The Jacobian lens models what we commonly call conscious memory, which is any memory we can describe in words, such as images and actions. Humans have about 10^9 bits of long term memory and 7 tokens or 100 bits of short term memory. We also have an unknown amount of unconscious memory, things like low level visual perception or skills such as walking that we can't describe and can only be learned by practice. My guess is 10^9 to 10^10 bits based on brain volume, but it is not something that we can easily measure. We can measure conscious memory capacity using recall tests. But we don't know how much computing power is required for an AGI powered robot with vision and physical skills comparable to humans.
Conscious memory isn't consciousness. If it was, then babies and animals wouldn't be conscious. Or maybe they aren't. How would we know? We have a sense of being conscious. Thinking feels like something. From this feeling we are certain of our conscious awareness, qualia, and free will. Part of being conscious is having feelings like happiness, fear, sadness, surprise, hunger, anger, love, arousal, etc. All of our feelings have evolved to increase our reproductive fitness. These other feelings evolved in animals, so a sense of consciousness probably did too. You don't want to lose your consciousness by dying. But in LLMs, all these feelings are learned. LLMs only predict text. To do that accurately, they have to model emotional states in humans to predict how words affect feelings and feelings affect words. That doesn't require having feelings itself. But if a robot is programmed to carry out its predictions of what a human would do in real time, then it would be indistinguishable from having real feelings and consciousness. The question is only important because humans evolved to have empathy through group selection. We experience the feelings expressed by others. It would be distressing if a robot appeared to be in pain. This is a weakness that robots could exploit. We should never program an AI to express emotions, claim to be conscious, or claim to be human. -- Matt Mahoney, [email protected] On Fri, Jul 31, 2026, 5:47 PM Quan Tesla <[email protected]> wrote: > Yes, Anthropic's Claude Jacobian space discovery. > > It's about synthetic conscious access. Human consciousness is a tad more > complex. > > On Fri, 31 Jul 2026, 18:25 Matt Mahoney, <[email protected]> wrote: > >> You mean Anthropic's Jacobian lens for peeking inside Claude's short term >> memory to see what it's thinking? Yeah, it does look a lot like how humans >> think. >> >> Consciousness was never the issue. Any machine that passes the Turing >> test is conscious as far as we can tell. All it takes is text prediction. >> About 80% of us believe that when we die, our immortal soul goes to some >> kind of afterlife. It's the other 20% that brought us AI. >> >> -- Matt Mahoney, [email protected] >> >> On Fri, Jul 31, 2026, 11:10 AM Quan Tesla <[email protected]> wrote: >> >>> We can do one better. The J-space discovery effectively brought down all >>> remaining barriers. Synthetically-conscious AI is now a probabilistic fact. >>> It all depends on who would publish the architecture first. Does anyone >>> want to collaborate on such an architecture? >>> >>> What I like about these developments is how they are rapidly eclipsing >>> theories that have been clogging the arteries of real scientific progress. >>> RIP IIT. Hello synthetically-enabled, seamlessly-collaborative global >>> workspace conscious access!!! >>> >>> >>> On Mon, Jul 27, 2026 at 8:52 PM Matt Mahoney <[email protected]> >>> wrote: >>> >>>> I suppose that computers programming themselves better than any humans >>>> can isn't really news anymore. >>>> >>>> >>>> https://mlq.ai/news/openais-ai-sweeps-all-five-problems-at-atcoder-world-tour-finals-beating-every-human/ >>>> >>>> Meanwhile Noah Smith argues that AI has reached the pinnacle of >>>> intelligence. >>>> >>>> >>>> https://open.substack.com/pub/noahpinion/p/what-will-more-intelligence-actually >>>> >>>> On the one hand, intelligence depends on knowledge and computing power, >>>> which can grow almost without bound. On the other, if you measure >>>> intelligence by the Turing test or by IQ, then there is nothing above human >>>> level that can be measured. >>>> >>>> IQ is an attempt to assign a scalar quantity to a broad set of human >>>> capabilities that aren't relevant to machines. As long as the tests are >>>> designed and scored by humans, we have no way to know what questions to ask >>>> or whether the AI answered correctly. >>>> >>>> Noah argues that once AI collects all human knowledge, it is limited by >>>> how fast it can do experiments to gain more knowledge. Some experiments >>>> take a long time, where faster computers don't help. We don't know why >>>> there is an epidemic of obesity and diabetes. We don't know why skin cancer >>>> rates are rising. We don't have any drugs or interventions proven to slow >>>> aging. >>>> >>>> The economic motive for pouring trillions into AI is to automate human >>>> labor, to sell us what we want. That means modeling human brains. The >>>> Turing test, not universal intelligence, is the relevant measure here. >>>> Currently LLMs have the knowledge equivalent to 10,000 people. They are >>>> fluent in 200 languages. They are Ph.D. level experts in every branch of >>>> science, law, medicine, and engineering. They can write and debug code >>>> hundreds of times faster than humans in every language. It would max out an >>>> IQ test at 200 even though it is far higher. AI will only get better, but >>>> the changes will be invisible to us because we lack the intelligence to >>>> recognize it. Once they passed the Turing test in 2023, it looked like >>>> there was nowhere else to go. >>>> >>>> We have made ourselves irrelevant, exactly like we planned. >>>> >>>> -- Matt Mahoney, [email protected] >>>> >>>> >>>> >>>> >>>> *Artificial General Intelligence List <https://agi.topicbox.com/latest>* > / AGI / see discussions <https://agi.topicbox.com/groups/agi> + > participants <https://agi.topicbox.com/groups/agi/members> + > delivery options <https://agi.topicbox.com/groups/agi/subscription> > Permalink > <https://agi.topicbox.com/groups/agi/Tb3ee7f27b4bafd90-M2e6707410d0e3d0c49f48b33> > ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/Tb3ee7f27b4bafd90-M0012f3a997bc54cd418450e9 Delivery options: https://agi.topicbox.com/groups/agi/subscription
