone should be using AI", which will soon sound as bizarre as "everyone should be surfing the web" did 20 years ago, be purposeful. If you have a licensed LLM, add specific learning objectives to development plans.

The objective becomes tangible, like: "learn and practice using Copilot by completing training modules plus 10 hours of self-guided learning". (Other LLMs are available. Copilot isn't terrible, though. Yahoo! got you to web pages, in a fashion, before Google).

As AI embeds into everyday tools, expecting people to rock up with AI skills will become moot. Beyond tech teams developing AI models, encourage a willingness to learn with the provided tools and systems. That means more investment in tech and training before reaching any productivity holy grail.

Give them just enough AI

Careful Trouble's Rachel Caldicott encourages us to think about limiting how much we use LLMs. While measuring token use isn't easy, you could indicate an appropriate number of prompts per week by role. With free LLMs, you'll know when you hit your token limit.

Limitations can benefit us. They make cognitive thinking mandatory, preserving our grey cells, as LLMs could be rotting our brains.

On the flip side, some firms restrict LLM access based on need or seniority. Understandable given stretched IT budgets and wasted licenses. We can't afford another "everyone needs Photoshop" just to make clip art.

But soon your people will expect the latest tech stack, or their skills will stagnate. Once the economy improves, without proper tools and training, they'll leave.

Depending on your organisation's needs and culture, providing basic Copilot licenses to all might be more democratising than deciding who is deserving of an enterprise OpenAI license. Without tools, shadow AI will flourish as people bring their own, or use less secure free LLMs.

Another approach: provide licenses but limit use cases to manage data and carbon outputs. Settling on a company-wide LLM strategy is challenging. Veolia offers its 220,000 employees a choose-your-own-AI-model where features vary by role. All get summarisation and text creation, but only some get the token-heavy image creation. For enterprises, token-based billing may beat per-seat licensing for flex and efficiency.

Small language models are getting big

While AI vendors trumpet that their models have trillions of parameters, a quiet trend emerges. Small Language Models (SLMs) prove that, like much in life, the best things come in small packages.

The major LLM providers (OpenAI et al) dominate our attention. Yet the real money might be in AI wrappers, functional tools linked to existing LLMs. And the next evolution: AI agents. Tools like Manus, Operator and Kairos target specific use cases – like auto-creating and sending invoices – using narrow datasets.

As DeepSeek R2 hovercrafts into view, a model delivering the same 'gist' as big LLMs but using a fraction of the resources has big AI quaking in their mainframes.

And for some industries, "drop an AI from a great height" isn’t the right strategy for customers. The engineering excellence team at Lloyds Banking Group prioritises security and data privacy in tech innovation. AI is shaping behind the scenes with testing, caution and caveats. But it's not the reason to trust your bank with your money.

Small and steady wins the race

SLMs typically contain millions to a few billion parameters, compared to hundreds of billions in GPT-4 or Claude. But what they lack in size, they make up for in focus and efficiency, seemingly performing the impossible through specialised tasks coupled with distilled knowledge. They can run entirely on-premise or on edge devices, making them more secure. For regulated industries like healthcare or finance, this changes the game.

Smaller models gained momentum when Meta released Llama 2 in late 2023 with impressive performance results. Since then, models optimised for specific tasks have exploded. Some examples:

Llama 3 is now being tested at the International Space Station. This leaner model could manage tricksy data flows from space to Earth.

Why process documents using external APIs when you can run a smaller model on your infrastructure? Perplexity AI and Vectara deliver more efficient RAG (retrieval-augmented generation) methods to source from internal knowledge.

Developers don't always need the most powerful models, but they do need ones that understand their codebase and respond quickly. GitHub's Copilot shows smaller models can boost productivity (a fair clip anyway, 26% according to the vendor's survey).

My trusty LLM birthed this lovely analogy:

You wouldn't use a sledgehammer to hang a picture frame.

Indeed not, though you might foolishly use one to crack a nut, Claude. Same with language models: you don't need a Swiss army knife LLM for "everything". The Pareto rule applies: We spend 80% of our time doing 20% of tasks, so let's develop tools for that crucial 80%.

AI has spoiled us with expectations of immediacy (or was that Amazon Prime?). Models fine-tuned on your corporate knowledge often outperform general-purpose giants, making them ideal for real-time applications like customer service chatbots or fraud detection.

Over the next year, expect more industry-specific SLMs packaged as AI agents for specific use cases and industries. Good news: This makes advanced AI accessible to companies of all persuasions and IT budget sizes.

How to think big about smaller AI

Plan for ‘just-in-time’ AI.

Consider how AI fits your processes. Define repeatable systems before writing use cases. The more specific your needs, the more a SLM could help. General Q&A with unstructured data might need larger models, but for defined tasks, smaller models may work better.

Take the slow and steady path to AI adoption.

Rather than startling the horses with return-to-office mandates paired with “AI is your new coworker”, take a people-friendly approach. Globally, more of the public opposes (35%) than supports (30%) AI currently. This likely includes some of your staff.

It's OK to WAIT. That's Working on AI Transformation. Slow, AI transformation that works triumphs fast AI that falls over and breeds mistrust.

AI for forward-looking organisations today isn't about “my AI’s bigger than yours” or “my AI-first transformation is thrusting forward harder”. It’s not the size but what you do with it that counts. (But you knew this already, right?).

It's about finding the right tool for the job. Sometimes that's using a small screwdriver to tighten the tiny screw causing the big leak.

Cast many small pebbles in the water and see which ones create the biggest ripples.

What's your experience working with small language models and managing AI rollout? This post is free. Please share with anyone who may enjoy it and add your comments.