02.07.2025

ForwardThinking

The language problem with large language models

When I launched this newsletter at the beginning of December, it was never my intention to make every commentary about AI. It just seems to have worked out that way. There is just so much to talk about, and as I noted in this space last week, we are dealing with a constantly moving target. This week, I want to talk about the language component of large language models.

It may seem obvious enough, but how you communicate with these models matters, and you’ll learn as you begin using them more often that precision matters, writing skills matter – and that may be the most surprising thing I’ve learned about LLMs.

I know I shouldn’t be surprised that language is an important part of interacting with a model built on language, but it turns out that communicating what you want isn’t as easy as it sounds. We already know how important precision is in human-to-human communication, even if we aren’t always good at it. Think about how many great books, movies or plays are based on a miscommunication that turns tragic or disastrous – just about every Shakespeare play – and that’s because it turns out communicating is hard.

Consider this example from an event in December: I asked several people I met at an event happy hour where they were from. I was thinking about what company they worked at, but I got several answers including the cities where they currently live, where they grew up, and perhaps most hilariously from one man who answered, his mother.

It’s a simple enough question, right? Yet there were a myriad of answers and any one of them could be correct. Now imagine, you aren’t talking to a person standing in front of you with all of the subtleties, context clues and cultural nuances of human communication, but are instead having a conversation with a bot powered by a large language model. Then it gets even trickier. I’ve learned that the more precise your prompt, and the more information you give the model, the higher quality the answer you get. Seems logical enough, but I write for a living. My job is to communicate clearly in writing.

Now consider the fact that many people don’t write well, and suddenly we are being told that in the future writing well won’t matter because LLMs will make that skill obsolete. Evidence suggests, however, that the exact opposite is true. In the future, writing well could matter even more because you have to tell the bot exactly what you want it to do, and if you don’t give it enough information, it’s not going to give you the results you want.

This could become compounded with AI agents, AI systems that autonomously execute tasks, requiring even more precise instructions than chatbots. That’s going to require some expertise about how the process works and, yes, an ability to write well, to communicate that to the agent.

If humans struggle with ambiguity, working with LLMs could amplify that problem. Consider that engineers often trained in math and logic versus narrative writing, may face steeper learning curves. When the process of creating a program could be abstracted to describing what you want, and having the AI create it for you, suddenly software development could be less about math skills, so much as writing. That’s mind blowing really.

Regardless, we have created these models built on language, and in the process we have made the process of communicating even more essential than ever before, and that’s an outcome I didn’t expect.

-Ron

What's new on the blog

How Belinda Neal helps keep Goldman Sachs’ engineering team on track

As a technology leader at Goldman Sachs, Belinda Neal has to operate within the constraints of a highly regulated industry, yet she continues to find ways to innovate.

Here's how she sees implementing AI at Goldman:
"AI allows us opportunities to create more efficient business processes. It allows us to enhance our client experience, and then more broadly with the developers, it allows us to make our engineering groups and communities much more effective at developing technology at a rapid rate."

Read the full article

Here’s what’s holding back your generative AI project

It's fair to say that most enterprise companies are working on generative AI applications, but studies continually show that the vast majority are struggling to get from experiments and PoCs to scaled projects.

I looked at three recent studies from Deloitte, BCG and Kearney-Futurum to find out what the key factors were that are holding companies back.

Read the full article

Analysts express skepticism over DOJ suit to stop $14 billion HPE-Juniper Networks merger

I spoke to a couple of analysts to get their take on the DOJ lawsuit to stop the $14 billion HPE-Juniper Networks deal, and they both felt the DOJ might be missing the point of this coupling.

Read the full article

News of the week

Cloud infrastructure market reached $330 billion last year

The big three cloud vendors -- Amazon, Microsoft and Google -- reported earnings this week, and the cloud results were mostly good. For what it's worth, each of the the big three actually missed their Wall Street revenue projections slightly, but the market still reached $91 billion for the quarter and $330 billion for 2024, just mind boggling numbers (especially for companies that continue to lay off employees).

Synergy Research, a firm that follows all things cloud, reported that the totals for the quarter were up $17 billion or 22% over the prior year. If you're wondering about market share percentages, Amazon led the way with 31%, followed by Microsoft with 21% and Google with 12%, per Synergy.

Yet all of the vendors reported that in spite of growing demand, mostly due to the resource requirements of generative AI, they were constrained by a lack of datacenter capacity and chip supply shortages. To alleviate the datacenter component, the three companies are planning massive spending sprees this year with Amazon shelling out $100 billion and Microsoft and Google announcing plans to pony up $80 billion and $75 billion respectively.

OpenAI Deep research AI Agent could be big leap for researchers

OpenAI surely likes to keep its name in the news, and seems to be releasing something new every week. This week it was the Deep Research AI agent, which can conduct complex research on the web, reducing hours of human effort into one that can be done at the speed of a bot.

Wharton professor and AI advocate, Ethan Mollick, writing eloquently about the new agent, says it does a credible job of doing academic-style research, but also points out that doesn't mean just anyone can use this effectively. It still requires someone who understands the subject matter to write a high quality prompt, and to push back at and poke the results.

While keeping in mind this is a very early release, and it could run into substantial roadblocks at scale, the fact it can produce a set of working results in fairly short order has the potential at least to change the nature of research in the future.

Enterprise software companies announce a slew of new layoffs

Another week, and more downsizing at big tech companies. Workday announced it was laying off 1750 people this week, Salesforce 1000, and Okta 180. As these companies continue to try to squeeze out operating cost savings, their employees are paying the price. The layoffs come as enterprise software vendors are making big, expensive bets on artificial intelligence.

Vercel announces new approach to serverless built for the AI age

Serverless computing has always held the promise of using just the right amount of computing to complete a job and no more, and while there were trade-offs for that efficiency, it has provided a way to limit cloud costs to some extent.

This week, Vercel, a company best known for its Jamstack front-end web development platform, announced a new serverless-like product called Fluid compute. As Runtime's Tom Krazit describes it, the product is designed specifically for AI workloads "to eliminate the idle period when an AI app is waiting for a model to answer a question." In doing so, this small amount of compute could add up to a lot of dollars saved, if it works as described.

FastForward #8: The language problem with large language models

You may also like...