Introduction: Navigating the AI Hype Cycle

Every technological revolution brings waves of hype, followed by inflated expectations, disillusionment, and eventually, practical applications. AI is no exception. Gartner expertly outlines this process showing various types of AI and their current expectations versus maturity (see image below).¹

With these cycles comes an onslaught of "experts" all weighing in on how the market will react to AI and what the future will be. Most of these are wrong.²

Knowing which aspects of the current AI evolution will still be here in 5-10 years is a tough prediction to make:

• Do people trust AI?

• How will it be governed?

• Should it be open-source, open-weight, or locked down and controlled by a select few companies?

• Which algorithms and trends actually provide value?

After a year of building in the AI space with Attrove, these are my observations on where conventional wisdom might be missing the mark—and what I'm seeing on the ground with real users that contradicts much of the popular narrative.

The True Strengths and Limitations of Language Models

Language models (LLMs) are incredible. I'm sure you've had a "wow" moment at some point when typing something into ChatGPT and almost immediately getting hit with an impressive response. Ask it to explain quantum computing or write a sonnet about your morning coffee, and the results can be astonishing.

The new language of computers is "English" and anyone can quickly jump in and use them (the genius of "chat" GPT is the familiar chat interface we all have been trained to use since the nascent days of mobile phones & the internet). OpenAI just passed 400 million users this month—that kind of growth is unprecedented.³ You can now analyze massive amounts of text in a few seconds & provide (reasonably) accurate responses.

However, there are significant limitations that many overlook in their excitement.

Precision tasks like counting, consistent reasoning, and mathematical operations continue to be tough for language models. The famous "strawberry" problem is a good example of how language models have trouble with seemingly simplistic tasks: how many 'r's are in the word "strawberry"? It probably took you a few seconds but I'm betting you landed on 3. LLMs tend to struggle and output 2.⁴ Try asking it to count characters in a longer passage and watch it falter even more dramatically.

Understanding these strengths and limitations helps us design better AI systems rather than forcing LLMs to do things they're not optimized for—something we've incorporated deeply into Attrove's architecture from day one.

Beyond Chatbots: The Evolution of AI Interfaces

The natural allure of conversation is strong—humans have evolved to communicate through language. This is the magic of language models: they feel human because they communicate in words (not code, symbols, or 80's style robot voices). Chatbots are open-ended which is both powerful and confusing. You can take the conversation almost anywhere (most LLMs have some guard rails around unlawful information); that also makes them hard to control. What if you only want users to ask about their past orders on an eCommerce site? What if users don't actually know what questions they can ask?

The blank canvas problem is a tough design to overcome.⁵ It also taxes users mentally in that they need to keep coming up with follow-on questions or think of novel ways to pull out information they want. After the initial novelty wears off, many users experience "prompt fatigue"—the cognitive burden of constantly having to formulate the perfect question.

You know what's better than pulling out data in an open-ended context? Pushing relevant data straight to you.

Moving from the quintessential reactive "pull" to the proactive "push" will be critical as AI user experience evolves. Instead of asking "What do I need to know today for my day?" How about a gentle nudge that you can read or listen to a brief overview of what you care about? Think podcast meets personalized news feed—where the AI tells you, "Three people on your team submitted updates on Project X" or "There's a scheduling conflict tomorrow that needs your attention."

While the push model has tremendous value, it's much harder to nail. How does the AI know what you want (as opposed to anyone)? This is where the real magic happens.

Time to dive into legacy data science.

The Enduring Value of Traditional Machine Learning (ML) and Natural Language Processing (NLP)

Not every workflow caters well to language models. They tend to be slower, more expensive, and have limitations like we discussed above. However, that doesn't mean that data science has failed us: we have many more tools in our proverbial tool belt.

Ever use Instagram or TikTok? Those apps are highly addicting and amazing at recommending relevant content for you. They do this through a series of inputs that all go into your user profile⁶ and help cater the app experience towards your personal tastes. This uses a lot of machine learning and modeling to take a series of inputs (such as likes, watch time, etc.) and outputs a series of information used to show you the next image, video, or feed.

These recommendation engines don't use LLMs for the core of their functionality—they rely on more traditional machine learning approaches that are lightning-fast, highly scalable, and deterministic. They excel at exactly what LLMs struggle with: numerical processing, pattern matching across massive datasets, and consistent output.

Thus, finding an appropriate way to personalize content has tremendous power—and it works amazingly well in conjunction with emerging language models. The future isn't about choosing one approach, but skillfully combining both.

The Hybrid Future: Using the Right Tool for the Right Job

Parsing raw text and distilling down topics, themes, sentiment, questions, etc. work well with LLMs. Personalizing content to show you exactly what you care about works well with traditional machine learning methods. Put these two together and you start to get a nice "AI Stack" for personalizing text content and workflows.

This hybrid approach enables systems that can understand nuanced content (LLMs) but recommend it with the precision and personalization of traditional ML. Think of it as combining the best of both worlds: the linguistic intelligence of modern AI with the numerical reliability of established data science techniques.

Attrove's Approach: Pragmatic AI Integration

Our method of parsing workplace communication involves using an LLM to give each message some structure. We analyze conversations, documents, and messages, extracting key elements like topics, action items, and context. This structure can then be fed into traditional ML algorithms that help rank and recommend items that each user cares about.

From there we can go back to the LLM with a series of messages and text that are personalized and relevant. Want to go one step further? Give an agent the relevant work context and boom, you've got a nice status update PowerPoint draft ready without having to manually comb through dozens of messages and updates.

For example, rather than forcing users to ask "What updates happened on Project X yesterday?" our system proactively surfaces the most relevant updates based on your role, past engagement patterns, and the importance of different information streams—combining LLM comprehension with ML prioritization.

Conclusion: Embracing Nuance in the AI Revolution

The future of AI isn't about LLMs vs. traditional ML, but about understanding when to use each. Data science will continue to discover better techniques, and incorporating the appropriate technique for each task will provide tremendous value.

So do I think the chatbot is the endgame for most AI use cases? Not even close. It's just the beginning of a much richer ecosystem of AI interfaces and experiences that will evolve rapidly over the coming years.

As we continue building Attrove, we're committed to this hybrid approach—using the right tools for each specific challenge rather than forcing everything through a single technological lens. After all, the most powerful systems aren't those that blindly follow trends, but those that pragmatically combine techniques to deliver genuine value to users.

1: https://www.gartner.com/en/articles/hype-cycle-for-artificial-intelligence
2: https://www.forbes.com/sites/bernardmarr/2024/09/24/the-10-biggest-ai-trends-of-2025-everyone-must-be-ready-for-today/
3: https://www.cnbc.com/2025/02/20/openai-tops-400-million-users-despite-deepseeks-emergence.html
4: https://venturebeat.com/ai/the-strawberrry-problem-how-to-overcome-ais-limitations/
5: https://www.koistudios.io/articles/3-ai-chatbot-design-patterns-you-should-avoid-and-what-to-do-instead
6: https://support.tiktok.com/en/using-tiktok/exploring-videos/how-tiktok-recommends-content