Efficiency

1. What’s new in AI

A few recent well-publicized developments in Large Language Models (LLMs) that have shifted a previously predictable direction of travel into something more exciting.

These were:

• DeepSeek’s R1 model
• OpenAI’s Deep Research
• The fact that multimodal (pictures and text) LLMs like OpenAI’s 4o are now the norm

Putting aside for a moment the broader geopolitical implications, DeepSeek’s new R1 model represented a credible challenge to the assumption that to continue improving the smartness of models requires an almost infinite increase of AI datacenters. R1 demonstrated that less capital-intensive approaches can yield comparable results, but it did so by using previously developed models in its’ development, by a process known as distillation. This recycling of models is not such a bad thing, as “Chat LLMs” is a category that is already somewhat saturated, and approaches used to train language models from scratch could be reaching inherent capacity limits, if it turns out that most available language data has run out.

Then there’s the reinvention of AI into “agents”, which are mostly scaffolding around LLMs’ ability to do next-token-prediction, simulating a kind of continuity of intention, ideally up to fulfillment of that intention. Ironically, this construction arguably is not true agency, although it can be really useful as an automation. And automation is what sells Deep Research: given a prompt, it produces a report made up of several autonomously-selected sources, traversed and collapsed into one summary with references. The value turns out to not so much be novel insights, for which the prompt is the thesis, but the formatting into a cohesive whole. This is closer to what could be the enduring usefulness of language models (see GPT Use-Cases) of translating one format to another - rather than a true copilot, LLMs run on plain language instructions to do something, and the discretion of what that is still rightfully lies on the user.

So while these tools represent a meaningful evolution of the current AI landscape, a little bit of perspective is warranted - writing out the instructions and checking them against a goal feasibly requires more context than first appears. But under such direction, there is a definite force-multiplication underway, especially when adding in the multimodal capability of the latest models, allowing for applications that are transverse across different interfaces; say, giving verbal directions from a map, drawing a complicated step of an instruction guide manual, etc…

2. Long and short term horizons

The impact of AI on education and labour are already shown to be pretty drastic, but the promise of one or another product developed by one organization can fade pretty quickly (Remember Pi?), or become the defacto standard (OpenAI), or vanish entirely (Neeva).

That leaves a lot of space for the anticipated proliferation of open-source language models that could be purpose-built to pursue strategic, non-consensus goals, such as integrating externalities into the business model and comprehensive accounting of impact of activities.

The shift of mindset from risk management to meeting the inevitable uncertainty with conviction is human-driven, but having an on-demand knowledge base to validate ideas, translate larger policy targets into actionable goals for a company, or be directed to do administrative tasks like write compliant reports. This potential enabler of efficiency that makes novel operating models become viable, combined with other new technologies to be covered in a future article. When it comes together, it might look strange, as new paradigms do at first.

Bottom line

The unlock that truly helps businesses embrace sustainability for their own prosperity and that of their stakeholders will not show up as a tool advertised to do this as such, but more a gradual increase of capabilities exemplified in R1 and Deep Research, to be assembled around the direction of founders and operators with vision, and the desire to try something new.

What’s more, AI/LLMs within their current moment of sustained interest can be a momentous vehicle pushing towards the reinterpretation of certain accepted orthodoxies. It will be interesting to try and imagine how this plays out while many ideals and goals for the future are in question.