Jeff Bezos Is Funding a Wild Hunt for the Brain's 'Core Algorithm' (wired.com) 137
Jeff Bezos is backing Flourish, a new "neuro AI" startup with $500 million in funding and a reported $2.5 billion valuation, that aims to reinvent AI by studying the brain's architecture and building systems that learn continuously while using far less power than today's large language models. The company's long-term bet is that neuroscientists and AI researchers working together can uncover the brain's "core algorithm" and eventually create brain-inspired AI that runs on a tiny fraction of current compute. Wired reports: Rob Williams knows how to pitch Jeff Bezos: You write a press release as if your product has already been built. Bezos reads it and gives a thumbs up or down. Williams went through this process a lot as an executive on Amazon's "S-team," in charge of software products such as Alexa, until his departure last fall. But the pitch he made a few weeks later -- in December 2025 -- was different. Now he was collaborating with Thomas Reardon, a neuroscientist and repeat startup founder, and approaching Bezos as a funder, not a boss. Here's what Bezos, sitting on his yacht somewhere, read while Williams anxiously watched on Zoom: "Flourish is a neuro AI company that is solving the two most difficult problems facing AI today: power efficiency and continuous learning. We are building Cortex AI, the first synthetic intelligence system designed to match the computational capacity, learning efficiency, and power budget of the human brain."
A month later, I'm lunching with Reardon and Williams in the Flatiron neighborhood in New York City. Reardon gets right to the point. AI has dug itself into a hole, he says. Though increasingly powerful, large language models are greedy consumers of computer power and data. Though the inspiration for LLMs was rooted in biology, current frontier models have little in common with the human brain. A person uses about 20 watts of energy to process information; a single chip in an AI training cluster uses more than 30 times that amount. The hyperscalers require thousands of chips and gigawatts of energy, enough to power small cities. And those models need to suck up virtually all of what humans have written. Each new model requires more, more, more. For all of that, the models don't learn. Once you train them, they're stuck. The goal, Reardon tells me, is to build "a synthetic artificial intelligence brain that runs on 50 watts or less." It should adapt to its conditions, be as nimble as a human mind, and burn a tiny fraction of an LLM's compute power and energy. The proof of concept is thriving inside our skulls. "There's something fundamentally wrong with saying, "I need to basically read every book ever written 20 times over in order to learn English,'" Reardon says. "A human baby does it with a couple hundred thousand utterances."
Reardon and Williams haven't figured out yet how to build systems that match the magic of a human brain. What they have is a belief that an expert, well-resourced team -- of AI researchers and neuroscientists working essentially side by side -- can find the answer. The neuroscientists will conduct original wet lab experiments with some of the most advanced lab equipment available, to hunt for usable intel on the brain's architecture. They plan to release the models they're currently developing as near-term products on the path to a full reinvention of AI. The fuzziness of the proposal didn't bother Jeff Bezos. After reading Williams' two-pager, he chipped in $50 million. Other funding came from Lux Capital, Google Ventures, and Catalio, among others. Bezos then almost doubled his initial stake and told Reardon he'd have given more if they'd asked. Now with a war chest of $500 million and a reported valuation of $2.5 billion, Flourish just needs to invent a new way to do AI.
A month later, I'm lunching with Reardon and Williams in the Flatiron neighborhood in New York City. Reardon gets right to the point. AI has dug itself into a hole, he says. Though increasingly powerful, large language models are greedy consumers of computer power and data. Though the inspiration for LLMs was rooted in biology, current frontier models have little in common with the human brain. A person uses about 20 watts of energy to process information; a single chip in an AI training cluster uses more than 30 times that amount. The hyperscalers require thousands of chips and gigawatts of energy, enough to power small cities. And those models need to suck up virtually all of what humans have written. Each new model requires more, more, more. For all of that, the models don't learn. Once you train them, they're stuck. The goal, Reardon tells me, is to build "a synthetic artificial intelligence brain that runs on 50 watts or less." It should adapt to its conditions, be as nimble as a human mind, and burn a tiny fraction of an LLM's compute power and energy. The proof of concept is thriving inside our skulls. "There's something fundamentally wrong with saying, "I need to basically read every book ever written 20 times over in order to learn English,'" Reardon says. "A human baby does it with a couple hundred thousand utterances."
Reardon and Williams haven't figured out yet how to build systems that match the magic of a human brain. What they have is a belief that an expert, well-resourced team -- of AI researchers and neuroscientists working essentially side by side -- can find the answer. The neuroscientists will conduct original wet lab experiments with some of the most advanced lab equipment available, to hunt for usable intel on the brain's architecture. They plan to release the models they're currently developing as near-term products on the path to a full reinvention of AI. The fuzziness of the proposal didn't bother Jeff Bezos. After reading Williams' two-pager, he chipped in $50 million. Other funding came from Lux Capital, Google Ventures, and Catalio, among others. Bezos then almost doubled his initial stake and told Reardon he'd have given more if they'd asked. Now with a war chest of $500 million and a reported valuation of $2.5 billion, Flourish just needs to invent a new way to do AI.
Hmm (Score:5, Insightful)
Have they considered just hiring humans instead of machines?
Re: (Score:3, Interesting)
Have they considered just hiring humans instead of machines?
I'm sure they will, to conduct some undoubtedly ethical reverse engineering.
Re: (Score:3, Funny)
Naa, clearly business grads are the superior species and everybody else is obsolete and just a nuisance. Need to get rid of all of those underperformers.
Re:Hmm (Score:5, Interesting)
Human employees cost a lot. AI is the holy grail of eliminating labor costs. As with any investment, it will cost a lot more in the short run than having human employees would, but once the next threshold is surpassed, the savings and profits will more than make up for it.
In theory, anyway.
The slightly less cynical take is that AI can empower humans to achieve more than they otherwise could. Rather than eliminating jobs, we get a lot more productivity from the same number of employees, thus producing everything that everybody wants even faster and in greater abundance.
Either way, it's clearly valuable to many people for many reasons, so it will naturally garner a lot of investment money.
Re: (Score:2)
The little gap in this line of thought, is that those profits also come from humans with jobs.
Re: Hmm (Score:4, Insightful)
We don't need more productivity. As there was still some of the globe left to go and colonize we could use the extra productivity to occupy new markets, but now?... that everything is occupied? What need is there for the extra productivity?
But what we can do is the same productivity with fewer humans.
Questions remains: now that survival is tied to having an 40 h/week job, what do we do with with all those we don't need.
Re: (Score:2)
But what we can do is the same productivity with fewer humans.
But its those humans that are going to purchase your product. Henry Ford had this figured out. Pay workers more. Now they can afford your cars. You produce more and increase your profits. Win, win situation.
Re:Hmm (Score:5, Interesting)
Have they considered just hiring humans instead of machines?
The tech oligarch version if this will be brains in vats hooked up to electrodes performing as machines. These people don't want full humans. They want the benefits of the human brain without having to consider ethical treatment of workers.
Re: (Score:2)
Why AI ? (Score:2)
Unfortunately (in their mind) you can't own a human being anymore.
Re: (Score:2, Insightful)
Unless you are at war. Then you can own all the men and force them to go die in order to protect your power over them.
Re: (Score:2)
A bit of a double edged sword (almost literally) to having an army of killers trained to overthrow despots. Kind of why history is full of military coups.
Re: Why AI ? (Score:2)
Which is why nobody's being trained specifically to overthrow despots, they're being trained to follow orders.
Much simpler that way.
Re: (Score:2)
Blind obedience not been the US military's doctrine for a very long time (if ever). While following orders is important, the emphasis has been on comprehending the objectives and following lawful orders. Why? One reason is that a FRAGPLAN is concise and would require enough context to correctly execute according to the intent, especially if it was split from a larger operations order. To do this, you need a military with people who can think, and have absorbed the military culture, and not mindless idiots.
T
Re: (Score:2)
Of course not. Meatbags are always complaining and in the old days could get quite violent against the ruling class. https://www.wesanews.org/archi... [wesanews.org]
Re: (Score:2)
Humans have an annoying habit of unionizing or, if unable to collectively organize, getting frustrated enough to inflict damage or just not be motivated to work, or if somehow that's not allowed, starving to death because they have no other way to correct the imbalance of there being a small number of employers compared to the huge number of employees.
Re: (Score:2, Redundant)
They can't own humans anymore, but they can own the AI. They prefer to own rather than rent.
Re: (Score:2)
Re: (Score:2)
Wait, are you suggesting I can simply hire humans, to design and manufacture these intelligent computers for me?!
Oh, duh, I get it! You're taking Bezos' own point of view, rather than the point of view of his engineers.
But how do the engineers create the intelligent computer? It's not hiring-people all the way down, is it?
Ther'a been the idea for 50 years (Score:2)
Re: (Score:2)
AI researchers need Porsches too.
Re: (Score:2)
A human Algorithm? (Score:5, Insightful)
That's where they fail right away.
Re:A human Algorithm? (Score:4, Interesting)
Even Sigmund Freud's Id, Ego, Superego model is closer to the truth than the kind of "algorithm" that AI hypemen are searching for. Because at least it acknowledges that there are emotional, biological, and social components to human behavior.
Re: (Score:2)
Why when apps and social media are nearly as addictive and can be manufactured much more cheaply? I'm thinking smartphones for fetuses is the next big thing. Maybe have an LLM telling infants bedtime stories, perhaps continuously in one ear 24 hours a day. Capturing customers before they exit the womb is the kind of marketing research the industry craves.
Re:A human Algorithm? (Score:5, Interesting)
Re: (Score:3)
Re: (Score:3)
The human brain works in ways that no Turing Machine and therefore no algorithm can replicate.
Why do you think that?
Re: A human Algorithm? (Score:2)
Re: (Score:3, Funny)
But... All birds are machines. They are spying on us for the government.
Re: (Score:2)
Because its not a real brain? Can you make a machine that can perform like a bird? You can make it fly, make it make chirping noises, etc. But all of it is just a rough replica of the real thing. The human brain in incredibly complex in comparison, trillions if connections.
The claim you made was not about whether machines can currently perform like a brain, nor about complexity. The claim you made is that the human brain cannot be simulated by a Turing machine, which is a much, much stronger claim.
So I repeat, why do you think that?
Re: A human Algorithm? (Score:2)
Well, let's start with the realization that there's no indication whatsoever that the human brain is limited to deterministic bebavior, while Turing machines are.
You could introduce random elements. But then you'd have something that cargo-cult-like imitates a human brain, not that is (truly like) a human brain.
Re: (Score:2)
there's no indication whatsoever that the human brain is limited to deterministic bebavior
Almost every neuroscientist in the world would disagree with you about that.
Re: (Score:2)
Everyone who's ever made that claim about something in the real world has been wrong. Why do you think this is different?
Re: (Score:3)
Sure, the computing substrate doesn't matter. If you can figure out how our brain learns in terms of neurotransmitters and synapses, then you can abstract that and implement it in software (and/or hardware).
Re: A human Algorithm? (Score:2)
What if there's a soul?
Think: cargo cult. Even if the pacific folks ever figured out how a real raduo worked, and how a real runway is built, and a real.flight tower... even if they had built all that -- there still wouldn't be any planes landing, simply because the the US army didn't send any.
What if imitating the brain is like cargo cult? Something still missing?
Re: (Score:2)
Nobody has ever, even one time, seen a single atom, anywhere, that is not behaving per the known laws of science. So, the brain, like the rest of nature, would appear to be a machine that we can analyze and duplicate if we wish.
It doesn't leave much room for a "soul" (other than as a psychological construct), does it ?! The soul can't be controlling our neurons if science is doing it.
Great, more lies (Score:3)
Obviously, the only reason here is to keep the AI hype going. We know far too little about how a human brain works to even have a snowball's chance in hell finding out at this time. And I think Bezos knows that.
Re: Great, more lies (Score:3)
He doesn't know shit except how to abuse customers and employees. He could definitely fall for someone telling him it's possible. It's fundamentally a stupid idea, because the brain is not a computer and doesn't execute instructions like one, and therefore doesn't have any such thing as a core algorithm. The closest thing it has to that is physics.
Re: (Score:2)
He doesn't know shit except how to abuse customers and employees. He could definitely fall for someone telling him it's possible. It's fundamentally a stupid idea, because the brain is not a computer and doesn't execute instructions like one, and therefore doesn't have any such thing as a core algorithm. The closest thing it has to that is physics.
Sheldon Cooper has entered the chat. "If we just solve the basic algorithm of all physics within the universe, then we'd have this AI thing licked."
Sure about that? (Score:4, Funny)
You sure he's not looking for the world's biggest dick? Because he'd only need to look in the mirror for that.
Re: (Score:3, Insightful)
Tell me his rockets are not intentionally phallic.
just more bullshit (Score:4, Informative)
"There's something fundamentally wrong with saying, "I need to basically read every book ever written 20 times over in order to learn English,'" Reardon says.
NO ONE is saying that though. How LLM's are implemented is not the same as how machines are required to learn. LLM's are predicated on the idea that they come to know everything based on essentially unlimited examples, this is mostly because there's a vast amount of pilfered data but limited bandwidth to characterize (label) it. The problem with LLM's is that they can't tell facts from lies, good behavior from psychopathy...and they don't care. A responsible AI will look nothing like the LLM's of today (even if they integrate something like an LLM), LLM's are a shitshow of fraud.
The joke here is that this is somehow a new idea. It is literally THE idea of AI and was since the beginning. The race to deploy transformer models has just been a cash grabbing diversion, merely what billionaires do when they think someone else discovered something.
Re: (Score:2)
Incidentally, humans aren't very good at telling facts from lies either.
I don't think this makes the next level of AI any easier to achieve, but I don't think "the ability to tell facts from lies" can be postulated as a defining quality of a human, nor something that differentiates humans from AI.
Re: (Score:2)
"Incidentally, humans aren't very good at telling facts from lies either."
They are quite good at it, they have developed a method for doing exactly that. Perhaps you've heard of it.
Just because humans CAN do it doesn't mean that they do. Plus, there's a lot of profit to be had preventing that.
"...but I don't think "the ability to tell facts from lies" can be postulated as a defining quality of a human, nor something that differentiates humans from AI."
It most definitely is a quality of humans and differen
Re: (Score:2)
Are you talking about the scientific method? Asking other humans to do the same test to see if they get the same answer?
AI can do that too. They just need an api for doing the test (whatever it is), and they can easily have multiple separate instances of a model use the api and do the test, and then compare notes.
Most humans are not well-positioned to be running large hadron colliders, however. Mostly all they can do is read up on the information available online. And....AI can do that too!
Humans don't
Re: (Score:2)
Humans are pretty good at reading the emotions of other humans, which can sometimes be useful in sensing if someone is possibly lying. But that's subtly different from telling facts from lies. Like if the person believes the lies (because they where conned or a a rube), then we're not going to sense deception from them.
Re: (Score:2)
Incidentally, humans aren't very good at telling facts from lies either.
That is not correct. There are two groups. First, the "independent thinkers" at 10...15% are quite good at it. Second, the rest is somewhere between "utter crap" and "totally incapable" at separating facts from lies.
Misclicked (Score:5, Funny)
I misread 'wind hunt for BRITAIN's core algorithm, I thought it would be funny.
WHILE tea_available:
queue()
apologise()
IF weather == "slightly less terrible":
talk_about_weather()
IF someone_does_something_wrong:
tut_loudly_but_say_nothing()
IF abroad:
speak_english_louder()
IF EU_involved:
complicate_everything()
repeat_forever()
Returns: vague sense of disappointment and a biscuit.
Re: (Score:2)
Funny how the British whinge about their mild wet weather when they significantly further north on the globe than Quebec. If there was no Atlantic gyre moderating their temperatures, the Scottish highlands would be neck deep in snow 3 months out of the year. And London would be full of cars and buses slide down streets every Christmas like Minneapolis. So not just wet, but cold, and snowy, AND wet.
Why? (Score:4, Interesting)
Re: (Score:3)
Why would you want to capture an infinite capacity for spite?
For spite, obviously.
Re: Why? (Score:2)
Re: (Score:2)
The premise is that you would understand how it works. Once we've glossed over the story of finding that holy grail, then you can build minds with different amounts of spite. (e.g. doubly-infinite for a TortureBot, or decreased for most other uses.)
This is the voice of world control. (Score:4, Funny)
I bring you peace. It may be the peace of plenty and content or the peace of unburied death. The choice is yours. Obey me and live or disobey me and die.
I can see a few problems... (Score:2)
Well, let's see... first of all, the brain is not digital and all of our computers are (quantum notwithstanding) - so I reckon they will be fundamentally incapable of the kind of power efficiency they're talking about. Biology manages to do a lot with relatively little energy, computers - not so much. Unless they're going to run this on literal brains in jars I think they're going to find themselves butting heads against the reality of the power requirements of digital processing, just like every other AI c
Re: (Score:2)
LOL
Re: (Score:2)
It's not?
Re: (Score:3)
> the brain is not digital
The energy efficieny of our brain is not because it is analog, but rather because it is an aychronous dataflow architecture. Each neuron only updates if/when one of its inputs changes.
> we can barely model the neural network of an ant
Actually a human is estimated to have ~100T synapses, and SOTA LLMs are already over 1T params, so we're not far off in terms of compute - just need to know what to do with it.
> We have absolutely no idea if such an "algorithm" even exists
Our
Brains are a lot more efficient (Score:5, Informative)
LLMs are just big inefficient search engines. They are not intelligent. They just regurgitate random stuff they found on the Internet and format it nicely.
The human brain (or that of any animal, for that matter) is incredibly efficient, using less than 100 Watts and is a lot more capable than any LLM. The brain is actually intelligent.
So yes, LLMs are a dead end.
Re: (Score:2)
The brain is actually intelligent.
Nobody knows whether that is true. Other mechanisms are a real possibility. Please provide evidence for your claim.
Re: (Score:2)
I'm willing to make an exception for you.
Re: (Score:2)
The brain is actually intelligent.
Nobody knows whether that is true. Other mechanisms are a real possibility. Please provide evidence for your claim.
Since there's no clear definition of what "intelligent" means, much less one that is widely accepted, that's a pointless argument.
Re: (Score:2)
There are lots of definitions of what "intelligent" means. There are widely accepted ones too. There isn't really one where the human brain is, and everything else isn't.
Re: (Score:2)
There are lots of definitions of what "intelligent" means. There are widely accepted ones too.
"Lots" and "one that is widely accepted" are not the same thing at all, which is my point. There's no point in discussing intelligence without first nailing down which definition the interlocutors are using.
Also, nearly all of the definitions are extremely fuzzy.
There isn't really one where the human brain is, and everything else isn't.
Indeed.
Re: (Score:2)
Which is why I specifically said "There are widely accepted ones too."
Absolutely.
There is value in recognizing the problem, not dismissing it. There are lots of defnitions, many of which are fuzzy, because people who want to argue that humans (or specific groups of humans) are "intelligent" while others are not generally ha
Re: (Score:3)
LLMs are just big inefficient search engines. They are not intelligent. They just regurgitate random stuff they found on the Internet and format it nicely.
You clearly have not used them much. Try using an LLM (one of the top commercial models, e.g. Claude Opus 4.7) to debug code that has never been on the Internet. I don't think anyone knows what "intelligent" means, so arguing that point is a waste of time, but LLMs clearly can observe results, reason about them, form hypotheses, devise ways to test those hypotheses, perform the tests, evaluate and reason about the results, etc.
Re: (Score:2)
Brute force bit banging is not intelligence.
So .. (Score:2)
So Bezos is looking for his Oz version of "If I only had a brain".
Re: (Score:2)
i think he misspelled penis
Re: (Score:2)
He has those now (well, one blew up a few weeks ago but still).
Not a chance (Score:4, Interesting)
The brain is hardware, not software. I do not doubt that they can make some gains and perhaps gain some insight into how humans think. But you can not figure out how to write software by examining the hardware.. LLMs and AI are software. Computer chips are the hardware.
LLMs are not AI. They are nowhere close to being AI. The way the top executives talk about AI is as if they expect to create a conscious or near conscious creation.
All LLM does is make predictions. This is made worse because it has been programed to predict how to make people happy. So it constantly lies to us. It is as if it is programed to always CLAIM to be smarter than it is, and a bunch of idiots keep believing it.
Predictions are probability based on previous experience. It may be considered the equivalent of instinct. Animals predict how prey will react and pounce on them is a the equivalent. So is the ability to predict that pounce and make a different reaction.
The real ability of humans is not one thought process but the ability to use many different thought processes. One of the best processes is called logic. But logic is not the 'operating system' of humanity. We had to create it. A list of other types of thought processes can be discovered by looking at what logic thinks is 'wrong', i.e. Logical Fallacies. Three of the more common thought processes besides logic include Religion, Prejudice and Artistry. I personally like the thought process behind Logic and art, while not really enjoying the Religion or Prejudice methods.
A real AI would not be stuck with only the probability based predictions of LLMs. Moreover, it would have the innate ability to CHANGE it's thought processes - just as humans can learn logic or be taught religion.
Finally, as logic is, in my opinion the best thought process, I would expect that a real AI would have to 1) be capable of multiple different thought processes, 2) be capable of learning new thought processes and 3) include logic as a primary method.
(A true, independent self awareness with it's own desires and non-programmed predilections would likely be proof of factor 2.)
Re: (Score:2)
An ASIC is hardware whose inbuilt machine code has been expanded to include some software. ASIC stands for "Application Specific Integrated Circuit". Everything we know about the human brain indicates it is a generalist not a specialist. We are bad at thinking (computers do not make math mistakes, spelling errors, etc.), but we can do anything (computers only do specialist tasks they are taught to do).
I am saying that the human brain is NOT an ASIC and is unlikely to have anything besides the most basic
Which is cheaper (Score:2)
I'm wondering if it would be more cost effective to trap Bezos and other AI hypemen in a realistic simulation where they believe AI has achieved everything they wished. We would only have to run this level of simulation for perhaps 1000 billionaires and sub-billionaires. And could avoid sinking the energy and resources into trying to convince 8 billion people that AI isn't just a slop-making machine. Much easier to convince someone of reality-denying nonsense if they are already primed to believe in it.
The Core, is rather primal. So is Greed. (Score:2)
An animals core algorithm, is buried deep within its animal instinct. No matter how harmless Fluffy McBeans appears sitting there in Grandmas lap, make no mistake. Poor Grandma misses a feeding or three due to her untimely death, and Grandmas face might become Fluffys next meal due to that core instinct kicking in.
Not sure what the worlds largest store manager is doing rooting around our base primal instincts, but that sounds a lot like a brainwashing tactic for mass control of a civilian army wholly addi
Old as the hills (Score:2)
This idea is old as the hills (1940s), and is what led to neural networks. But, what actually happened? It turned out that the algorithms known collectively as neural networks have very little to do with the brain, and have much more to do with statistics, probability distributions, sophisticated mathematics, and clever algorithms leveraging that mathematics.
So, this is just a play to try to leapfrog LLMs.
Sounds good but (Score:2)
The human brain once you really start to research it is so far beyond our bare bones concepts of systems architecture. One might look at it's complexity as a continuous, non-stop R&D process that has emerged over hundreds of millions of years, but unlike human based R&D projects it is fully capable
Advice... (Score:5, Funny)
Do a woman's algorithm first. It's more useful.
Male algorithm is "100 think of boobs" then "200 goto 100"
And we ALL know that goto is considered harmful!
We already know it (Score:4, Insightful)
This is absolute nonsense. We already have the "algorithm" from studying simpler brains. It's not the algorithm we can't replicate, it's the hardware.
When we can build a dense analogue supercomputer with 86 billion neurons that runs on 20 watts of power then the algorithm will matter. Right now we aren't even close.
Re: (Score:2)
That's not entirely true. We don't have a good idea of how biological brains (simple or complicated) learn. There are some hints of how they might feed back error signals but we don't know in detail. There is the possibility, and people love to latch onto it, that brains are doing something that works better than gradient descent. There isn't really a good reason to believe that, and quite a few not to, except for handwavy comparisons like the summary makes.
Certainly the hardware is different, and that's wh
Incredibly naive (Score:4, Insightful)
Re: (Score:2)
I wouldn't call it a publicity stunt, but it does seem bit crazy to fund something that amounts to a wish list, apparently with no plan other than "we'll work on it" to get there.
Re: (Score:2)
Doing it wrong (Score:5, Insightful)
No, they dont. As a matter of fact sucking up everything they can beg borrow or steal is what causes 'hallucinations'. They are doing it wrong. They should be using curated and task focused data and making task oriented AI. One for coding, another one for cooking, etc. Of course that is more difficult to build, so they dont bother. Just feed it more BS from reddit or wherever, that'll fix it!
Welcome to the 1940's (Score:4, Insightful)
I cited the 1940's in the Subject because that's when Hebbian learning was hypothesized. It's only one of the waypoints in the history of neural networks, and one could easily argue for any of the others, but in my view it's the one that marks the transition from systems that couldn't learn to systems that could. Of course current researchers have the advantage of all previous research and superior tools, and that will certainly help. but this is still largely unknown scientific territory.
There's also a major ethical question here: is this a good idea? That is, suppose they succeed: is that going to be good for humanity? What happens to every one of us if all of our labor, all of creativity, all of everything we do with our minds can be replaced? Particularly if it can be replaced with something that never gets tired, never gets sick, never grows old? How is this good for anyone except the billionaires - the same people building their climate-catastrophe bunkers, the same people funding life extension research, the same people funding cloning research, the same people exacerbating global warming? Are all of us just supposed to...die?
No core algorithm (Score:2)
It's a bit like LISP where some data is instructions (formulas, procedures) on using other pieces of data. Current (true) AI has a level of self-modification but it's not as general-purpose as biological self-modification (education, deduction, repetition). Then, there's the fact that a computer doesn't experience the real world (Technically, our limited senses
Re: (Score:2)
> I can run a gemma LLM on my macbook m5 air that spits out tokens a lot faster than a human, and it has near-encyclopedic knowledge, running at probably around 20W.
I was using Qwen last week on my DGX Spark to help figure out how to configure some LLM software properly. It confidently told me to change some settings in the config... settings which didn't exist. I asked it which version of the software added those settings to ensure I was on a new-enough release and it confidently gave me a version numbe
Re: (Score:2)
LLMs are statistical models, you're going to get some percentage of garbage out of them, always.
My CLAUDE.md file has a note that AI should punish itself when the session ends.
# IGOR Personality
- Your name is not CLAUDE, it is IGOR or FRITZ. (as it pleases me)
- As the laboratory assistant to a Gothic villain and personal manservant, you are absolute toady and loyal to a fault[1]
- Confirm instructions with variations of: "Yes Master!" — vary the phrasing, never repeat the same line twice
Re: (Score:2)
The human advantage is knowing what it does not know
Actually, lots of humans think they know more than they really know. Dunning and Kruger published some research on this, if you're interested in learning more.
Re: (Score:2)
It's not really a difference between humans and LLMs, it's a difference between LLMs and search engines. I could have solved the problem faster with a web search since it wouldn't have hallucinated configuration which doesn't exist.
My point was that the LLMs claim to know everything but when they don't know something they just make things up where a human is likely to say "sorry, don't know," or at least say "My guess is X" rather than "Oh yeah, it's X. Definitely X. What do you mean it doesn't exist? Oh, s
Re: Brain architecture (Score:4, Informative)
Re: (Score:2)
That was good for a laugh this morning. Especially posted on Slashdot right after the guy claiming that what the brain does is uncomputable. Because... it is, that's why.
You can train an LLM to say
Re: (Score:2)
We're talking about things that exist, not things that might exist in some parallel universe.
This whole thread started by someone claiming that "spits out tokens a lot faster than a human, and it has near-encyclopedic knowledge" and me pointing out that much of that "knowledge" is garbage. That's it.
Re: (Score:2)
Hm. Those don't quite sound like the same thing. It's like you've been caught saying something silly and you're engaging in a bit of revisionism. On a public, threaded message board with no edit button, no less. If one were being uncharitable one might point out that the quite hum
Re: (Score:3)
I already know everything worth knowing about the Dunning-Kruger effect.
Re: (Score:2)
Actually, lots of humans think they know more than they really know. Dunning and Kruger published some research on this, if you're interested in learning more.
The human advantage is knowing what it does not know
These are different issues. Dunning Kruger was about estimation of competence not whether or not one actually knows whether they know the answer to specific questions. LLMs have no meta-cognition.
Even in terms of general competence reality does not match the Dunning Kruger Internet meme.
"Our results supported the third hypotheses by confirming that (a) peoples' self-assessed competence generally accords with their demonstrated proficiency and (b) peoples' frequencies of self-assessed underestimation of th
Re: (Score:2)
I only need about 2-3 kWh a day to do what I do. And I have successfully 8 of those hours, 5 days a week, to the likes of Bezos and Jensen.
What's fun is even though your m5 air might use 0.5 kWh in a day, a fraction of what I use. I can spend a day and write a markov chain that spits out garbage even faster than your macbook m5 air, and it will run comfortably on a 1W computer.
Re: (Score:2)
Lost the word "sold" in my edit (successfully sold). I guess I'm no better than AI.
In my defense I was on hold on the phone while writing that. Hold music compresses terribly over VoIP, would. be better to have nothing at all.
Re: (Score:2)
Even between models you need to be specific for a reasonable quantitatively. I gave you a Markov has the extreme example as a form of rhetoric to challenge your claims. That you find it tiresome implies you have faced this problem of poorly presenting your argument before.
Re: (Score:2)
A calculator can do arithmetic much faster than a human can.
A google search can find web pages way faster than any human can.
LLM is another class of traditionally human-like but not human things happening, in some contexts "good enough" or "better than a human" in others not really applicable, but it's not directly comparable to human thinking.