Perhaps I haven't been clear. I have no issue with the research OpenAI is performing, nor with anyone's beliefs in AI's imminence or their personal role in bringing it about. However, no one knows whether what they're doing is even on the right path towards AI, and certainly not when it will be achieved, plus the topic has been subject to overoptimism for decades now, so I do take issue with publicly calling what you do "working on AGI" or "pre-AGI" even though you have no idea whether that is what you're doing. Hopes and aspirations are good, but at this stage they fall far short of the level required for such public proclamations. My issue is with the language, not with the work.
I think you're issue with the language is not shared by most people. In research we rarely know beforehand what research is on the right or wrong path. But we are comfortable with someone saying they are researching something even if they don't know beforehand whether the research will be useful or a wild goose chase. For example most people's first though to hearing "I'm researching ways to treat Alzheimer's" isn't "Only if it passes phase 3 trials!".
Yeah, in this release they're not saying they're doing research towards AI, or even that they're researching AI. They're saying that they're "building artificial general intelligence" and developing a platform that "will scale to AGI." (emphasis mine) They're also calling what they're actually building "pre-AGI."
> We’re partnering to develop a hardware and software platform within Microsoft Azure which will scale to AGI.
This sentence might by itself imply they are farther along than they are, but in the context of the whole article I never got the impression they were close to actually building an AGI.
> The most obvious way to cover costs is to build a product, but that would mean changing our focus. Instead, we intend to license some of our pre-AGI technologies, with Microsoft becoming our preferred partner for commercializing them.
This read pretty straightforwardly to me. Pre-AGI seems like a shorthand for useful technologies like GPT-2.
Reading the article I never got the impression they'd solved AGI, or were even close. The context of the article is a partnership announcement not a breakthrough. I could see how a few people who are very unsophisticated might get a little confused as to how far along they are. But I assumed they were writing for people who had heard of OpenAI which pretty much eliminates anyone this unsophisticated.
They don't know what connection, if any, what they're doing has with AGI. For all we know right now, some botanist researching the reproductive system of ferns is as likely to bring about a breakthrough in AI as their research is. To me this feels like peak-Silicon Valley, the moment they've completely lost touch with reality.
People may also not be confused if Ben and Jerry's start an ice cream ad with mentions of AGI and the change of human trajectory and Marie Curie, and name it Pre-AGI Rum Raisin, but that doesn't mean the text isn't a beautiful and amusing example of contemporary Silicon Valley self-importance and delusion, and reads like a parody that makes the characters in HBO's Silicon Valley sound grounded and humble. Especially the "pre-AGI" bit, which I'm now stealing and will be using at every opportunity. Maybe it's just me, but I think it is quite hilarious when a company whose actual connection with AGI is that, like many others, they dream about it and wish they could one day invent it, call their work "pre-AGI." Ironic, considering they're writing this pre-apocalypse.
Although, perhaps you would agree that someone saying "my work on Alzheimer's might help your friend" would be behaving in a cruel and unprofessional way unless the treatment was indeed in human trials?
Do you hate all marketing or just around AI in particular? Would you bat an eye at MS investing $1B into a project + ads about say a new GPU architecture promising how "games will never be the same" because the new hardware (or even Cloud Integration) lets developers efficiently try and satisfy the rendering equation with ray tracing?
FWIW I thought you were clear, but there are only so many middlebrow dismissals one can make towards AI or AGI efforts and I think I've seen them all plus the low-value threads they generate. (I've made some too, and suspect we might get brain emulations before AGI, but I try to avoid the impulse and in any case it doesn't stop me from hoping (and minor contributing) for the research on the article's load-bearing word "beneficial" to precede any realistic efforts of building the actual thing. At least the OpenAI guys aren't entirely ignorant of the importance of the "beneficial" problem.)
I don't hate this copy at all; I absolutely love it! I think it is a beautiful specimen of the early-21st c. Silicon Valley ethos, and it made me laugh. Pre-AGI is my new meme, and that means something coming from a pre-Nobel Prize laureate.
What I'm interested in is how many dismissals of AI, most end up justified, can the field take before considering toning down the prose a bit, especially considering that the dismissals are a result of setting unrealistic expectations in the first place.
> the topic has been subject to overoptimism for decades now,
But so has every other big idea that went on to become reality, like planes (da Vinci was drawing designs for planes over 400 years before the first working ones).
> no one knows whether what they're doing is even on the right path towards AI
This is completely wrong. That would be like saying "no one knows if working on a wing is on the right path to flight".
Look at the way deep learning works. Look at the way the brain works. They share immense similarities. Some people say "neural nets" aren't like the brain, but that's not true--they are just trying to not over-exaggerate the differences which laymen commonly do. They are very similar.
> But so has every other big idea that went on to become reality, like planes (da Vinci was drawing designs for planes over 400 years before the first working ones).
And so has every other big idea that didn't become reality, and that was the majority. Again, I have no problem with AI research whatsoever, but the prose was still eyebrow-raising considering the actual state of affairs.
> They are very similar.
They are not. The main role of NNs is learning, which they still mostly do pretty much with backpropagation gradient-descent (+ heuristics). The brain does not learn with backpropagation.
The most famous is the Philosopher's Stone: a substance that can convert base metals into gold.
But in itself, that was not the point. It would also transform the owner or user -- it was a hermetic symbol, a mechanical means to "pierce the veil" and to see the deep mystical and magical truths, the Real Reality. It was immanent, a thing in the world, that enabled the transcendent, to go beyond, above, outside of the world. Its discovery would have been the single most important moment in the history of the world, the moment in which humans had a reliable road to divinity.
Hmm. Sounds familiar, doesn't it?
But out of alchemy came modern chemistry, and also some parts of the scientific method. After all, as some smart people worked out, you could systematically try all the permutations of materials that your reading had suggested as possibilities. That meant measuring, weighing, mixing properly, keeping detailed notes. Fundamental lab work is the unglamorous slab of concrete beneath the shining houses of the physical sciences. There were waves of hysteria and hype, but after each, something useful would be left behind, minus the sheen of unlimited dreams.
Hmm. Sounds familiar, doesn't it?
These days it is possible for a device to transmute base metals into gold. But the operators have not, so far as I can deduce, ascended to any higher planes of existence. They have eschewed the ethereal and remained reliably corporeal.
> Neuroscientists have long criticised deep learning algorithms as incompatible with current knowledge of neurobiology. We explore more biologically plausible versions of deep representation learning, focusing here mostly on unsupervised learning but developing a learning mechanism that could account for supervised, unsupervised and reinforcement learning. The starting point is that the basic learning rule believed to govern synaptic weight updates (Spike-Timing-Dependent Plasticity) arises out of a simple update rule that makes a lot of sense from a machine learning point of view and can be interpreted as gradient descent on some objective function so long as the neuronal dynamics push firing rates towards better values of the objective function (be it supervised, unsupervised, or reward-driven). The second main idea is that this corresponds to a form of the variational EM algorithm, i.e., with approximate rather than exact posteriors, implemented by neural dynamics. Another contribution of this paper is that the gradients required for updating the hidden states in the above variational interpretation can be estimated using an approximation that only requires propagating activations forward and backward, with pairs of layers learning to form a denoising auto-encoder. Finally, we extend the theory about the probabilistic interpretation of auto-encoders to justify improved sampling schemes based on the generative interpretation of denoising auto-encoders, and we validate all these ideas on generative learning tasks.
