Sergey Davidovich

The through-line

Davidovich's consistent argument, running back more than a decade, is that the valuable thing a machine can do with data is not predict an outcome but ask questions about it. When he started, "when you said AI, it most likely referred to predictive analytics," where "the output is in most cases a number or a category" and "nobody could imagine that machines could come up with creative ideas on their own" [1]. His bet was that a different category of capability was needed, one that lets businesses "understand root causes" and "understand the factors behind outcomes," and that this requires different machinery: "You need the ability to ask questions. Not one question, but billions of questions. Each hypothesis, each question that the machine would come up with needs to be tested on the data in order to discover the truth" [1].

The framing has survived the generative AI wave largely intact, and he now applies it to autonomous agents and to the question of where AI-driven optimization goes next [3]. What changed is the technology available to execute it, not the diagnosis: the failure mode he identified in predictive analytics, a system that answers rather than interrogates, is the same failure mode he now identifies in agents that are given a prompt and a database and left to get on with it [1].

On why ideas beat predictions

The insight that reoriented his product thinking came from work with a large retail chain, helping them find the factors shaping decisions about opening new stores and inorganic moves. The lesson was that "people want ideas and want the ability to shape the future in many cases more than being able to predict the future," and from there that "the ability to ask questions is extremely valuable" [1]. Prediction tells you what will happen; hypothesis generation tells you which levers exist.

On code as the building blocks of ideas

The other founding realisation was that the raw material for machine creativity was already lying around in public. Code on the web, on GitHub and in open source repositories contains "the building blocks of ideas," because "almost any idea that you have you can express as code" and "people have already written those building blocks" [1]. The bet followed directly: explore novel combinations of those blocks, scan them fast enough and systematically enough, and "maybe one of those ideas may lead to a breakthrough" [1].

On why most AI agents fail

He places today's agents on a spectrum, and finds almost all of them at the wrong end. The most naive version is a bare prompt: "You are a salesperson, your goal is to sell these products. Here's a bunch of products. Good luck" [1]. One level up, MCP gives the agent access to a product database or a CRM so it has more context and "in principle supposed to do better" [1]. The level that matters, and where "most agents fail today," is when you stop merely informing the agent and instead "let it ask questions and understand the nuances" [1].

His worked example is technical support. An agent troubleshooting an issue for one customer will struggle, whereas looking "across 500 customers" to find "a certain pattern, a certain regularity behind those failures" would pinpoint the root cause and be far more effective [1]. Doing that means combining structured and unstructured data, and going beyond querying the structured data: "You need to be able to scan it across a broad range of ideas" [1]. This is the reasoning that pushed SparkBeyond into agentic territory [1].

On measuring agents like people, and on feedback loops

Agents inherit the KPIs of the humans they stand in for: a salesperson is measured on conversion rate, a support agent on escalation rate, resolution rate and customer satisfaction [1]. He treats this as an advantage on two counts. It permits comparison against human-driven baselines, and it supplies ground truth for reinforcement learning and fine-tuning policies, which is what makes "systems that continuously improve" possible [1]. On the fashionable question of human oversight he is blunt: "the human in the loop is optional" [1]. What is not optional is observation of consequences. Once an agent can "observe the outcome of your actions in the real world, that's the feedback loop that you need," and he points to systems like AlphaGo, which needed no human in the loop because they "operate in an environment where they immediately get feedback for every action that they make" [1].

On bias, trust and coherent nonsense

Asked what he has learned about human resistance to AI recommendations, he concedes there is "no easy solution" [1]. Bias is not a defect to be scolded away; it is functional. Biases are "short circuits that enable us to make decisions fast," something needed for survival, and switching from intuition to reasoning is energetically expensive, a distinction he maps onto LLMs having "two modes, with thinking or without thinking" [1]. The trouble is that correct answers are often counterintuitive, may contradict personal experience, and run into the bias that we weight our own experience above collective experience [1].

Transparency helps most, since explained answers are easier to trust than bare ones, but he immediately complicates that remedy: LLMs are "very good at convincing us humans" and making us believe them "even when the explanation or the rationale is actually incorrect but coherent" [1]. He calls this coherent nonsense a psychological weak point in humans. His answer is to anchor explanations in evidence, which is what SparkBeyond does by empowering LLMs "with very granular and high precision data that allows them to contextualize and provide those factual anchors into their explanation" [1].

On critical thinking and what the next generation should build

Thinking as a father of children aged six and eight, he finds it hard to ignore the trend line: AI is showing superiority across more and more benchmarks, and "it is very well possible that we are less than a decade away from automation of knowledge work" [1]. That raises the question of "the role, the purpose, the meaning of us as humans," and here he draws on Bill Gates' reminder that jobs and careers are a relatively recent invention, that for many hundreds of years there were no careers, and that people were not therefore depressed or purposeless [1].

Two skills stand out for him, and they reinforce each other. The first is critical thinking, urgent both because machines are so persuasive and because "we also live in the post truth era," where it is "very easy to shape our own opinions, shape opinions of a generation through an algorithm" [1]. The second is reasoning itself, a muscle worth building "even if you do not need this for a job, for your career" [1]. Knowledge, by contrast, he treats as commoditised: "you don't need vast knowledge but you really want the ability to think," because knowledge is now on demand [1]. He does not claim to have the answer to how young people establish what is true, saying "I don't think I'm best positioned to answer this question," but he is candid about the stakes: he is "very worried about how fragile democracies are" and how possible it is to shape the mindset of an entire generation, questions he sees as belonging to ethics, psychology, governments and regulation rather than to technology alone [1].

On what bold means

Boldness, for him, is a matter of temporal range: "imagining what the world is going to look like one decade forward," which is hard because "we live in exponential times" [1]. The point is not accurate prediction. It is to "discover certain aspect of the future that you're extremely passionate about and excited about" and let that energise the work of realising it. His metaphor is spatial: boldness is about "how far into the future you throw your mental anchor" [1].

Takeaways

• Predictive analytics answers what will happen; discovering root causes requires machinery that generates and tests billions of hypotheses against data [1].
• Businesses generally want ideas and the ability to shape the future more than they want forecasts of it [1].
• Public code repositories are a library of the building blocks of ideas, and novel recombination of them at scale can surface breakthroughs [1].
• Most agents sit between a bare prompt and database access via MCP; the level that works, and where most fail, is letting the agent ask its own questions across combined structured and unstructured data [1].
• Root causes often appear only across a population: scanning 500 customers for regularities beats troubleshooting one case in isolation [1].
• Agents should be judged on the same KPIs as their human counterparts, which supplies both a baseline and ground truth for reinforcement learning [1].
• Human in the loop is optional; an observable real-world feedback loop is not [1].
• LLM explanations can be coherent and wrong, so explanations must be grounded in granular, high-precision data rather than merely fluent [1].
• With knowledge available on demand, critical thinking and the ability to reason are the two skills worth building for the next generation [1].