EP 556: Choosing the Right AI: Agents, LLMs, or Algorithms?
🎯 Summary
EP 556: Choosing the Right AI: Agents, LLMs, or Algorithms? - Comprehensive Summary
This episode of the Everyday AI Show, featuring Michael Abramov (CEO of KeyMaker and KeyLabs), serves as a practical guide for business leaders navigating the confusing landscape of modern AI technologies, emphasizing experimentation over rigid adherence to buzzwords.
1. Focus Area
The primary focus is demystifying and providing a framework for selecting the appropriate AI technology—specifically contrasting traditional Algorithms, Large Language Models (LLMs), Agents, and Agentic AI—based on specific business needs rather than hype. The discussion also touches upon the critical role of high-quality data preparation (Michael’s expertise) and the competitive dynamics of the AI talent market.
2. Key Technical Insights
- LLMs as Black Boxes: An LLM is fundamentally defined as a “black box” model capable of answering questions, serving as a foundational component.
- Agents as Decision Trees: Agents are conceptualized as automated workflows comprising multiple models linked by hard-coded logic (if/else statements) to perform complex tasks (e.g., planning a trip by checking bank balance, weather, and flights sequentially).
- Agentic AI as Dynamic Logic: Agentic AI is a concept where the decision-making logic (the if/else structure) within an agent is dynamically determined and orchestrated by a separate, supervisory model, rather than being pre-programmed.
3. Business/Investment Angle
- The Danger of Hype and “Corporate Blade”: There is significant overpromising on social media regarding immediate employee replacement. Furthermore, many startups building wrappers or plugins around foundational LLMs (like Perplexity wrapping ChatGPT) risk being quickly “blade-cut” (obliterated) when major players integrate those features directly.
- Data Quality is Paramount: Regardless of the AI chosen (algorithm, LLM, or agent), the quality, structure, and correct perception of data used for training or decision-making remain the most critical factors.
- Risk in Agentic AI Adoption: The Gartner prediction that 40% of agentic AI projects will be canceled by 2027 highlights risks stemming from high costs, unclear business value, or insufficient risk controls, especially when agents operate with complex, autonomous loops.
4. Notable Companies/People
- Michael Abramov (KeyMaker/KeyLabs): Guest expert specializing in data labeling and preparing training datasets for various AI models.
- Amazon Web Services (AWS): Mentioned due to the departure of a key Gen AI leader (Vasi Filiman) amid intense talent competition.
- Meta: Highlighted for winning a significant copyright lawsuit regarding the use of copyrighted books to train its Llama model, establishing a “fair use” precedent in that specific ruling.
- OpenAI: Noted for releasing deep research models (GPT-4 deep research, GPT-4 mini deep research) and webhooks for its API, expanding advanced capabilities to developers.
- Anthropic: Mentioned for securing an $8 billion stake from Amazon and also winning a similar copyright ruling regarding training data use.
5. Future Implications
The conversation suggests that the industry faces both known unknowns (speculation around AGI leading to mass unemployment vs. universal leisure) and unknown unknowns—paradigm-shifting innovations that are currently unimaginable, much like engineers in the Zeppelin era could not conceive of the airplane. The immediate future requires businesses to focus on practical application rather than theoretical endpoints like AGI.
6. Target Audience
This episode is highly valuable for Business Leaders, CTOs, Product Managers, and AI Strategy Professionals who need actionable guidance on evaluating and implementing AI solutions beyond the current buzzwords.
Comprehensive Narrative Summary
The podcast addresses the overwhelming complexity facing business leaders trying to select the right AI tool—be it a traditional algorithm, a foundational LLM, or the newer concept of autonomous agents. Host Jordan Wilson sets the stage by acknowledging the alphabet soup of AI terminology.
Michael Abramov advises against chasing trends, likening the choice to selecting a life partner—it must suit evolving needs. His core recommendation is radical, personal experimentation: leaders must dedicate time daily to play with available tools, ignoring the jargon (LLMs vs. Agents) initially, and focus only on solving their most painful, immediate problems. Once a solution is validated personally, it should be shared with the team.
Abramov then clarifies the technical distinctions: LLMs are the core predictive models; Agents are systems where multiple models interact via hard-coded decision trees (if/else logic); and Agentic AI is when a supervisor model dynamically programs those decision trees. He uses the analogy of daily life decisions (waking a child) to illustrate the concept of decision trees inherent in agents.
A significant portion of the discussion centers on data integrity. Abramov stresses that being “data-driven” is often misunderstood, leading to “pseudo-data-driven” conclusions where individuals force data to support their pre-existing perceptions—a problem his data-labeling company frequently encounters.
Regarding the current rush toward agentic AI, Abramov supports the Gartner finding that many projects will fail due to unsustainable models or being outcompeted by giants. He introduces the “corporate blade” concept, where large tech companies rapidly absorb successful startup innovations built on top of foundational models.
Finally, when asked about the path beyond current technology, Abramov distinguishes between known unknowns (AGI scenarios) and unknown unknowns—truly revolutionary concepts that cannot even be predicted based on current technological trajectories, using the analogy of airship engineers failing to predict
🏢 Companies Mentioned
💬 Key Insights
"Now, there are some unknown unknowns. A good example of it is imagine the times when people were flying on zeppelins or on balloons... They could never imagine an airplane, right? That was an unknown unknown for them."
"Perplexity is a multimillion-dollar startup, and I think it's a multi-billion dollar valuation already, but all it does is it's wrapping ChatGPT and then it adds a little bit better search capabilities on top of it... And if they like the idea, they can do it and just blade-cut out all of these startups that did interesting things."
"40% of agentic AI projects will be canceled by 2027, either due to high costs, unclear business value, or inadequate risk controls."
"People what they did was they were super data-driven. They took data, and they have built conclusions on the data, and they would come to me and say, 'Hey, this is the problem, and here is the data that proves it.' And that was super funny to see that the data didn't prove it at all. It was their perception of the data that proved it."
"Oh, that's the most important thing."
"I think that just goes to show and emphasize ultimately how—even like what your company does, like data—and making sure that you have the structured data to help answer those if-else conditions, right? Like, whether it comes to traditional algorithms, large language models, agentic AI, whatever it is, how important is having your data correct?"