What Went Wrong at OpenAI? [audio]
slate.com
slate.com
This piece suggests that they have since focused (at least partially) on creating profitable products/services, because they need to show Microsoft that this investment was worthwhile.
Does anyone with more context know if this is accurate, and if so, why they changed their approach/focus? What are they working on and is GAI still a goal?
I wonder why this problem doesn't get much more attention in the field. It seems like models to solve even extremely domain specific, limited tasks already are running into economic / computional constraints.
Given that openAI wanted to be a place that develops general, artifical intelligence without having so much short term commercial focus I wonder why there is not more research going into approaches that are less computationally intensive. Which IMO is in itself a metric for intelligence
OpenAI has made huge initiatives in bringing in diversity and really opening their work (deepmind rarely ever does) and I think that profiting in the way they set it up only ensures they can do bigger and better research.
I understand that its a bit ironic, but I think the reporter needs to understand that strangers (esp. reporters) don't usually get this types of access regardless of whether it's "openai" and since the reporter is probably not under a NDA, any company would put tight restrictions of what strangers see and do.
I agree that OpenAI sometimes overhypes stuff (I think the gpt-2 case was unintentional) and some of the criticism she points out, but overall the piece in my mind was more of a hit-piece than showing that OpenAI does try to have a diverse group of people (fellows/scholars) and was probably the only nonacademic non-profit before needing large amounts of compute.
That contrasts with how Hao got a number of interviews:
"Many who work or worked for the company insisted on anonymity because they were not authorized to speak or feared retaliation. Their accounts suggest that OpenAI, for all its noble aspirations, is obsessed with maintaining secrecy, protecting its image, and retaining the loyalty of its employees."
Does "OpenAI" mean unfettered access for reporters? Nah, but within the context of the article, her pictures of secrecy at the org add color.
I'm curious what elements drew you to conclude it was overall a hit-piece.
I found the headline and a part of it trying to seem sensational e.g. "What Went Wrong?", which in spite of that it was a really informative article for me.
If anything seemed scandalous to me, its OpenAI's change from non-profit status and Altman's proclamation that
"In sharing his 2020 vision for the lab privately with employees, Altman’s message is clear: OpenAI needs to make money in order to do research—not the other way around."
Contrast that with:
"For one thing, the sticker price was shocking: the venture would start with $1 billion from private investors"
One billion is a lot of dollars! How in the world are they already in need of more funding?
I also wonder what Altman does there. For being CEO, he is mentioned only in passing in the article.
[0] https://www.technologyreview.com/s/615181/ai-openai-moonshot...
In what way?
Deepmind on the other hand doesn't open source stuff (I think it was alphafold that they promised to open source and didn't for over a year... i think they may have open sourced it recently).
>"So there were two main theories that came out of this initial founding of the field. One theory was humans are intelligent because we can learn. So if we can replicate the ability to learn in machines, then we can create machines that have human intelligence. And the other theory was humans are intelligent because we have a lot of knowledge. So if we can encode all of our knowledge into machines, then it will have human intelligence. And so these two different directions have kind of defined the entire trajectory of the field. Almost everything that we hear today is actually from this learning branch and it’s called machine learning or deep learning more recently."
Is there still development of the other branch, the "encode all of our knowledge into machines, then it will have human intelligence" branch then? If so what is the branch of AI called then?
Knowledge representation and reasoning is one of the main fields of AI research, on the same broad level as machine learning or robotics and with its own journals and conferences (KR 2020 will be held in Rhodes, Greece in September). It enjoys much less recognition than machine learning in software development circles because it doesn't receive such broad coverage in the lay press as machine learning does, but it's an active area of reserach. Google's Knowledge Graph is probably the best known example of appplications of the techniques that have originated in research from that field.
I don't really know why the author says that machine learning and KRR are "the two main theories" in the field. Perhaps she has access to historical information that I ignore. She says, a little earlier than the passage you quote that "[AI] was started 70 years ago", which must mean the workshop at Dartmouth College in 1956, where the term "Artificial Intelligence" was first introduced (by John McCarthy, perhaps more recognisable as the creator of Lisp to a programmer audience).
There's sure been many binarily-opposed "camps" in AI, like the symbolicists vs the connectionists, or the "scruffies" vs the "neats" and so on. While I recognise the "machine learning vs knowledge representation" as one of the classic dichotomies, I don't really think it's such an ancient and fundamental dichotomy as the interviewee makes it sound.
I wonder if the interviewee is mixing up the "ML vs KRR" distinction with a more subtle distinction between different forms of machine learning. I'm thinking of Alan Turing's original description of a "learning machine" from the classic 1950 Mind paper ("Computing machinery and intelligence", where he introduced the "imitation game"). Turing's learning machine would learn incrementally, from a small original knowledge base and from human instruction and from contact with the world, whereas today's machine learning tries to learn everything from scratch, in an end-to-end, no-human-in-the-loop, approach. This distinction, "incremental vs all-at-once learning" seems to fit the interviewee's description of the "two main theories" better.
There's a paper, "The child machine vs the world brain", by the Australian AI scientist Claud Sammut, that goes into some detail in this distinction, based on Turing's paper and later developments in data mining and big data:
https://www.semanticscholar.org/paper/The-Child-Machine-vs-t...
I recommend reading at least its introduction and then digging in to the references if you're interested in the history of AI in general and machine learning in particular and different ideas on those subjects that have been explored and abandoned over the years.
Warning: contains ancient lore.
I wonder if the journalist was being reductionist simply because having only two competing branches makes for a slightly more digestible and compelling narrative for the average reader?
I don't really know- I'm from the machine learning side of things. I study symbolic machine learning for my PhD and I have a background in logic programming so I know what KRR is, but I'm no expert and I'm not up-to-date with the work in the field.
I think the reason the lay press doesn't cover KRR and other AI fields is that they are, well, fields of academic research and as such not terribly interesting to the press and to most software developers. My intution is that most software developers don't really know much about AI in general, not to mention particular sub-fields of it.
This includes machine learning. It's just deep learning that is an exception to this (quick reminder: AI ⊋ machine learning ⊋ deep learning). We could spend a long time discussing the reasons for that but, basically, I think you don't hear about KRR that much because it's not deep learning :)
>"Pursuing a G.I., particularly with a long term view, was the central mission of open A.I.. And yeah, there was the traditional Silicon Valley talk of changing the world, but also this sense that if HDI was done wrong, it could have very scary consequences."
If the concern was truly avoiding AGI was done wrong which presumably includes its development being in the hands of a select few tech giants. Wouldn't it be better to simply wind the operation down rather take money one of those few tech giants leading in AI development then running a company with motives that are odds with each other?
Just off the top of my head doesn't it seem that Microsoft with its new billion dollar investment now stand to benefit from that first billion dollars of investment made to the non-profit OpenAI more so than anybody else?
The reporter was invited to do a piece on them, and while visiting had trouble reconciling their secrecy with their ethos of openness. She was not allowed to interact with the actual researchers where they were doing their work, and her lunch was pushed out of the building so she couldn’t overhear their all-hands meeting. (My take is that their openness extended to the curated fruits of research, but their process itself was guarded from any communication channel they couldn't control i.e. the reporter).
This seems related to the second part, where they discuss the pressures toward profit, from strings attached to corporate investments, which they suggest would be different under traditional long-term gov investments. And they talked about the paradox of adding a for-profit branch to a non-profit org, without resolution.
I’m a bit unsettled recently when listening to podcasts and stories like this that seem to end on a note of “shrug, capitalism, isn’t this an interesting problem?”. I’d be more encouraged to see folks talking about post–game-theoretic social structures that can categorically solve for these issues, that can allow us to transition out of capitalistic dynamics rather than trying to fight them in order to get work done. This seems to be the rallying call of the nebulous ideas behind “game~b”. Wondering if anyone here has been seeing that yet.
Maybe it’s just me but companies are not public benefit enterprises, even if structured in some way as a not for profit.
This wave of coverage and the dialog around it seems to come from the view that OpenAI somehow owes the world something. When in fact it only owes its stakeholders none of whom are reporters.
Our primary fiduciary duty is to humanity. We anticipate needing to marshal substantial resources to fulfill our mission, but will always diligently act to minimize conflicts of interest among our employees and stakeholders that could compromise broad benefit.
I recently had Google Maps "yelling" directions over a podcast,looked for a solution, and discovered Google Maps has volume adjustment for that (on Android, profile > settings > navigation settings > guidance volume). I toggled from harder to softer (there's a joke there) and all was good.