[freddealmeida]

I often wonder about this. Four things are needed to truly build advanced AI's (read deep learning, deep reinforcement learning): new algorithms, complex data sets, and advanced GPU based computing (optimally GPU in any case) but also an open community.

I think AI research is one of the most open, and this openness is really at the center of its growth. So I am happy OpenAI has started since it is within this vein of sharing that the community has already built. But I certainly don't fully grasp how it can open AI to the world unless it can share rather valuable data sets (often impossible to get data such as personal health record), and make computation much much cheaper.

Let me illustrate my concern; Alphago required not just 30m game sets and complex understanding of both a policy and value network design, but also 1000 CPU and 200+ GPU instances. Something on the order of a few million dollars to build and utilize.

I look forward to the work coming from OpenAI. I hope it lives up to the hype. But I believe AI will more than likely remain squarely in large enterprise since the cost of developing applications will be high for the short and possibly for the near long term.

[AndrewKemendo]

Four things are needed to truly build advanced AI's (read deep learning, deep reinforcement learning): new algorithms, complex data sets, and advanced GPU based computing (optimally GPU in any case) but also an open community.

Actually we have no idea what the constituent parts of AGI are.

What you mention are the current state of the art for narrow AI projects like classification and segmentation - which is basically 100% of machine/deep learning currently, but are not generalizable yet.

As an example the pre-eminant biologically inspired computing researcher Richard Granger is skeptical (and I agree) that parallel silicon will be able to scale to the flexibility that we see in biological learning (aka General intelligence).

Based on what I see so far from OpenAI I don't see them getting to AGI. They haven't stated it as an explicit goal, I think because they don't have a pathway (nobody does by the way).

[studentrob]

> Actually we have no idea what the constituent parts of AGI are.

Yup. We don't even have a good test for knowing AGI.

Frankly, we don't even know if we are AI or if everything is predestined.

We don't know if true randomness exists.

Currently no AI system can define its own goals. I'd like to know how Open AI would solve that. To me, Open AI seems like it will just end up as a giant open sourced machine learning toolset. That's great but not the initially stated goal, and they risk souring investors on future AI tech that actually has merit when they fail to achieve AGI.

[AndrewKemendo]

Yup. We don't even have a good test for knowing AGI.

Very true and I think significantly overlooked.

The closest thing I have ever found was the Universal Anytime Intelligence Test [1]

[1] http://users.dsic.upv.es/proy/anynt/

[studentrob]

That's a huge amount of links and information.

Can you summarize?

[AndrewKemendo]

Can you summarize?

Not really - it's worth reading the whole thing.

Best I can do is state that the test basically has multiple different environments with different rules that are independent of how capable or fast they are.

[maaku]

You are correct to point out that machine learning is NOT general intelligence, and what OpenAI is working on really have very little to do with AGI and super-intelligence, sadly.

But how can you say "we have no idea what the constituent parts of AGI are" or "they don't have a pathway (nobody does by the way)"? There is an active and vibrant (if sometimes eclipsed) AGI community. There is an annual AGI conference. There are a half dozen or so actively developed AGI projects with comprehensive architectures with attached roadmaps. It's an active area of research, but it's not like we have no idea how to build a general intelligence, or what such an architecture might look like.

[AndrewKemendo]

Uh, I go to the same conferences - in fact I'll be at AGI 16 this year and I was at AGI 14. Ben was my research advisor for my Masters.

I stand by my statements. The community, or even a handful of researchers haven't come up with a competent path to AGI. That's indisputable.

it's not like we have no idea how to build a general intelligence, or what such an architecture might look like

Show me one, I'd love to see it.

Listen, I love everyone working on them and many are my friends; but none of the attempts today have anything near the specificity of a project management roadmap to say with any certainty that AGI is even a probable outcome. Not OpenCOG, not Numenta, not MicroPsi.

That's not a hit on any of them either. The people and areas they are working on are awesome, amazing and fundamental to research but none of them would claim that they have a solid roadmap. Even the roadmap sessions at the conferences usually go nowhere because we just don't know enough about how generalizable intelligence works yet.

[maaku]

> I stand by my statements. The community, or even a handful of researchers haven't come up with a competent path to AGI. That's indisputable.

The keyword there is competent. You're making a subjective evaluation. Given your CV you must surely be aware that Goertzel has a 1,000 page book (two volumes, actually) laying out in great detail his roadmap to human-level intelligence. The leaders of other major projects in this space have their own ideas which they talk freely about at the AGI conferences, and are written down to varying degrees.

[argonaut]

> book... roadmap... ideas

Notice a pattern?

Meanwhile, in deep learning (and FWIW I don't think any deep learning researchers are under the illusion deep learning provides a path to AGI), there are:

working systems that outperform humans at narrow visual tasks (image classification, segmentation, etc.), a working Go bot, early prototype systems that caption images, the list goes on and on.

[argonaut]

You can say that. Someone working on AGI != they have a good pathway to AGI. And anyone telling you they have a good pathway to AGI is being ridiculously and naively over-optimistic (recall that back in the 1960s people thought they had a pathway to AGI - repeat this every other decade).

None of those projects have yet shown any real progress towards AGI. A roadmap and "comprehensive architecture" are just plans and conjectures, not results. The AGI conference is also still fairly niche.

EDIT: Agreed with below. I also want to clarify that I'm not saying AGI is unworthy of research. It's just total early stages right now. Be aware that progress is incremental, and maintaining momentum (and research funding) is predicated on delivering tangible results and tools incrementally.

[studentrob]

Yup. There are also organizations who meet and research about contacting aliens. Doing research is not evidence of validity or progress.

One thing to note though is these efforts are often put forth by some hard working folks and the product can be something great.

For example, "Thinking Machines" [1] [2], founded in 1983 in Cambridge and defunct in 1994. The folks who spent effort there would later take their knowledge of parallel programming and built a tool called An Initio as a successor company. The tool was way ahead of its time. It'd maximize resource usage on a machine in parallel without much technical knowledge needed by the developer. They made a boat load off of deploying this tool on internal datasets at many major companies for what was then called "data warehousing".

[1] https://www.technologyreview.com/s/406781/thinking-machines/

[2] https://en.wikipedia.org/wiki/Thinking_Machines_Corporation

[tlb]

I believe AI will more than likely remain squarely in large enterprise since the cost of developing applications will be high.

For perspective, the same thing was said about search engines in 1995. At the time, server-class computers (big DEC and Sun boxes) were expensive. Altavista had a computer nobody could buy. But it turned out, with some cleverness you could use commodity PCs instead. NVidia is doing impressive work to bring the price/performance of hardware down. I think the vast majority of applications will be possible on a startup budget.

I also think data sets will become more common and less critical over time. Unsupervised and reinforcement learning require less external data. And so much data is available openly.

[vasilipupkin]

"For perspective, the same thing was said about search engines in 1995"

yeah, and since then, we've basically had 2 major search engines developed. Google and Bing :) and bing development cost billions

[jsemrau]

And even with all that investment Bing is not even working that well. It is good, but not on a Google level.

[maaku]

Do you realize how much computing power is behind google search?

[taneq]

I'd be interested in the amount of computing power per searchable web page. Sure, there's a huge amount of computing power behind Google's web services but I'd lay odds that it's much more efficient than it used to be. The internet is huge.

[doctorpangloss]

> how it can open AI to the world unless it can share rather valuable data sets

I think the spectacularly publishable performance from industry comes entirely from these massive datasets.

AlphaGo needed way more energy and data than a person did to perform as well as it did. To a set of pedantic but extremely well-meaning researchers, that could be interpreted as a huge failure.

Clearly whatever it's doing isn't the kind of intelligence that people have, since people don't need megawatts of power and 30 million game sets. But that's what big companies have that academic departments don't.

Journals publish both approaches though, and we all win from great discoveries.

[vhold]

I think you're coming it at it from the perspective that making Lee Sedol was free? Imagine how many calories, all the fuel used to create those calories, his own direct transportation footprint, the time and money spent across creating him and his entire life. AlphaGo might not be relatively that inefficient.

[sqeaky]

Let's do some Math!

Assuming 2000 Calories per day a megawatt is about 240.000 calories per second we can do some silly calculations. Assuming AlphaGo consumed one megawatt for 1 whole day then it consumed about 20,736,000,000. This is probably low. This is enough Calories for a human for 28,405 years. If we split the Calories accross 500 that accounts for 57 years apiece.

I had no clue where this would turn out when I start. If we presume AlphaGo can run on a 1,000 Watt machine we can still divide all these numbers by 1,000 and they still hold, so 28.4 years of human calories.

[daveguy]

Re:computation. I wonder if there are talks between OpenAI and BOINC to open up some of their computational resources. With the new OpenAI environments platform they include a peer review/validation process. That may help concerns that BOINC resources would be used commercially.