In short:
The US Federal government has mandated that AI model developers must build their models as to ideologically comply with and reinforce the Biden Harris ethos or be cut off from capital, federal contracts, and working with federal partners.
Additionally, those who comply will be heavily subsidized.
Welcome to the start of the American capture of the infosphere. Mark this date.
I read it very differently. Recent AI developments have headed us into crazy town, towards a enevitable seeming break down of our social structures. A “Fake news 2.0” amongst other challenges.
It’s ludicrous to watch our tech leaders release new AI models without ethical consideration or any form of reponsibility.
I work within a university and socially impactful research goes through rigorous ethics committees before being approved. Where was this with Open AI’s release of ChatGPT for example? Sam Altman just tweets after release words to the effect of, oh regulators should probably be considering the ethics around this. What a farce! Where are the ethics of those releasing these half baked tools?
Having the White House step in and provide leadership and bring it’s power to bear is totally appropriate and responsible in this moment.
By all means, if you believe (as it sounds like you do) that this will result in the affirmation and expansion of your worldview this is good news for you.
If you suspect, at any level, that maybe the current formulation of what is "ethical" and what is not may, at any point in the future, be called into question, than designing the fabric of super intelligent computational processes to presumptively adhere to your preconceptions is at best ill conceived at at worst malicious.
> I work within a university and socially impactful research goes through rigerous ethics committees before being approved
This is a gross exaggeration. Only a small portion of academic research requires IRB approval, and it's mostly limited to experiments on humans or animals.
If we were to start requiring IRB pre-approval for all research that might be "socially impactful", it would be the end of progress as we know it.
The first step would be to determine what counts as "socially impactful". You couldn't limit it to just AI. What about security and privacy? Why not all of CS? And then, why not Economics or History?
Source: worked for various universities doing research, first as graduate assistant, then scientist, then tenure-track professor, for 19 or 20 years.
I think there's a big difference between "socially impactful" and "enevitable seeming break down of our social structures"; that difference being the potential (direct) downside. It seems unlikely that most history or economic research would have the same downsides as AI.
Are you working in medical/health? There are very different ethics bars for different fields. Do you see no value in having ethics committees?
Is your argument that they took appropriate precautions before releasing their paid tools publically?
My point isn’t that these AI research houses should have the same ethics committee as a university, it’s that they should have ethics committee appropriate to the risks around the work they are doing.
What kind of AI do you want employed in a government setting that doesn't follow the policies of the government?
Additionally it sounds like the capital is net new. Which means if the government took no action, the would be no capital. Is it inappropriate for the government to offer programs to encourage things it wants?
At the other end of the spectrum, the government punishes companies that do things the administration doesn't like. In the middle, the government does nothing. It would seem to me that the alternative here is something everybody doesn't want, no?
LLMs are already being used as mediums of information access and will continue to grow in that role.
Information filters/access has a huge role in informing public sentiment, that makes LLMs a massive propaganda vector to be employed for whichever purpose it's controller's wish.
Their effectiveness IN that role is also likely to be beyond anything we've ever seen before.
Literally everything an LLM can do can also be done by a human (and substantially worse), but for some reason you inherently trust them?
Hammers are better at hitting nails into wood than a human hand, but we're not all losing our minds at the fact that you can also smash a human skull in with a hammer. The laws apply to the human not the object, and that remains true here.
Almost, but not quite. I'm quite sure the government itself will run IA without any compromise against whatever problems it thinks it has (either foreign or domestic). So this would regulate privately run AI but not government run black/undisclosed AI.
In this scenario they would quite likely grant Palantir-GOV an exception to do whatever was necessary to give them a leg up no matter how indiscriminate or discriminating it were, but also stipulate that this is not to be shared with other nation states, except those the GOV okays.
Like when the GOV procures munitions it's not saying, "please be sure they don't kill the enemy's innocent livestock".
You are probably correct in that the government MUST recognize that non handicapped AI models are more useful, but to believe so requires looking past what the current representatives are actually saying and doing.
If a person repeatedly says they are going to close a door, walks over to the door, grabs the handle, and begins to pull it closed... Is it really outside the realm of possibility that they may, in fact, close the door?
Sure. But they’re showing you the front door being closed and aren’t letting you in that there is a back door they’re not telling you about which they intend to keep open. Politicians may shake ands and kiss babies, but they're total "killers" and not the nice people looking after you or their constituents as people think.
Considering the preliminary outline of the AI BoR specifically states that AI models must be prevented from reaching conclusions or promulgating information the authors don't like, regardless of if the model has supporting information in it's training data, that result is almost certain.
+ Safe and Effective Systems
-> What should be expected of automated systems
--> Risk identification and mitigation
What they should really be focusing on is Hazard/Reliability engineering. "Risk" is too broad of a term for safety because it includes "economic harm" which ... well it's by decree.
+ Algorithmic Discrimination Protections
The only task these system perform is discrimination, and it's impossible to tell what the system can infer -- the Federal government will do anything to avoid naming the root causes of harmful inequality experienced by protected classes (and addressing that instead of holding a butter knife up to the neck of a shiny object).
+ Data Privacy
This is actually impossible without trust. Homomorphic encryption isn't a thing. This isn't happening.
+ Notice and Explanation
-> Why this principle is important
Chicago kept a secret watch list of violence-prone individuals. https://chicago.suntimes.com/2017/5/18/18386116/a-look-insid... is this important? If it was derived from OSINT, and it was being used for public safety purposes why is the federal government meddling in it -- do they have a monopoly on secrets? They shouldn't, if the public safety of Chicago is to be seen.
edit: I'm not sure what the criticism is here in light of the Notice and Explanation requirement -- there was N&E at most 4y after the practice began -- and it wasn't really a gaffe, presumably such a risk assessment system is still being used. I guess the argument is that this should have been required, or made earlier?
+ Human Alternatives, Consideration, and Fallback
This is probably what we need more of at first, and then less of later -- once all these goofy flat "simple" decision models have been replaced.
Here's a clear drawing of the ideal system without all the pesky words: https://ibb.co/7KrgBk4
I'm promise you I'm not trying to be mean or hyperbolic here but...
What makes you think "they" want you to survive?
What policy or statement have they made that affirms their interest in your life or the lives of your interest groups?
Do they want to/have they already disarmed you?
Do you control a means of producing your own food? Your own energy?
Are they allowing people into your areas to compete with you in labor markets and for control over limited resources?
Are they meaningfully protecting the future of the planet for you and your children?
Are they encouraging or discouraging you from having children at all?
Meatbags are inefficient ancestors that are not needed on the planet anymore. Small population (say a 1000 bodies) will be kept around in a designated area. Rest will be culled to save planet’s resources.
It seems quite clear that we are just a few years from being able to wire up less than 20k of nvidia hardware and produce a GPT-4 class model. All the methods and datasets are out there. Plenty of pre-trained base model weights too. The horse is out of the barn.
I would bet money against that. Replicating GPT-4 pre-training with current hardware would cost about 40-50m in compute. Compute will continue to decrease in cost and algorithmic improvements may allow for more efficient training, but probably not 3 orders of magnitude in a few years. I think there will be plenty of open source models that will claim GPT-4 quality, and some of them will be close, but they will be models that used millions of dollars (probably from some corporate benefactor but possibly from crowdsourcing) in compute to train. You will probably be able to fine-tune and run inference on fairly cheap hardware, but you can't cheat scale. It's going to take a major innovation to move away from the expensive base model paradigm.
Especially if you consider that as compute costs decrease, so does the ability of scale players to process larger datasets.
If we extrapolate that relation, you eventually reach a point where the biggest player can collect and process the most information and produce an ever-evolving model to maintain that relation.
Better hope it's creators have your best interests at heart.
I did my own calculations based on plotting loss on benchmarks compared to models with known parameters and training data, as well as using a quote from Sam Altman that said that GPT-4 would not use very many more parameters than GPT-3. Based on this, I estimated that GPT-4 probably used about 250B parameters, and since I had an estimate for the total compute I was able to estimate that the training data was about 15T tokens. 250B parameters times 15T tokens times 6 (https://medium.com/@dzmitrybahdanau/the-flops-calculus-of-la...) means the compute was about 2.2510^25 FLOPs. I estimated that A100s cost about $1/hr and can process about 5.410^17 FLOPs at 50% efficiency per hour. Therefore, the compute cost would be (2.2510^25)/(5.410^17) or about $40 million.
Interestingly, my own calculations lined up pretty well with this calculation, although they approached the problem from a different direction (a leak by Morgan Stanley about how many GPUs OpenAI used to train GPT-4 as well as an estimate of how long it was trained):
https://colab.research.google.com/drive/1O99z9b1I5O66bT78r9S...
1. We don't know what the number of parameters is, could be 175B, could be 250B, could be 400B. Ok, let's stick with 250B.
2. Training data: GPT-3 was trained on 300B tokens. It already used most of the high-quality data available on the internet, but let's say they somehow managed to find and prepare three times as much high quality data for GPT-4. This means GPT-4 was trained on about 1T tokens.
3. 5.4e+17 FLOPs/hour means 150TFlops, which is half of the BFLOAT16 max theoretical output, sounds reasonable.
4. $1/A100/hr is reasonable.
OK, so we need to divide your cost estimate by a factor of 15: Total cost to train GPT-4 comes out to be around $2.7M.
Regarding Altman's statement about "more than 100M to train GPT-4" - I'm pretty sure he was talking about the total cost to develop GPT-4, which includes a lot of experimentation and exploration, many training runs, and many other administrative costs which are not relevant to the cost of a single training run to reproduce the existing results. Just salaries alone: ~200 people worked on GPT-4 for let's say half a year, at $400k/year: 0.5 * 400k * 200 = $40M.
Government gets involved in cutting edge technology and insists on (and funds) a "diverse workforce" initiative in an area where less than 1% of software engineers have the capacity to work on these systems at a low level. The end result is often (not always) roles created where these new hires are so far removed from core activities that, as a matter of job security, they have to justify their roles with initiatives and requirements that cause all manner of dysfunction in the organization; and that establish a power dynamic where the company's mission shifts from customer and safety focus, to pacifying employee agendas.
Your analysis flat out refuses to acknowledge the inherent value-add that diversity in general brings to the field of AI. Also the only part in the whole thing where the word "diverse" is used is when talking about supporting higher education to create a diverse workforce. This is about education and support not quotas. You said "less than 1% of software engineers have the capacity" and that's exactly what it's trying to solve
The notion that diverse groups produce better innovation is a thoroughly researched and widely accepted tenet of modern organizational behavior research.[0] I grabbed the top google results for it, but I learned about it in school a decade ago.
And like the other commenter noted, the Institutes component seems focused on fostering new AI talent all over the country, so ergo a part of that will be making its resources accessible to many different groups of people. A claim that this initiative would be more effective in its stated goals by confining its educational resources to a single population demographic seems like it would also demand research and concrete evidence. Would that even be legal?
"Diverse" is one keyword in a whole paragraph about introducing an NSF funded program to provide educational resources for new workers to skill up on AI. America is a diverse country: Of course a program that serves all of America with that stated goal would create a diverse workforce. I'm a little confused why the top discussion on this thread has seized on that single word with such outrage.
Unfortunately that relationship has been disproven many times as it rests on a very weak definition of 'innovation' (Often substituting novelty in the form of simple recollection of terms or base concepts, regardless of how well those teams can convert those concepts into timely efficient and effective solutions.).
If you were taught this a decade ago you likely are familiar with the "Striving, Storming, Thriving" study which demonstrates that relation especially well.
I recall a recent hand-wringing study out of Harvard jam-packed with soothing language but showing that as teams diversify their ability to engage in shared risk-taking declines [0]
Well, that's interesting to learn - thank you. I'll have to reach out to my old professor since he keeps up with the state of the art, I'm curious what the rest of the current literature looks like.
On the other hand, I still don't share the anguish over the term 'diverse' in this particular OP for the reasons I stated - if this program is designed to foster AI talent all over America, if successful it will by definition foster diverse talent.
Like if the program wasn't designed to help America build a diverse AI workforce, then it would be designed to create a demographically constrained, homogenized workforce. I wouldn't expect us to be happy about that and I wouldn't think of it as a good allocation of federal tax money.
Something not being designed to do X does not mean it is designed to do Y.
The hackles raised here over "diversity" are, as you note, likely related to demographic considerations as opposed to say geographic considerations.
The argument is usually that the goal of government programs should be to accomplish a task as effectively and efficiently as possible. In the case of AI doing so is especially important in an era of rising geopolitical multipolarity. Sacrificing this for the sake of appeasing identity interest groups is short-sighted.
As we briefly explored, arguments that demographic diversity augment this efficacy have yet to meaningfully materialize, while concrete evidence to the contrary has both in the literature and in large-scale behavioral patterns (the neighborhood trust studies come to mind). It's one of, if not THE, most contentious area of sociology and public policy right now because of the obvious implications.
IMO that's why people latch onto it and why it's getting so much attention here. It's a hot button issue of grave importance that hasn't been solved yet and discussions around it are extremely painful and frequently stifled.
I appreciate the reply and I think you have accurately highlighted one reason this general area is such a trigger point on the board.
In the context of this specific initiative, my view would still be: America is an inherently diverse place, so a program that seeks to develop an AI workforce across America but does not create a diverse one has not achieved “maximum effectiveness” and has not effectively served all of America.
That may very well be true "in your view" but, if you'll excuse my pessimism, that doesn't mean very much if it results in China or India developing more effective models faster and more cheaply.
We may be diverse, but that needn't mean every sector of every niche in every industry must be so or that it would be better if it were.
Our position in the world is relative. Our power comes from the relative delta between us and "them", whomever it is we happen to be competing with.
To forget that fundamental relation, even in service to our personal sense of morality, is solipsism.
Well, that is your view. We are both just voicing our opinions, so without any concrete evidence neither of us can fairly state one as an outright fact, pessimistic or not. I will respond to your points, but since our conversation began as a discussion about academic behavioral research and is now 100% an exchange of opinion between non-experts, I'm going to get off at this station.
>if it results in China or India developing more effective models faster and more cheaply
Why should we project that possibility? The most successful research centers for new AI development have been American organizations like OpenAI and Google with clearly stated diversity policies. And of course, this OP is not about private companies, it's about a public resource program. Do you have any evidence that China and India are developing more effective models faster and more cheaply and that it's because of some specific lack of attention paid to ensuring public resources are broadly accessible to all demographics?
>We may be diverse, but that needn't mean every sector of every niche in every industry must be so or that it would be better if it were.
Rather than address this whole point, I'll revisit that we are talking about a federally funded public education program. I do not think a federally funded program to develop a bigger AI workforce in America would be more successful if it did not serve all Americans, and especially not if some of the funding were diverted to gatekeeping its resources to specific demographics.
This is strictly personal opinion, but mine is: America is not China or India and that extends to the values enshrined in our founding documents and constitutional amendments. It is my view one of those values is equal opportunity for all, and it is my view we should not compromise our values for a perceived opportunity to better "compete" on an arbitrarily constrained playing field with other players we feel do not share those values. I think modern history has repeatedly proven that democratic and equitable organizations and countries prevail over those that are not and I don't see where in the lesson plan of America's history we would glean the insight that we should compromise our values now, in this case.
India and China are awful examples for your case, they aren't homogenous and have quotas for minorities that make affirmative action look like a joke. America is much more laissez faire about racial equality (which is a bad thing imo)
Saying that China or India are not "diverse" seems to assume that all Chinese people are actually members of the same group simply because they're all "Chinese people".
(China recognizes 55 ethnic minorities, and the Han majority itself speaks several mutually unintelligible languages although they kind of just ignore that.)
I run into this line of thinking quite often and it truly perplexes me.
Harvard’s incoming freshman class is a good example of the utter failure of mandating “reflective diversity” in higher education (or above). In nominal terms, 18% of these freshman are black. They are heavily, heavily incentivized and provided substantial affordances in the admissions process, even when compared to other minority groups. And the result is actually several percentage points ABOVE “reflective diversity”! Success!
Except going by historical trends compiled by “The Journal of Blacks in Higher Education” [1], the overwhelming majority - up to 75%! - of these incoming black freshmen are wealthy scions of the elite from Africa/the Caribbean or legacy admissions.
A top-down policy supposedly designed to promote opportunities for the descendants of slaves, has not only resulted in further disenfranchisement, but has served to coat the entire situation with a pernicious sense of false righteousness that has effectively boxed out the intended beneficiaries from engaging in the public discourse with their concerns. Which, in turn, is causing actual discriminatory beliefs to arise (check out “ADOS” or “FBA” hashtags on Twitter for a taste).
And fwiw, this trend is much broader and taller than just the Ivies. I can’t find the source again at the moment, but in one recent year, an outright majority of all publicized black hires in the C-suite weren’t American born blacks. Outright homogeneity (which doesn’t exist) is hardly less “reflective” of native-born Americans than the current situation.
The saddest aspect to all this is we’d get much more actual “reflective” diversity if we pulled all the distortionary diversity money out of the upper levels of the system and put it into massively improving the communities and EARLY educational outcomes of these underrepresented groups.
Unless you believe in the absolute primacy of genetics, community/home life and early education are the ONLY reliably predictive measures for success (after controlling for wealth). Our attention should be on those things, not on shady initiatives that disincentivize Americans of all stripes for the benefit of millionaire/billionaire global elites who are far more similar to their economic peers than they are to Americans with whom they share a skin color.
There’s a lot controversy because the “diversity” that actually occurs in the workplace is different from the diversity of organizational theory.
For example, a company that employs you graduates of elite schools might be racially diverse, but if many of the employees graduated from 10 colleges and all live in San Francisco, their world views could be quite similar.
“Diversity” can also be a political signal word. Progressives love diversity for reasons unrelated to management theory.
In fact “diversity” can used to enforce rigid ideological orthodoxies.
I’m sure that’s not what organizational scientists want, but it’s often what results.
Real evidence would be if it was something that got naturally prioritized and didn't need to be shoved down people's throats. But instead many would rather believe there's some massive conspiracy that companies are involved in, so that they need to be coerced into doing what's apparently unequivocally in their best interests.
In that case, what does it indicate that most successful and productive to date research centers for AI, like Google and OpenAI, have comprehensive diversity hiring policies that far exceed anything required by law?
And of course, the federal government is not mandating any behavior by private businesses - it is setting the charter for a publicly funded public resource initiative. The federal government often sets in writing directives for its agencies that the average Joe would view as common sense. For example, I think it's common sense that an effectively disseminated public resource initiative to expand America's AI workforce nationwide would result in building a diverse workforce.
Diversity of ideas is superior to diversity of skin color.
Any initiative that seeks to create a "diverse workforce" by skin color consideration (even if coupled with other considerations) is inherently flawed.
Goodhart's Law tells us these initiatives boil down to a checkbox or statistic - some number that must be met. Which naturally leads to underqualified or unqualified candidates being brought on because of skin color.
Diversity of ideas is what you want. People of all skin colors come from all places, economic statuses, educational backgrounds, etc.
I fully believe the current US government and most businesses can't create any kind of public mandate, regardless of the content, without throwing in a section about diversity. They don't fully understand AI but they do know how to cover their collective butts.
Once value is created by merit, the diversity and equity police come in demanding their share because their input will be so valuable. Hilarious and sad.
My read was that part of the value add of these new institutes is paying more people to do AI work, thus creating a larger AI workforce in the US and giving the US a competitive advantage.
It seems like a no-brainer to try and make that growing workforce come from diverse backgrounds because that will just help AI skills permeate throughout society even more, thus increasing the workforce advantage.
Considering the nature of AI systems, it seems this would create monopolies in various domains and AI devs will never exceed a single digit share of the IT workforce.
>try and make that growing workforce come from diverse backgrounds
If one's race is in question, that is racism no matter how you ice that (very rotten) cake. Same goes for sexism.
Give equal pay and treat everyone equally, do not care for and be blind to one's race and sex. That is equality. Trying to manipulate the distributions is discrimination.
As a Japanese (aka Asian, aka minority) American, I will say this: One party constantly calls me by a wide variety of labels, the other party treats me like just another man and American.
I know which party I'm voting for, and it doesn't start with the letter D.
Like culi, I read that as specific to the goal of expanding the "AI workforce" with new workers. I don't follow where you project the "creation of new roles far removed from core activities," from what I can see it has nothing to do with changing how individual businesses manage their hiring and staffing.
The remark about diversity comes near the end of the block, after all the discussion about inter-agency collaboration, and seems, to me, to specifically refer to the goal to "bolster America’s AI R&D infrastructure and support the development of a diverse AI workforce."
America is a very diverse country. A network of NSF institutes that is designed to expand America's AI R&D workforce will by definition create a diverse workforce if it effectively serves all Americans. I don't see why that's particularly objectionable at face.
I definitely cannot figure how to read that paragraph and project forward to this future state:
>the company's mission shifts from customer and safety focus, to pacifying employee agendas.
From what I can read, the Institutes aren't going to be directing companies to do anything.
> they have to justify their roles with initiatives and requirements that cause all manner of dysfunction in the organization
It's not necessarily the new hires causing the issues. If a company has a successful team that isn't diverse yet also needs to be seen as diverse, it will inevitably hire into non-core areas.
I'm skeptical of legislators legislating in domains they're not knowledgeable in, while being advised by the players with the upper-hand in the field financially.
This is really just government investment. Government investment gave us the iPhone basically.
I mean look at it:
GPS — government technology
the internet — a government technology
ai voice assistants — Apple literally hired the head of this research straight from DARPA after DARPA had released its work to the public
touchscreens — DARPA
accelerometers — DARPA
speech recognition — more DARPA and MIT tech
One of the only non-gov't technologies core to it is microprocessors which was made by Bell Labs. But even that was only possible because of a gov't mandated monopoly on Bell Labs and legislation that forced it to invest in research. Basically what public universities are
The government is innovative. It's just not good at taking credit for it
I mean just think about it for a second. Why on earth would a private company ever invest heavily in R&D and then release that knowledge to the public so everyone (including their competitors) can benefit. All our major technological innovations have come from the public sector and that should not at all be surprising if you just think about the incentive structures
Edit: if you're interested in this topic, economist Mariana Mazzucato has a whole book on debunking public vs. private sector myths. It's quite good and draws on a lot of data and evidence to make its case and show how absurd some of our assumptions really are: https://marianamazzucato.com/books/the-entrepreneurial-state
> I'm skeptical of legislators legislating in domains they're not knowledgeable in
The 2A folks would like to have a word with you...
In all seriousness - everyone assumes what they hear/read is 100% accurate and factual until it's about a topic they know something about. Then... we know the truth.
Why would some average legislator be any more qualified to regulate AI than Firearms? The answer is - they aren't.
> In all seriousness - everyone assumes what they hear/read is 100% accurate and factual until it's about a topic they know something about. Then... we know the truth.
And to bring it back around to AI, I've noticed a lot of people taking ChatGPT's responses as gospel but if you ask it about something you're knowledgeable in you can very easily catch it making stuff up and passing it off as fact.
This is scarier than AI. (The generic 'bias' and 'safety' can be interpreted wildly)
The US's enemies will not put limits on their AI, it will cause their software to be better.
I'm most afraid of 'safety' making GPT stop letting you use it for medical diagnosis. The medical cartels have competed regulatory capture and AI/LLMs are a genuine threat. My wife has already used GPT to diagnose a difficult patient that was passed around by multiple physicians(and specialists) over a 2 year span. (This was January GPT3)
The local models and powerful hardware cannot arrive soon enough. I'm afraid for our patients and my personal health, we should not under estimate the greed of the medical cartels, their lobbying efforts, and the fear over 'safety' they can draw from. In my lifetime the medical cartels always win. Its urgent to have the 'cat out of the bag'.
Don't forget the education cartel. GPT4 technology could essentially give everyone a 24/7 tutor in any subject they're studying at a tiny tiny fraction of the cost of a one-to-one human teacher. By almost every metric one-to-one learning overwhelmingly out-performs classroom learning (as the teacher has more time per student), but teachers unions are likely to heavily resist any move away from the traditional classroom model.
Schools have already started implementing blanket bans on ChatGPT, claiming that it prevents cheating. Receiving real-time feedback on an essay from a GPT certainly gives students an advantage, but I would hesitate to call it cheating. There is a growing uneasiness, particularly among individuals like Paul Graham, that GPTs diminish the importance of writing skills. However, I don't see it that way at all. In a future dominated by language models, written language is more relevant than ever.
When I was in school, the existence of YouTube and Google made many classes feel like a farce. You'd go to class, and the teacher would explain a concept so poorly that you would be forced to go home and watch a YouTube video about it. LLMs feel like an extension of this, further diminishing the relevance of in-class learning to actual education. Time will tell if this finally puts an end to the farce.
As they should, considering there are no guarantees of information quality coming out of these models. Anyone who thinks these things are an authoritative source on any kind of information is insane. Putting it in charge of teaching our children is just giving our kids an inferior service to save a buck.
> The US's enemies will not put limits on their AI, it will cause their software to be better.
You're goddamn right. And I'm going to tell something else, if the next ChatGPT-like AI tool is chinese or russian but turns out is not trying to make quit meat, I'll take that over Captain America-GPT any day of the year.
> The US's enemies will not put limits on their AI, it will cause their software to be better.
Absolutely ridiculous statement. We know that China is requiring every chatbot to be registered with the state. The idea that they're not going to put their own limits on that stuff has absolutely no basis in reality and it's not an excuse for lax regulatory frameworks of this technology.
Great, then we can ask US language models about Tiananmen Square, and Chinese language models about US crime statistics. Maybe diversity is a benefit after all?
The "get your hands off my AI" comments in this thread seem totally disconnected from what this announcement actually is. It's some funding for responsible AI research, a (voluntary) commitment from some leading AI companies to a public bake-off, and a plan to have some guidelines for using AI within the federal government. That's it.
That's an awful lot of minimizing you're doing here.
- Funding AI that adheres to the governments ideas of what is "ethical" and "responsible" despite massive public disagreement on how those terms manifest.
- Getting the largest tech firms in the world on board towards that goal.
- Explicitly imposing limits on what information models are allowed to share and train on if they want access to government capital and the capital of those aforementioned partners and all other businesses including capital markets that contract with the government.
The amount of arrogance in this field is fucking staggering. 50% of AI researchers think there's a 10% chance this tech makes us go extinct, and they rant and rave on the huge impacts this is going to have on all different aspects of our economy and society, but in the same breath they'll complain about government oversight.
yeah but funding comes with strings, and implies the ability for funders to pick the winners and losers of what is basically a new industrial sector based on who wants to play ball.
so where does the funding actually come from? sure, NSF, great! but according to https://new.nsf.gov/news/nsf-announces-seven-new-national-ar..., "U.S. Department of Homeland Security’s Science and Technology Directorate" is on the list. isn't it a realistic cause for concern that these people are going to be on steering committees?
Did you know that there are already 18 NSF AI institutes? This is adding 7 more. There is also of-course privately sponsored research going on at Google, Facebook, Open AI, etc., where most of the cutting-edge stuff is happening anwyway. This is not in exclusion to all other AI research. It's a new investment by the government into one particular area.
> This is not in exclusion to all other AI research
well, even setting aside DHS, IBM is also listed as a funding partner. are you saying that after IBM gives cash to the NSF, then the NSF is fair and unbiased about publishing work that might hurt IBM and help their rivals? IBM is no doubt on this list because they want to buy influence and future government contracts before they miss the boat completely, whereas google/fb/openai feel they don't need to.
i think we should follow the money, look for the conflicts of interest. it's not like you have to make research illegal to exclude players. you just gradually push them out of the loop, restrict their access to capital & research.
It's funny how language works. If this said "promote AI innovation" that would mean a particular thing. Adding the word "responsible" shouldn't make it mean roughly the opposite, and yet.
None of this matters. Zuck changed the game by dropping a cannonball on a rolling deck and pseudo-releasing Llama. You can try to rein in the monster by filling AI companies with ethicists and all that shit but everyone has a god in a bottle now and no one has a need to hold back.
104 comments
[ 3.2 ms ] story [ 203 ms ] threadAdditionally, those who comply will be heavily subsidized.
Welcome to the start of the American capture of the infosphere. Mark this date.
It’s ludicrous to watch our tech leaders release new AI models without ethical consideration or any form of reponsibility.
I work within a university and socially impactful research goes through rigorous ethics committees before being approved. Where was this with Open AI’s release of ChatGPT for example? Sam Altman just tweets after release words to the effect of, oh regulators should probably be considering the ethics around this. What a farce! Where are the ethics of those releasing these half baked tools?
Having the White House step in and provide leadership and bring it’s power to bear is totally appropriate and responsible in this moment.
Edit: spelling
By all means, if you believe (as it sounds like you do) that this will result in the affirmation and expansion of your worldview this is good news for you.
If you suspect, at any level, that maybe the current formulation of what is "ethical" and what is not may, at any point in the future, be called into question, than designing the fabric of super intelligent computational processes to presumptively adhere to your preconceptions is at best ill conceived at at worst malicious.
Enjoy the basilisk.
This is a gross exaggeration. Only a small portion of academic research requires IRB approval, and it's mostly limited to experiments on humans or animals.
If we were to start requiring IRB pre-approval for all research that might be "socially impactful", it would be the end of progress as we know it.
The first step would be to determine what counts as "socially impactful". You couldn't limit it to just AI. What about security and privacy? Why not all of CS? And then, why not Economics or History?
Source: worked for various universities doing research, first as graduate assistant, then scientist, then tenure-track professor, for 19 or 20 years.
Some things in CS are already regulated such as encryption via EAR: https://web.stanford.edu/group/export/encrypt_ear.html
Is your argument that they took appropriate precautions before releasing their paid tools publically?
My point isn’t that these AI research houses should have the same ethics committee as a university, it’s that they should have ethics committee appropriate to the risks around the work they are doing.
Additionally it sounds like the capital is net new. Which means if the government took no action, the would be no capital. Is it inappropriate for the government to offer programs to encourage things it wants?
At the other end of the spectrum, the government punishes companies that do things the administration doesn't like. In the middle, the government does nothing. It would seem to me that the alternative here is something everybody doesn't want, no?
We're standing here designing neutron bombs and arguing about who gets to decide what the targeting criteria is.
That should matter to absolutely everyone.
LLMs are already being used as mediums of information access and will continue to grow in that role.
Information filters/access has a huge role in informing public sentiment, that makes LLMs a massive propaganda vector to be employed for whichever purpose it's controller's wish.
Their effectiveness IN that role is also likely to be beyond anything we've ever seen before.
Hammers are better at hitting nails into wood than a human hand, but we're not all losing our minds at the fact that you can also smash a human skull in with a hammer. The laws apply to the human not the object, and that remains true here.
[0] https://youtu.be/bhYw-VlkXTU
This would, for a completely random example, restrict what Palantir's AIs could do.
That would then effect results delivered to clients of Palantir, including other nationstates.
But that's not an issue, right? No government would ever contract out it's ultra-high end scale-of-observation intelligence tasks... Right?
Like when the GOV procures munitions it's not saying, "please be sure they don't kill the enemy's innocent livestock".
If a person repeatedly says they are going to close a door, walks over to the door, grabs the handle, and begins to pull it closed... Is it really outside the realm of possibility that they may, in fact, close the door?
edit: I'm not sure what the criticism is here in light of the Notice and Explanation requirement -- there was N&E at most 4y after the practice began -- and it wasn't really a gaffe, presumably such a risk assessment system is still being used. I guess the argument is that this should have been required, or made earlier?
This is probably what we need more of at first, and then less of later -- once all these goofy flat "simple" decision models have been replaced.Here's a clear drawing of the ideal system without all the pesky words: https://ibb.co/7KrgBk4
What makes you think "they" want you to survive?
What policy or statement have they made that affirms their interest in your life or the lives of your interest groups? Do they want to/have they already disarmed you? Do you control a means of producing your own food? Your own energy? Are they allowing people into your areas to compete with you in labor markets and for control over limited resources? Are they meaningfully protecting the future of the planet for you and your children? Are they encouraging or discouraging you from having children at all?
This is a genuine question.
(But there's also no reason to disfavor it, because welfare is largely meant to benefit non-workers like elderly and children.)
If we extrapolate that relation, you eventually reach a point where the biggest player can collect and process the most information and produce an ever-evolving model to maintain that relation.
Better hope it's creators have your best interests at heart.
Source? My educated guess it’s somewhere between 10 to 100 times cheaper than that.
Actually:
https://www.wired.com/story/openai-ceo-sam-altman-the-age-of...
At the MIT event, Altman was asked if training GPT-4 cost $100 million; he replied, “It’s more than that.”
Granted, OP did say pre-training.
Interestingly, my own calculations lined up pretty well with this calculation, although they approached the problem from a different direction (a leak by Morgan Stanley about how many GPUs OpenAI used to train GPT-4 as well as an estimate of how long it was trained): https://colab.research.google.com/drive/1O99z9b1I5O66bT78r9S...
Sam Altman has also stated that GPT-4 cost more than $100 million to train, and replication can cost 2-4x less compute. https://www.wired.com/story/openai-ceo-sam-altman-the-age-of...
If you know of an organization that can replicate GPT-4 for $400k to $4m I would love to know so that I can invest in them.
1. We don't know what the number of parameters is, could be 175B, could be 250B, could be 400B. Ok, let's stick with 250B.
2. Training data: GPT-3 was trained on 300B tokens. It already used most of the high-quality data available on the internet, but let's say they somehow managed to find and prepare three times as much high quality data for GPT-4. This means GPT-4 was trained on about 1T tokens.
3. 5.4e+17 FLOPs/hour means 150TFlops, which is half of the BFLOAT16 max theoretical output, sounds reasonable.
4. $1/A100/hr is reasonable.
OK, so we need to divide your cost estimate by a factor of 15: Total cost to train GPT-4 comes out to be around $2.7M.
Regarding Altman's statement about "more than 100M to train GPT-4" - I'm pretty sure he was talking about the total cost to develop GPT-4, which includes a lot of experimentation and exploration, many training runs, and many other administrative costs which are not relevant to the cost of a single training run to reproduce the existing results. Just salaries alone: ~200 people worked on GPT-4 for let's say half a year, at $400k/year: 0.5 * 400k * 200 = $40M.
Do you have actual quantified evidence that supports this? Also if you are going to cite a study, it should be reproducible.
And like the other commenter noted, the Institutes component seems focused on fostering new AI talent all over the country, so ergo a part of that will be making its resources accessible to many different groups of people. A claim that this initiative would be more effective in its stated goals by confining its educational resources to a single population demographic seems like it would also demand research and concrete evidence. Would that even be legal?
"Diverse" is one keyword in a whole paragraph about introducing an NSF funded program to provide educational resources for new workers to skill up on AI. America is a diverse country: Of course a program that serves all of America with that stated goal would create a diverse workforce. I'm a little confused why the top discussion on this thread has seized on that single word with such outrage.
[0]https://hbr.org/2013/12/how-diversity-can-drive-innovation
If you were taught this a decade ago you likely are familiar with the "Striving, Storming, Thriving" study which demonstrates that relation especially well.
I recall a recent hand-wringing study out of Harvard jam-packed with soothing language but showing that as teams diversify their ability to engage in shared risk-taking declines [0]
[0] https://www.hbs.edu/faculty/Pages/item.aspx?num=61993
On the other hand, I still don't share the anguish over the term 'diverse' in this particular OP for the reasons I stated - if this program is designed to foster AI talent all over America, if successful it will by definition foster diverse talent.
Like if the program wasn't designed to help America build a diverse AI workforce, then it would be designed to create a demographically constrained, homogenized workforce. I wouldn't expect us to be happy about that and I wouldn't think of it as a good allocation of federal tax money.
Something not being designed to do X does not mean it is designed to do Y.
The hackles raised here over "diversity" are, as you note, likely related to demographic considerations as opposed to say geographic considerations.
The argument is usually that the goal of government programs should be to accomplish a task as effectively and efficiently as possible. In the case of AI doing so is especially important in an era of rising geopolitical multipolarity. Sacrificing this for the sake of appeasing identity interest groups is short-sighted.
As we briefly explored, arguments that demographic diversity augment this efficacy have yet to meaningfully materialize, while concrete evidence to the contrary has both in the literature and in large-scale behavioral patterns (the neighborhood trust studies come to mind). It's one of, if not THE, most contentious area of sociology and public policy right now because of the obvious implications.
IMO that's why people latch onto it and why it's getting so much attention here. It's a hot button issue of grave importance that hasn't been solved yet and discussions around it are extremely painful and frequently stifled.
In the context of this specific initiative, my view would still be: America is an inherently diverse place, so a program that seeks to develop an AI workforce across America but does not create a diverse one has not achieved “maximum effectiveness” and has not effectively served all of America.
We may be diverse, but that needn't mean every sector of every niche in every industry must be so or that it would be better if it were.
Our position in the world is relative. Our power comes from the relative delta between us and "them", whomever it is we happen to be competing with.
To forget that fundamental relation, even in service to our personal sense of morality, is solipsism.
>if it results in China or India developing more effective models faster and more cheaply
Why should we project that possibility? The most successful research centers for new AI development have been American organizations like OpenAI and Google with clearly stated diversity policies. And of course, this OP is not about private companies, it's about a public resource program. Do you have any evidence that China and India are developing more effective models faster and more cheaply and that it's because of some specific lack of attention paid to ensuring public resources are broadly accessible to all demographics?
>We may be diverse, but that needn't mean every sector of every niche in every industry must be so or that it would be better if it were.
Rather than address this whole point, I'll revisit that we are talking about a federally funded public education program. I do not think a federally funded program to develop a bigger AI workforce in America would be more successful if it did not serve all Americans, and especially not if some of the funding were diverted to gatekeeping its resources to specific demographics.
This is strictly personal opinion, but mine is: America is not China or India and that extends to the values enshrined in our founding documents and constitutional amendments. It is my view one of those values is equal opportunity for all, and it is my view we should not compromise our values for a perceived opportunity to better "compete" on an arbitrarily constrained playing field with other players we feel do not share those values. I think modern history has repeatedly proven that democratic and equitable organizations and countries prevail over those that are not and I don't see where in the lesson plan of America's history we would glean the insight that we should compromise our values now, in this case.
(China recognizes 55 ethnic minorities, and the Han majority itself speaks several mutually unintelligible languages although they kind of just ignore that.)
Harvard’s incoming freshman class is a good example of the utter failure of mandating “reflective diversity” in higher education (or above). In nominal terms, 18% of these freshman are black. They are heavily, heavily incentivized and provided substantial affordances in the admissions process, even when compared to other minority groups. And the result is actually several percentage points ABOVE “reflective diversity”! Success!
Except going by historical trends compiled by “The Journal of Blacks in Higher Education” [1], the overwhelming majority - up to 75%! - of these incoming black freshmen are wealthy scions of the elite from Africa/the Caribbean or legacy admissions.
A top-down policy supposedly designed to promote opportunities for the descendants of slaves, has not only resulted in further disenfranchisement, but has served to coat the entire situation with a pernicious sense of false righteousness that has effectively boxed out the intended beneficiaries from engaging in the public discourse with their concerns. Which, in turn, is causing actual discriminatory beliefs to arise (check out “ADOS” or “FBA” hashtags on Twitter for a taste).
And fwiw, this trend is much broader and taller than just the Ivies. I can’t find the source again at the moment, but in one recent year, an outright majority of all publicized black hires in the C-suite weren’t American born blacks. Outright homogeneity (which doesn’t exist) is hardly less “reflective” of native-born Americans than the current situation.
The saddest aspect to all this is we’d get much more actual “reflective” diversity if we pulled all the distortionary diversity money out of the upper levels of the system and put it into massively improving the communities and EARLY educational outcomes of these underrepresented groups.
Unless you believe in the absolute primacy of genetics, community/home life and early education are the ONLY reliably predictive measures for success (after controlling for wealth). Our attention should be on those things, not on shady initiatives that disincentivize Americans of all stripes for the benefit of millionaire/billionaire global elites who are far more similar to their economic peers than they are to Americans with whom they share a skin color.
[1] https://www.jbhe.com/news_views/52_harvard-blackstudents.htm...
For example, a company that employs you graduates of elite schools might be racially diverse, but if many of the employees graduated from 10 colleges and all live in San Francisco, their world views could be quite similar.
“Diversity” can also be a political signal word. Progressives love diversity for reasons unrelated to management theory.
In fact “diversity” can used to enforce rigid ideological orthodoxies.
I’m sure that’s not what organizational scientists want, but it’s often what results.
And of course, the federal government is not mandating any behavior by private businesses - it is setting the charter for a publicly funded public resource initiative. The federal government often sets in writing directives for its agencies that the average Joe would view as common sense. For example, I think it's common sense that an effectively disseminated public resource initiative to expand America's AI workforce nationwide would result in building a diverse workforce.
Diversity of ideas is superior to diversity of skin color.
Any initiative that seeks to create a "diverse workforce" by skin color consideration (even if coupled with other considerations) is inherently flawed.
Goodhart's Law tells us these initiatives boil down to a checkbox or statistic - some number that must be met. Which naturally leads to underqualified or unqualified candidates being brought on because of skin color.
Diversity of ideas is what you want. People of all skin colors come from all places, economic statuses, educational backgrounds, etc.
This is about projecting leadership and relevancy to centrists they need to win in 2024. “Look, we saved your jobs!”
It seems like a no-brainer to try and make that growing workforce come from diverse backgrounds because that will just help AI skills permeate throughout society even more, thus increasing the workforce advantage.
If one's race is in question, that is racism no matter how you ice that (very rotten) cake. Same goes for sexism.
Give equal pay and treat everyone equally, do not care for and be blind to one's race and sex. That is equality. Trying to manipulate the distributions is discrimination.
I know which party I'm voting for, and it doesn't start with the letter D.
The remark about diversity comes near the end of the block, after all the discussion about inter-agency collaboration, and seems, to me, to specifically refer to the goal to "bolster America’s AI R&D infrastructure and support the development of a diverse AI workforce."
America is a very diverse country. A network of NSF institutes that is designed to expand America's AI R&D workforce will by definition create a diverse workforce if it effectively serves all Americans. I don't see why that's particularly objectionable at face.
I definitely cannot figure how to read that paragraph and project forward to this future state:
>the company's mission shifts from customer and safety focus, to pacifying employee agendas.
From what I can read, the Institutes aren't going to be directing companies to do anything.
It's not necessarily the new hires causing the issues. If a company has a successful team that isn't diverse yet also needs to be seen as diverse, it will inevitably hire into non-core areas.
I mean look at it:
One of the only non-gov't technologies core to it is microprocessors which was made by Bell Labs. But even that was only possible because of a gov't mandated monopoly on Bell Labs and legislation that forced it to invest in research. Basically what public universities areThe government is innovative. It's just not good at taking credit for it
I mean just think about it for a second. Why on earth would a private company ever invest heavily in R&D and then release that knowledge to the public so everyone (including their competitors) can benefit. All our major technological innovations have come from the public sector and that should not at all be surprising if you just think about the incentive structures
Edit: if you're interested in this topic, economist Mariana Mazzucato has a whole book on debunking public vs. private sector myths. It's quite good and draws on a lot of data and evidence to make its case and show how absurd some of our assumptions really are: https://marianamazzucato.com/books/the-entrepreneurial-state
The 2A folks would like to have a word with you...
In all seriousness - everyone assumes what they hear/read is 100% accurate and factual until it's about a topic they know something about. Then... we know the truth.
Why would some average legislator be any more qualified to regulate AI than Firearms? The answer is - they aren't.
And to bring it back around to AI, I've noticed a lot of people taking ChatGPT's responses as gospel but if you ask it about something you're knowledgeable in you can very easily catch it making stuff up and passing it off as fact.
The US's enemies will not put limits on their AI, it will cause their software to be better.
I'm most afraid of 'safety' making GPT stop letting you use it for medical diagnosis. The medical cartels have competed regulatory capture and AI/LLMs are a genuine threat. My wife has already used GPT to diagnose a difficult patient that was passed around by multiple physicians(and specialists) over a 2 year span. (This was January GPT3)
The local models and powerful hardware cannot arrive soon enough. I'm afraid for our patients and my personal health, we should not under estimate the greed of the medical cartels, their lobbying efforts, and the fear over 'safety' they can draw from. In my lifetime the medical cartels always win. Its urgent to have the 'cat out of the bag'.
When I was in school, the existence of YouTube and Google made many classes feel like a farce. You'd go to class, and the teacher would explain a concept so poorly that you would be forced to go home and watch a YouTube video about it. LLMs feel like an extension of this, further diminishing the relevance of in-class learning to actual education. Time will tell if this finally puts an end to the farce.
You're goddamn right. And I'm going to tell something else, if the next ChatGPT-like AI tool is chinese or russian but turns out is not trying to make quit meat, I'll take that over Captain America-GPT any day of the year.
Absolutely ridiculous statement. We know that China is requiring every chatbot to be registered with the state. The idea that they're not going to put their own limits on that stuff has absolutely no basis in reality and it's not an excuse for lax regulatory frameworks of this technology.
- Funding AI that adheres to the governments ideas of what is "ethical" and "responsible" despite massive public disagreement on how those terms manifest.
- Getting the largest tech firms in the world on board towards that goal.
- Explicitly imposing limits on what information models are allowed to share and train on if they want access to government capital and the capital of those aforementioned partners and all other businesses including capital markets that contract with the government.
so where does the funding actually come from? sure, NSF, great! but according to https://new.nsf.gov/news/nsf-announces-seven-new-national-ar..., "U.S. Department of Homeland Security’s Science and Technology Directorate" is on the list. isn't it a realistic cause for concern that these people are going to be on steering committees?
well, even setting aside DHS, IBM is also listed as a funding partner. are you saying that after IBM gives cash to the NSF, then the NSF is fair and unbiased about publishing work that might hurt IBM and help their rivals? IBM is no doubt on this list because they want to buy influence and future government contracts before they miss the boat completely, whereas google/fb/openai feel they don't need to.
i think we should follow the money, look for the conflicts of interest. it's not like you have to make research illegal to exclude players. you just gradually push them out of the loop, restrict their access to capital & research.
Police wants to lock you up and your boss wants to fire you using this kind of products:
https://www.aclu.org/news/privacy-technology/amazons-face-re...
And surprise surprise, it works even worse for women and minorities
https://proceedings.mlr.press/v81/buolamwini18a/buolamwini18...
LOL