Ron Karidi

The through-line

Since 2013, Ron Karidi has been working on a single proposition: that human problem-solving is capped by the humans doing it, and that the fix is a machine built specifically for ideation rather than for prediction or analysis. The framing he returns to is that "we human researchers and problem solvers are inherently limited in the skills and capabilities we have to solve complex course disciplinary problems" [2], and that the answer is a machine that reads the web, connects the dots across domains, generates hypotheses at industrial volume, and hands them back to people to judge. He is explicit that this is an unusual use of AI, one that "goes beyond prediction or understanding data or analyzing data with this core focus on problem solving" [1].

What shifts across the material is scope rather than conviction. The early framing is about the mechanics of ideation itself, the cognitive bottlenecks and the chasms between data, knowledge and invention [1]. The later framing keeps that engine at the centre but points it outward, at climate action and at the way innovation capital is allocated, arguing that investors should use AI-powered search to find breakthrough opportunities and then go and recruit the entrepreneurs to build them [2].

On the limits of human ideation

The starting question is why breakthroughs take so long. His diagnosis has three parts. First, bias: real disruption happens when "researchers need to get free of their individual and collective biases and come up with frameworks that simply contradict what they've been used to", and those moments, where a decade of work turns out to connect to a different problem entirely, are rare [1]. Second, throughput: "look how long it takes to convey two simple ideas if we need to go through millions of ideas before we get the right one this is definitely an obstacle" [1]. Third, coverage: knowledge created in the 21st century alone already exceeds everything humanity produced before it, so "there is no single person that understands everything and knows about all the research that was done in their domain", which matters most because cross-domain connection has been the strongest catalyst of invention over the last two centuries [1]. The machine is designed against exactly these constraints, overcoming "cognitive bottlenecks constraints such as speed of thought and bias" while it connects the dots across multiple domains of knowledge and expertise [2].

He is also blunt that complex problems do not arrive in a usable shape. "Complex systems have data associated with different facets of the problems", and his worked example is global warming, tangled up with the greenhouse effect, air pollution, rainforests, fires and aggressive agriculture, each carrying its own datasets, with "an enormous challenge to traverse this potentially almost infinite space of paths to connect all these datasets and come up with the right hypothesis" [1]. Data will not simply arrive "in a flat file where we can just run one of the state-of-the-art deep learning or other machine learning algorithms and get the solution" [1].

On the chasm between data, knowledge and invention

The most distinctive part of his argument is the split between data and knowledge. Data is bottom-up and inductive, and it runs out precisely when you need it most: for a solution you have just invented, "the first data points will come when we finish building this solution", and waiting for that evidence before discovering the design is wrong is unaffordable [1]. Knowledge is the mirror image, "often qualitative sometimes also quantitative", top-down and deductive, and it is what lets you cross the gap when data does not yet exist [1]. Beyond that sits a further gap he calls the knowledge to ideation chasm, since new knowledge itself requires new inventions, which is where machines become necessary rather than convenient [1].

The requirement he sets for such a machine is concrete: connect the dots "at the rate of millions of hypotheses per minute", organise everything fetched from the web into a knowledge graph, apply ideation principles, and link the resulting ideas to unmet needs [1]. Validation is the other half of the loop, applying hypotheses to data to see "which ones actually hold and which ones we need to throw to the basket and go back to the whiteboard" [1].

On the two engines

The architecture he describes maps directly onto that split. One engine "creates new ideas from data", the hypothesis engine; the other "mines the entire web more than 400 billion pages patents research papers and even used and connects the dots to provide answers to complex research questions", the knowledge engine [2]. Together they are what he calls the ideation machine [2].

In practice the two are used in sequence. Asking what negatively influences sustainable agriculture, drawing on research publications, patents, news and the wider web, surfaces soil erosion, soil salinity, climate change, nickel mining and deforestation, and shows how the dominant factor varies by geography, deforestation in Brazil, climate volatility and food deficits in sub-Saharan Africa, nickel mining in the Philippines [1]. That is the qualitative layer. Moving to relative importance means "overlaying data on top of knowledge": take low yield, ask what increases yield, get fertiliser as a candidate, link it to a repository of over a million datasets, then set one variable as dependent and the other as independent and let the hypothesis engine find the patterns [1]. The machine builds the data pipeline automatically and generates the hypotheses before the model is built [1]. The patterns it returns are readable rather than opaque, including one found by traversing Wikipedia showing that a country's being in Africa roughly triples the likelihood of low yield, alongside minimum nitrogen usage over twenty years and population density drawn from Wikipedia info boxes [1].

He also uses the machine to test whether an idea has momentum. Asking what reduces soil erosion and salinity surfaces compost, which helps with both, and biochar, and the follow-up question is whether biochar is new, trending or growing: around one publication a decade earlier, 64 by 2018, plus the top companies generating IP, competing market size estimates whose distribution is worth seeing precisely because they disagree, and the hotspots of innovation and implementation, with Ghana emerging as one [1].

On machines producing half-baked ideas and humans criticising them

He does not oversell what the machine generates on its own. Going beyond mining existing knowledge means extracting universal inventive principles from past inventions and applying things "that are not written anywhere", and he concedes immediately that "machine actually are going to be very good at coming up with half-baked ideas" [1]. The design response is a division of labour: pair the machine with people and "let humans do what they do best criticize" [1]. Refinement comes out of that dialogue, with humans setting the challenges, then acting as reviewers, filtering the good ideas and confirming the path to a solution [1].

The live examples are deliberately raw. Change the colour of a pesticide so it becomes visible and can be optimised with computer vision. Turn irrigation upside down, or make the land mobile. Use centrifugal force at harvest, spinning each fruit tree inside a net for sub-second picking [1]. His point about them is that "if you google for this sentence you will not find it it's actually generated dynamically", and that human responses feed straight back in, making the process a crowdsourcing loop rather than a one-way output [1]. The whole thing he describes as harnessing collective intelligence in three forms: computational intelligence, combining building blocks into hypotheses and testing them against data; web intelligence, reading the web to connect the dots; and human intelligence, to go beyond what is already written down [1].

On climate action as the proving ground

Sustainability is where he directs the machine, and he draws an explicit link between business impact and social impact when partnering with large companies to scale across many problems [1]. Sustainable agriculture is his chosen demonstration precisely because it connects to 17 UN sustainability goals [1]. Climate action was also named as the first area for making the investor-led model real [2].

The work presented alongside his framing includes Project World Stage, an attempt to find non-trivial areas for collaboration between nations around the Biden climate summit and COP26, built by analysing polluting countries and the momentum of their polluting activity, 45 technologies for tackling climate risk, 27 specific climate risks by country, and leaders' commitments through media mentions and a social network graph, then overlaying the four layers to find where they coincide, which surfaced energy storage as one candidate area [2]. Downstream of that sits a strategic alliance between SparkBeyond, the University of Cambridge and the Stimson Center, described as the Alliance for Climate Resilient Earth, which ranks climate solutions using AI, has corporations sponsor one transformational solution, runs a year-long technological development session at Cambridge with AI facilitation, and requires the sponsoring corporation to share the resulting technology with the developing world at no cost and with competitors on an open basis [2].

On redirecting how innovation gets funded

The most pointed structural claim he makes is about capital. He sees "an opportunity for investors to actually revert the direction of venture-backed innovation": rather than waiting for founders to arrive with ideas, investors would invest in AI-powered search to find breakthrough opportunities, then "proactively recruit entrepreneurs to realize these opportunities" [2]. That is the same argument he makes about researchers, applied one level up. He is equally clear that the mission exceeds any one company, saying it is "too big for us to be able to undertake on our own" and inviting partners in [1].

Takeaways

• The core bet is a machine built for ideation rather than prediction or analysis, one that reads the web, connects the dots and generates hypotheses for humans to judge [1][2].
• Three human bottlenecks define the design brief: individual and collective bias, the throughput of getting through millions of ideas, and the impossibility of any one person knowing their whole field, let alone adjacent ones [1][2].
• Data is bottom-up and often missing for anything genuinely new, since "the first data points will come when we finish building this solution"; knowledge is top-down and deductive, and bridges that gap [1].
• The platform runs on two engines, a hypothesis engine that generates ideas from data and a knowledge engine mining more than 400 billion pages of patents, research papers and web content [2].
• Machine-generated inventions are expected to be half-baked, and the intended workflow pairs them with human criticism to refine them [1].
• The method moves from qualitative to quantitative: surface causal factors from literature and patents, then overlay datasets and run hypothesis search to quantify them, as with yield, fertiliser, compost and biochar in sustainable agriculture [1].
• Idea maturity is testable too: publication trend, IP holders, competing market size estimates and geographic hotspots of implementation [1].
• Investors should flip the direction of venture-backed innovation, funding AI-powered search for breakthrough opportunities and then recruiting entrepreneurs to build them [2].