AI Learns to Think Fast: DeepSeek Running for the Masses

The currently over-hyped GenAI hysteria backed by bloated AI infrastructure implementations will predictably cause big problems:
- Massive 100+ billion node models use a boatload of power for both training and inference. MS and Amazon (et.al.) all publicly want to build and directly hook into whole new power plants – not a climate friendly approach1
- At some point us “AI Users” are going to have to pay for all the heavy duty AI infrastructure and the energy it’s using. Today there is extreme “greed” in the AI market with plenty of speculative money flowing in to power it up, but that will necessarily need to correct and when it does the disappointment will bring our whole economy down several notches
- BTW the term “AI Users” encompass all of us, whether we’ve bought into the current hype or not. E.g. Try doing anything new in MS without having Co-pilot slid into the mix of functionality (actually seems patronizing). AWS is making it too easy to include their AI services in every application being built (fast and very cool). Google is already cannibalizing their traditional search business by placing Gemini results front and center of almost every query (very telling and directly useful?) 2
- Hallucinations, an integral part of how GenAI works, will continue (accidents will happen and people will die).
Now we have China throwing down a gauntlet with DeepSeek. DeepSeek is based on open source, smaller infrastructure with cheaper GPUs, and overall takes a very home-grown “efficient” approach to GenAI. It appears to only activate the most relevant sub-part of a larger base model for inference (and probably training, although I’m not clear on that yet), which makes it faster and much cheaper to operate. DeepSeek might support wide “edge” AI usage by only forward deploying the necessary model sub-parts for each edge’s main (approved?) queries. DeepSeek being of Chinese origin, it’s also highly, um, filterable?3
Because DeepSeek operates mostly4 as an essentially smaller sub-model, it doesn’t need the über high-end AI infrastructure that us Westerners demand in play for completeness, correctness, cross-domain insights, etc. I think this DeepSeek approach presents an interesting perspective towards the proper uses of GenAI and maybe provides an astute insight into what GenAI is really good for with regards to assisting human endeavors. Let’s step back a bit and briefly review one of the best books on my shelf – “Thinking, Fast and Slow” by Daniel Kahneman.

Essentially, Kahneman’s premise is that human minds have evolved two main modes of thinking. We think “slow” when we really sit down and analyze situations, examine and sort out all the details, solve thorny puzzles, read and write poetry, consider and predict alternative courses of action, “strain our brains” for insight in making new connections, etc. However slow thinking mode requires a lot of calories, time and attention which as living beings we don’t have enough of to deal with everything all day long. As intelligent beings we rely on a default “fast” thinking mode based on learned heuristics to help us deal with our more commonplace situations and activities. Fast mode helps us avoid immediate danger, efficiently deal with daily life, create good times for slow thinking and ultimately better survive. Thinking about fast mode thinking can help explain a great deal of seemingly illogical human behavior.5
I’d propose that what GenAI offers is a type of “fast mode” thinking. I’ve been ungenerous to GenAI by resisting calling it “intelligent”. But now I think that GenAI actually could be seen as really good at augmenting and even surpassing our default human brain fast thinking mode. And by leveraging its inherent multi-billion node recall and deep “connection” relevance scoring, it can prep and jump-start slow mode thinking by first laying out all the related concepts to a given prompt, both explicitly and implicitly known (not unlike how our inherent “unconscious” fast mode informs and sub-consciously shapes our slow executive mode thinking).
GenAI clearly doesn’t perform what is described as slow mode thinking (and what most consider actual “intelligence”). But the DeepSeek efficient ML approach to activate only the most relevant parts of its larger “brain” seems congruent with our organic fast mode. And if we treat it as such, taking care to not confuse GenAI with slow mode thinking, DeepSeek’s increase in efficiency mirrors a likely step in the human evolution of efficient overall intelligence.
If what we really want is to create AGI – artificial general intelligence (and that’s what the DeepSeek founder, along with OpenAI et.al. really want to reach), what I currently think of as slow mode thinking, then larger, even more powerful AI clusters with horrendously complex and as yet undiscovered layer upon layer of executive functioning will need to be modeled6. Today’s “agentic” GenAI approaches don’t seem to offer much more than baked-in tricks to automate fast mode thinking, putting fast mode responsive activities directly into operation. That’s a good opportunity as computers can be both thorough and fast, but also risky as there is no flip-side “slow mode” AI to back it up (i.e. avoid hallucinations, executive reasoning, predict other slow mode thinking operators, discover and create actual innovation…).
Which for now leaves us at the frontier of figuring out how to best leverage and marry fast mode GenAI with the inherent slow mode side of our human minds. Eliminating the need for extensive prompt engineering and providing better knowledge search results are good steps. Agentic capabilities if corralled ingeniously might be another. Yet these will not be enough to keep “AI” once again from falling into the inevitable trough of disillusionment. GenAI will then slowly crawl back out of the muck within few years, probably first in gaming, augmented reality (AR) and embedded “edge” use cases where we no longer think of it as AI.
- I have heard the claim from AI vendors that AI is our best bet to solve climate change challenges in the future so full steam ahead releasing more carbon today! ↩︎
- Don’t cry for Google. They know what they are doing. I predict soon folks will be paying through the nose to have the Google GenAI models reference their site URLs in the AI “canonical” query results. As a sponsor, you might still get some click-through more than your competitors, but probably not nearly as much. Search users will soon probably have to implicitly agree (although not necessarily be aware) to having their top-level queries passed through to those “AI sponsors” even without a click-through. ↩︎
- And probably will be implemented and hosted with back-end abilities to identify and report on what an end-user is “really” trying to research and generate. ↩︎
- https://www.linkedin.com/news/story/chinas-deepseek-stuns-ai-world-7134418/ ↩︎
- A great deal of human psychological behavior can be modeled by considering fast modes, slow modes and the switching, interplay and tradeoffs between them. ↩︎
- And no, we aren’t there yet with an artificially implementable model or understanding of what is actually going on the human mind. But AI research continues! ↩︎