Gen AI Reduces Text
It’s been about two years since ChatGPT introduced a new era of computing, and we’re still trying to figure it all out.
The best definition of LLM based gen AI I’ve come across is from Laurie Voss, VP of Developer Relations at LlamaIndex, an open source provider of tooling to develop AI agents. He makes the case that LLMs should not be called AI because they are not intelligent.
Understanding he has lost that battle, however, he continues:
“But if we can’t call it what it is we can at least know, as practitioners, what it is: very fancy autocomplete.
“Is what you’re doing taking a large amount of text and asking the LLM to convert it into a smaller amount of text? Then it’s probably going to be great at it.
“If you’re asking it to convert into a roughly equal amount of text it will be so-so. If you’re asking it to create more text than you gave it, forget about it.“
In other words, gen AI is very good at reducing text. Not so good at maintaining an equivalent level of text, and not good at all at generating more text.
Kind of like chopping up word salad ingredients for you and putting them in a bag for you.
Gen AI for Developers
Yes, you say, but what about generating code, rather than text? The consensus on this seems to be that gen AI is like having a junior coding assistant (a “very fancy autocomplete”?).
A former colleague and friend of mine, John Davies, recently pivoted his company called Incept5 to help organizations use gen AI to automate various corporate functions such as sales & marketing and regulatory compliance, which seems to be a good application of gen AI.
In this recent LinkedIn post, John says gen AI dramatically speeds up the software development process for these corporate automation projects:
“At Incept5, we have created and are utilising an AI tool that significantly accelerates the engineering processes, including discovery, planning, coding and deployment, at the enterprise level. I mean five to ten times faster, not just a marginal improvement.”
He adds that “It’s not AI that replaces people, it’s people who use AI replacing people who don’t.”
Although Incept5 built tooling on top of LLMs specifically to further accelerate the engineering process, using gen AI to achieve productivity gain for developers is not unusual.
But I’m also seeing a good number of other vendors incorporating LLM based tools into the development lifecycle, such as AppMap and AugmentCode for example.
(By the way, John’s comments on DeepSeek are well worth reading. Check the above post and his comments to others’ LinkedIn posts on the topic. He says that DeepSeek improves results and lowers cost.)
Building AI Agents
AI agents are getting a lot of attention, including the recent “war of words” between Microsoft CEO Satya Nadella and Salesforce CEO Marc Benioff on the topic of co-pilots.
It’s no secret that Microsoft’s co-pilots have missed the mark, except perhaps for the Teams co-pilot, which takes a large amount of text and reduces it to a smaller amount of text (i.e. producing a summary).
But I am skeptical about the Salesforce agents as well. Articles about the Salesforce Agentforce agents mention the number of customers and deals, but never seem to mention a revenue figure.
I have a hard time believing Salesforce is succeeding where Microsoft failed, although that’s the claim. And of course both Salesforce and Microsoft have been known to sell vaporware.
But the fundamental issue is that it’s not easy to create an AI agent in the first place, as Phil Calcado of Outropy describes in considerable detail.
As Phil says, “AI’s stochastic (probabilistic) nature fundamentally differs from traditional deterministic software development.”
Phil also says that “Working with AI presents exciting opportunities and unique frustrations for a team like ours, with decades of experience building applications and infrastructure….The biggest open questions lie in structuring GenAI systems for long-term evolution and operation, moving beyond the quick-and-dirty prompt chaining that suffices for flashy demos.”
Because gen AI works with text, and agents inherit all the accuracy issues inherent in working with text, they are often basically guessing what they should do, and they are not going to get it right all the time. Wrapping an LLM into an agent doesn’t solve the hallucination problem. A human has to check an AI agent’s work.
Top Use Cases for Gen AI — Overall
Gartner VP Analyst Svetlana Sicular recently summarized the top use cases for gen AI, using a a long research project called Clio commissioned by Anthropic, the creator of Claude, one of the top gen AI products on the market. The research analyzes what people most often use Claude for.
Other than software development, which was at the top of the bar chart, all the other things on the list are basically consumer searches.
In other words, based on this study at least, gen AI is primarily being used today as an alternative to Google and Bing to search for information pertinent to daily life.
Homework help, technology troubleshooting, parenting and childcare tips, sports rules, recipes, pet care, and relationship advice are among the top searches.
This pretty squarely fits Laurie’s definition of gen AI as reducing a large amount of words (i.e. the world wide web training data set) to a smaller amount of words (i.e. answering a specific question or searching for specific information).
Top Use Cases for Gen AI – Corporate
CIO Magazine recently published their list of the top nine gen AI use cases for business, which places advanced chatbot at the top of the list.
The article says businesses are adopting gen AI in awareness of its issues, such as hallucinations, and picking spots where the issue won’t hurt them, such as in low level customer support situations.
They also report that CIOs are not entirely sold on AI “digital assistants” (aka AI agents).
Coding assistance for developers makes the list here as well, as does marketing support and drug discovery, and other operational corporate functions.
Cybersecurity is listed among the important use cases, as is analytics, business process automation, and data extraction.
With the exception of coding assistants and customer service chatbots, most of the use cases fall under the definition of reducing text.
A Word About DeepSeek
I was in the middle of writing this column when the DeepSeek news landed, stirring up pundits, investors, and analysts across the media landscape.
As best I can tell, however, the impact is primarily economic — a potential blow to the U.S. gen AI companies’ bottom lines and VC companies’ returns on investments.
DeepSeek “did more with less,” producing a superior product with a relatively small investment that runs on commodity hardware. And DeepSeek released the product as open source, meaning they have undercut not only the investment narrative of the U.S. companies but their pricing and revenue models as well.
The gen AI war is far from over, however, and the same questions remain about what gen AI is good for, how it will make money, and how it will help people.
The Intellyx Take
Gen AI is a revolutionary new technology – no doubt about that. But it has been over hyped for a bit more than two years.
Of course over-hyping is what the industry does. It’s nothing new. Twenty years ago everything was a web service – even a CD containing a list of addresses – to convince customers the vendors in question were aligned to the latest trend.
It’s a bit like a gold rush – everyone heads for the hills to stake a claim, but only a small percentage actually strike it rich. It takes a while for reality to sink in.
Because gen AI works with text, not binary instructions, text based applications are what’s working.
And because gen AI is subject to errors, inaccuracies, and hallucinations, a human always has to be there to check the results.
Copyright © Intellyx B.V. Intellyx is an industry analysis and advisory firm focused on enterprise digital transformation. Covering every angle of enterprise IT from mainframes to artificial intelligence, our broad focus across technologies allows business executives and IT professionals to connect the dots among disruptive trends. As of the time of writing, none of the organizations mentioned in this article is an Intellyx customer. Microsoft is a former customer. No AI was used to write this article. Image source: Adobe Image Express