| |
Before computers as we know them today were invented, they existed in a different form. |
| A computer was actually a job title for one who computes. |
| |
| Harvard "Computers", circa 1890 |
| Naturally, every computation was done by hand in the absence of semiconductors. |
| Productivity was limited, as was output, and there was no possibility of exponential growth. |
| But the one metric whereby the human brain continues to outperform the most advanced computers of today is in its power efficiency. |
| The human brain operates at roughly 1 exaFLOPs, which is roughly equivalent to the Frontier supercomputer – currently the No. 2-ranked supercomputer in the world. |
| But it does so at only 20 watts of power. |
| That compares to the 21 mega watts required to operate Frontier. |
| That means that the human brain requires about 1 million times less power than a supercomputer – a remarkable difference in power efficiency. |
| This factor has been the motivation for some to pursue the development of biological computers, in hopes that powerful supercomputers can be built and operate on a tiny fraction of the power required by today's data centers. |
| A Biological Breakthrough? |
Making waves around the internet in the last week was a development at Cortical Labs, a private biological computing company in Australia. |
| The media loved it, with splashy headlines like, "Human brain cells on a chip learned to play Doom in a week." |
| |
| Source: Cortical Labs |
| According to the announcement, not only did the "human brains on a chip" learn to play the computer game Doom in a week… |
| It did so… consuming significantly less energy than today's computer chips. |
| There is some truth to what was announced. |
| But there is a whole lot of context that was left out. |
| Devil in the Details |
| Cortical Labs has been developing a biological computing system based on human brain cells. |
| Specifically, it develops human neurons that have been derived from human induced pluripotent stem cells. |
| Yes, this could certainly raise some ethical concerns, specifically if the neurons were sentient. But there is certainly no evidence that they are. |
| Putting that aside, Cortical Labs developed its CL1 biological computer, which has – at its core – a biological semiconductor that contains about 200,000 neurons. |
| The CL1 system shown below is designed to power the biological semiconductor and keep the neurons alive for up to six months. |
| We can think of the unit as a life support system for the neurons. |
| |
| For comparison, however, a human brain has about 86 billion neurons, so the CL1 has about 0.0002% of the horsepower of our human brain. |
| Cortical Labs developed an interface for its biological computing platform that enables a developer to use the software programming language Python to utilize the CL1 system. |
| And collaborating with another software developer, Cortical's biological computer was used to play the computer game Doom, which was a huge game back in 1993, a game that I well remember playing myself. |
| Again, the media misinterpreted the development as human brain cells learned how to play a computer game, but the details matter. |
| Cortical Labs figured out a way to map the video feed of the game into patterns of electrical stimulation, which they could feed to the human neurons. |
| When the neurons were exposed to the electrical stimulation, they would spike in response and then respond with some kind of electrical signal. |
| There was no way for the CL1 and the neurons to "see" the gameplay at all. |
| It was just an electrical response to an electrical stimulus. |
| After a week of training, the human neurons were capable of some level of learning. |
| They could eventually seek out an enemy in the game, shoot, and navigate the mazes in the game. |
| But, by the company's own admission, after the training, the biological computer "plays like a beginner," showing only a slight improvement over completely random play. |
| To be very clear: It had no understanding of the game, no ability to plan ahead or be strategic with the game, and no path towards gaining the ability to play anywhere near a human level. |
| It was nothing more than a feedback loop demonstration – just an interesting research project. |
| And yet, Bloomberg interpreted the development as "an experiment that could one day challenge chips from companies like NVIDIA." |
| You've got to be kidding me. |
| Recommended Links
Red Alert Warning from Jeff Brown Jeff believes Nvidia is about to trigger a major crash on March 16. In fact, he says it's a "sure thing." During a special event tomorrow, Thursday, March 12, at 2 p.m. ET, in you'll discover: - Why this upcoming crash is a feature of the AI boom, not a bug
- Why the recent volatility is just the beginning
- The name and ticker of a wildly popular AI stock that could crash
- The name of a pick that could see huge gains
- And more…
Click here to register with a single click >>>
(When you click the link, your email address will automatically be added to Jeff's guest list.)
Up to 10x Returns From the AI Energy Boom? If you 10x'd the oil market during the last energy boom, you could've turned $1,000 into $47,000. Now artificial intelligence is driving a new boom in American energy that could be worth $7 trillion. A former money manager for Saudi Arabia's sovereign wealth fund is stepping forward with the best way to play this incredible energy boom. This has nothing to do with buying a stock, a crypto, or overhyped (and underwhelming) royalty deals. It's an overlooked strategy with incredible upside potential and the unique ability to insulate you from the boom/bust cycle. Click here now to learn more.
| |
| An Interesting Experiment… and Just That |
| No, that's not going to happen. |
| It is an interesting scientific experiment, probably better suited for an academic institution rather than a company in the private sector. |
| As we've been tracking closely in The Bleeding Edge for a decade, the cost-per-unit of compute using advanced semiconductor technology has been declining exponentially. |
| And that pace has been accelerating in the last few years. |
| Far more important to both industry and governments is how to scale computational performance at an accelerated pace. |
| There is no priority whatsoever to reduce overall power consumption. |
| Everyone knows that to achieve productivity gains and economic growth, computational resources must increase, and with that, so must energy production. |
| The well-proven adage that there are no high-income low-energy countries rings true. |
| All rich countries have an abundance of energy. |
| Scaling the Cortical Labs system is not economical, nor is it operationally efficient. |
| Growing, maintaining, and replacing (every six months or less) a data center of billions of neurons is arguably impossible. |
| And employing an electrical feedback loop system like this for meaningful computational tasks is completely unproven. |
| We should also keep in mind that the game Doom was so simple that it ran on a computer system with just 4 MB of RAM and a 386 processor, which is less powerful than some calculators available today. |
| Given these stark realities, what was the point of this announcement then? |
| And why all the fuss? |
| Follow the Money |
| Back in 2021, Cortical Labs did something similar with an earlier version of its biological computer. |
| It "trained" its system to play an even simpler game – Pong – as a proof of concept. |
| It used that development to eventually raise about $10 million in early 2023 to continue its research. |
| My guess is that Cortical Labs is getting very low on funds. |
| The timing of the announcement and media coverage is an obvious marketing tactic used to try to raise its next round of private funding. |
| While it may find some takers, an investment like this makes no sense at all. |
| For brain-inspired computing, a far more interesting and viable approach comes in the form of neuromorphic computing – designed to mimic the human brain and its neurons without having the difficulties associated with having to manage and care for living brain cells. |
The other exciting alternative is semiconductor companies that are using analog semiconductor designs for AI inference that have dramatically lower energy use, like Unconventional AI, Mythic, and EnCharge AI. |
| These approaches are where the investment dollars will flow… because they can scale at the speed of semiconductor technology. |
| Jeff |
| |
0 التعليقات:
إرسال تعليق