When I joined the AGI list over 20 years ago, there were a lot of individual AGI projects by people interested in understanding human cognition rather than world domination. I understood that the obvious application of AGI would be automating human labor. I estimated that to be worth global GDP divided by interest rates, or about $1 quadrillion. If the problem was easy, it would have already been solved. Nobody believed me. Now all those projects have been abandoned while we pour trillions into AI.
My cost estimate was based on hardware, software, and knowledge collection. All of the projects focused on software, which was the least expensive part. The most expensive was collecting human knowledge, which I estimated at 10^17 bits (10^10 people x 10^9 bits long term memory minus 99% overlap). It has to be collected by surveillance at 5-10 bits per second through speech and writing, costing in the hundreds of trillions. The Common Crawl is only 10^13 bits compressed. This is why AI still doesn't know how to do your job. The second area of interest was uploading or life extension. This led to endless debates about consciousness. I pointed out since 1999 that text prediction measured by compression is sufficient to pass the Turing test, and is therefore indistinguishable from consciousness. The big AI companies seem to agree that consciousness is irrelevant. I don't think we are going to see radical life extension or longevity escape velocity either. Global life expectancy has been increasing steady at 0.2 years per year for the last century and I don't think that will change much. Unlike Moore's law, the cost of medical testing on humans is rising exponentially. The third area of interest was the singularity. We had a separate mailing list for that, but it died years ago. Nobody on the AGI list believed that their project had any realistic chance of recursively self improving to reach infinite intelligence anytime soon. I argued that the premise that surpassing human intelligence would launch a fast takeoff was flawed, because intelligence is not a point on a line. We still argue now about whether we have passed it. The centi-billionaires and trillionaires interested in world domination have no interest in what we are doing. -- Matt Mahoney, [email protected] On Thu, Sep 24, 2026, 1:40 AM Zoltan Nadasdy <[email protected]> wrote: > I don't think the goal of the current AI development is to better > understand ourselves, our own cognition ... > The goal is simply to dominate the market. Those pursuing the > understanding of human cognition you would find in working in neuroscience, > psychology or cognitive science. > The two branches of AI development, one that aims to understand human > cognition and the other to engineer systems that emulate human cognition > have split a long time ago in the late 80s and early 90s. > > The subject of our biggest fear of unleashing AI should not be the > software itself but the human greed of unleashing it. > -z > > > > > > On Wed, Sep 23, 2026 at 7:23 PM Fleeky Flanco <[email protected]> wrote: > >> the term deceipt is really weird here? deceit implies goals , goals imply >> an inner world .. basically exactly what linas is arguing .. >> or do you mean something else ? we need a different term .. >> >> >> On Wed, Sep 23, 2026 at 8:11 PM Matt Mahoney <[email protected]> >> wrote: >> >>> On Wed, Sep 23, 2026, 4:32 AM David Bray, PhD <[email protected]> >>> wrote: >>> >>>> Open question to consider: >>>> >>>> What if the Turing Test was the wrong test - because at its heart it’s >>>> actually a machine deceiving a human vs. necessarily helping a human? >>>> >>>> https://apolitical.co/en/articles/rethinking-ai-success-from-imitating-us-to-helping-us-improve-195 >>>> >>> >>> Let's not lose sight of the goal of AGI. It is to automate human labor, >>> worth $60 trillion per year. It is the reason we are pouring trillions of >>> dollars into the problem. >>> >>> >> maybe for some greedy billionaires or cynics , but i honestly hope the >> actual goal is to better understand ourselves , our own cognition , and the >> cognition of other intelligent agents around us and in the universe .. >> >> >>> To automate labor, it must be able to do everything that humans can do. >>> The Turing test evaluates language modeling, the ability to understand >>> language and respond appropriately, which is a requirement of every job. It >>> implies human level knowledge of the world. It does not test embodiment, >>> like vision, hearing, and robotics. But language is a big part of it, >>> probably the most important part. And the rest is being solved too. We have >>> AI that can understand speech, describe images and video, and drive a car >>> with 1/8 the accident rate of humans. Humanoid robots can compete in >>> sports. All of these functions require integration with a language model. >>> >>> You are right that AI should not deceive humans that it itself is human >>> or conscious or that it has feelings. But it must be capable of doing so to >>> meet its other goals. We can instruct it appropriately and use text >>> compression as an alternative test. >>> >>>> >>>> From colleague and friend Vint Cerf on this as well: >>>> https://cacm.acm.org/opinion/turing-test-2/ >>>> >>> >>> A computer judge in the Turing test is really a competition between two >>> computers. The winner is the one that is better at predicting text. So >>> really it measures the same thing as the Turing test or text compression. >>> >>> -- Matt Mahoney, [email protected] >>> > > -- > --- > Zoltan Nadasdy > Dept. of Neurology, Dell Medical School, UT Austin ([email protected]) > Dept. of Cognitive Psychology, ELTE Budapest [email protected] > > http://brainstim.psy.utexas.edu > > ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T9e51fc70bbc4e9fc-M5c813b94aea652a682351e33 Delivery options: https://agi.topicbox.com/groups/agi/subscription
