And in cases where your knowledge sits far outside the bubble of consensus that AI tools draw from, it will likely look a lot more like co-writing.

Because authority is ultimately something earned from the community you are involved in.

The “average” outputs that AI generates can’t capture your unique life experiences.

This gravity toward the average will lead to…

that 20% of your company’s collective knowledge will drive 80% of your results.

have a keyword difficulty of 10 or less.

Utilizing cutting-edge fine-tuning methodologies

And let’s call it round one of the internet has not been great for that

very new technology, we need a new framework

combination of companies doing the right thing, regulation and public education

if content that they’re looking at might be generated or might not

we think that regulatory intervention by governments will be critical to mitigate the risks of increasingly powerful models

work together

The AI industry doesn’t have to wait for Congress

a governance regime flexible enough to adapt to new technical development

To ensure this, the U.S. government should consider a combination of licensing or registration requirements for development and release of AI models above a crucial threshold of capabilities, alongside incentives for full compliance with these requirements

we work to remove

to follow user instructions

OpenAI has an unusual structure

our api

also think it’s really important to decide to whose values we’re going to align these models.

I’m a big believer in the democratizing potential of technology,

is going to be small just because of the resources required.

I think it’s important to democratize the inputs to these systems, the values that we’re going to Alli align to.

with Microsoft releasing Sydney.

there are areas like copyright where we don’t really have laws.

to prioritize ethics and responsible technology as opposed to posing development.

My opinion is that the moratorium that we should focus on is actually deployment until we have good safety cases.

try to make these things actually enforced.

three principles, transparency, accountability, and limits on use.

These systems are almost like counterfeit people, and we don’t really honestly understand what the consequence of that is.

counterfeit people

Same with psychiatric advice

medical misinformation

haven’t been invented yet

pre-deployment and post-deployment.

we put so much burden that only the big players can do it.

regulatory capture.

you can still cause great harm with a smaller model.

you slow down American industry in such a way that China or somebody else makes faster progress.

economic transformation

safety concerns

Number one, you’re here because AI is this extraordinary new technology that everyone says can be transformative as much as the printing press. Number two is really unknown what’s gonna happen. But there’s a big fear you’ve expressed to all of you about what bad actors can do and will do if there’s no rules of the road. Number three, as a member who served in the house and now in the Senate, I’ve come to the conclusion that it’s impossible for Congress to keep up with the speed of technology.

what these systems get aligned to, whose values,

There is a real risk of a kind of technocracy combined with oligarchy, where small number of companies influence people’s beliefs through the nature of these systems.

this massive corporate concentration.

We have API developers pay us and we have ChatGPT.

we need to have some international meetings very quickly with people who have expertise in how you grow agencies in the history of growing agencies.

no way to put this genie in the bottle globally.

But I I, you know, you talked about defining the highest risk uses. We don’t know all of them. We really don’t. We can’t see where this is going regulating at the point of risk.

we’re not an advertising based model.

the safety for children of you

tools that humans use to make human judgments, and that we don’t take away human judgment

can predict future human behavior is potentially pretty significant at the individual level.

a model that could help create novel biological agents would be a great threshold.

capability thresholds

Andwe risk if we’re not thoughtful about it contributing to the development of tools and approaches that only exacerbate the bias and inequities that exist in our society.

 Excited to work with people who have particular data sets and to work to collect a representative set of values from around the world to draw these wide bounds of what the system can do.

Can can you speak just for a second specifically to language inclusivity?

language and cultural inclusivity

And what auto GPT does is it allows systems to access source code, access the internet and so forth. And there are a lot of potential, let’s say cybersecurity risks. There, there should be an external agency that says, well, we need to be reassured if you’re going to release this product that there aren’t gonna be cybersecurity problems or there are ways of addressing it.

the central scientific issue

I can’t envision or imagine right now what kind of a licensing scheme we would be able to create to pretty much regulate the vastness of, of the, this game-changing tool.

some sort of standard,

what do you consider a harmful request?

require independent audits.

specific tests that a model has to pass

a set of safety standards

a new agency that licenses any effort above a certain scale of capabilities and can take that license away and ensure compliance with safety standards.

basically focus on AI safety research.

ai, constitution

a nimble monitoring agency to follow what’s going on.

please don’t just use concepts. I’m looking for specificity.

So disclosure of the data that’s used to train AI, disclosure of the model and how it performs and making sure that there’s continuous governance over these models.

the I P C A UN body

So we have existing regulatory authorities in place who have been clear that they have the ability to regulate in their respective domains. A lot of the issues we’re talking about today span multiple domains, elections, and the like.

Guardrails need to be in place.

Different rules for different risks.

the conception of the EU AI Act is very consistent with this concept of precision regulation where you’re regulating the use of the technology in context.

we need to give policy makers and the world as a whole the tools to say, here’s the values and implement them.

that interaction with the world is very important.

constitutional AI

a reasonable care standard.

Thank you, Mr. Chairman and Senator Hawley for having this. I’m trying to find out how it is different than social media and learn from the mistakes we made with social media. The idea of not suing social media companies is to allow the internet to flourish. Because if I slander you you can sue me. If you’re a billboard company and you put up the slander, can you sue the billboard company? We said no. Basically, section 230 is being used by social media companies to high, to avoid liability for activity that other people generate. When they refuse to comply with their terms of use, a mother calls up the company and says, this app is being used to bully my child to death. You promise, in the terms of use, she would prevent bullying. And she calls three times, she gets no response, the child kills herself and they can’t sue. Do you all agree we don’t wanna do that again?

And so the quality of the sort of overall news market is going to decline as we have more generated content by systems that aren’t actually reliable in the content they’re generated.

the current version of GPT-4 ended to training in 2021.

And other countries are doing this, Australia and the like. And so my question is, when we already have a study by Northwestern predicting that one-third of the US newspapers are that roughly existed, two decades are gonna go, are gonna be gone by 2025, unless you start compensating for everything from book movies, books. Yes. but also news content. We’re gonna lose any realistic content producers. And so I’d like your response to that. And of course, there is an exemption for copyright in section two 30. But I think asking little newspapers to go out and sue all the time just can’t be the answer. They’re not gonna be able to keep up.

When some of those fake ads. So that’s number one. Number two is the impact on intellectual property.

Sen. Marsha Blackburn (R-TN):

OpenAI Jukebox

And I wanna come to you on music and content creation, because we’ve got a lot of songwriters and artists. Yeah. And a, I think we have the best creative community on the face of the Earth. They’re in Tennessee, and they should be able to decide if their copyrighted songs and images are going to be used to train these models. And I’m concerned about OpenAI’s jukebox. It offers some renditions in the style of Garth Brooks, which suggests that OpenAI is trained on Garth Brooks songs. I went in this weekend and I said, write me a song that sounds like Garth Brooks. And it gave me a different version of Simple Man. So it’s interesting that it would do that. But you’re training it on these copyrighted songs, these mini files, these sound technologies. So as you do this, who owns the rights to that AI generated material and using your technology, could I remake a song, insert content from my favorite artist, and then own the creative right to that song?

that people own their virtual you.

We’ve done it before with the IAEA.

When you think about the energy costs alone, just for training these systems, it would not be a good model if every country has its own policies and each, for each jurisdiction, every company has to train another model.

need a cabinet level organization within the United States in order to address this.

So we think that AI should be regulated at the point of risk, essentially, and that’s the point at which technology meets society.

I can’t recall when we’ve had people representing large corporations or private sector entities come before us and plead with us to regulate them. In fact, many people in the Senate have base their careers on the opposite that the economy will thrive if government gets the hell out of the way. And what I’m hearing instead today is that ‘stop me before I innovate again’ message. And I’m just curious as to how we’re going to achieve this

an enterprise technology company, not consumer focused. S

hyper targeting of advertising is definitely going to come.

And we probably need scientists in there doing analysis in order to understand what the political influences of, for example, of these systems might be.

we don’t know what it’s trained on.

If these large language models can, even now, based on the information we put into them quite accurately predict public opinion, you know, ahead of time. I mean, predict, it’s before you even ask the public these questions, what will happen when entities, whether it’s corporate entities or whether it’s governmental entities, or whether it’s campaigns or whether it’s foreign actors, take this survey information, these predictions about public opinion and then fine tune strategies to elicit certain responses, certain behavioral responses.

large language models can indeed predict public opinion.

so-called artificial general intelligence really will replace a large fraction of human jobs.

And the real question is over what time scale? Is it gonna be 10 years? Is it gonna be a hundred years?

that scientists be part of that process.

So I think the most important thing that we could be doing and can, and should be doing now, is to prepare the workforce of today and the workforce of tomorrow for partnering with AI technologies and using them. And we’ve been very involved for, for years now in doing that in focusing on skills-based hiring in educating for the skills of the future. Our skills build platform has 7 million learners and over a thousand courses worldwide focused on skills. And we’ve pledged to train 30 million individuals by 2030 in the skills that are needed for society today.

And so you see already people that are using GPT-4 to do their job much more efficiently by helping them with tasks. Now, GPT-4 will I think entirely automate away some jobs, and it will create new ones that we believe will be much better. This happens again, my understanding of the history of technology is one long technological revolution, not a bunch of different ones put together, but this has been continually happening. We, as our quality of life raises and as machines and tools that we create can help us live better lives the bar raises for what we do and, and our human ability and what we spend our time going after goes after more ambitious, more satisfying projects. So there will be an impact on jobs. We try to be very clear about that, and I think it will require partnership between the industry and government, but mostly action by government to figure out how we want to mitigate that. But I’m very optimistic about how great the jobs of the future will be.

not a creature,

this technology is in its early stages

I think that’s a great idea.

The National Institutes of Standards and technology actually already has an AI accuracy test,

Some of us might characterize it more like a bomb in a China shop, not a bull.

Ultimately, we may need something like cern Global, international and neutral, but focused on AI safety rather than high energy physics.

But we also need independent scientists, not just so that we scientists can have a voice, but so that we can participate directly in addressing the problems in evaluating solutions.

The sums of money at stake are mind boggling. Emissions drift, OpenAI’s original mission statement proclaimed our goal is to advance AI in the way that most is most likely to benefit humanity as a whole, unconstrained by a need to generate financial return. Seven years later, they’re largely beholden to Microsoft, embroiled in part an epic battle of search engines that routinely make things up.

We want, for example, for our systems to be transparent, to protect our privacy, to be free of bias and above all else to be safe. But current systems are not in line with these values. Current systems are not transparent. They do not adequately protect our privacy, and they continue to perpetuate bias, and even their makers don’t entirely understand how they work. Most of all, we cannot remotely guarantee that they’re safe. And hope here is not enough. The big tech company’s preferred plan boils down to trust us. But why should we?

We all more or less agrees on the values we would like for our AI systems to honor.

A law professor, for example, was accused by a chatbot of sexual harassment untrue. And it pointed to a Washington Post article that didn’t even exist. The more that that happens, the more that anybody can deny anything. As one prominent lawyer told me on Friday, defendants are starting to claim that plaintiffs are making up legitimate evidence. These sorts of allegations undermine the abilities of juries to decide what or who to believe and contribute to the undermining of democracy. Poor medical advice could have serious consequences to an open source large language model recently seems to have played a role in a person’s decision to take their own life. The large language model asked the human, if you wanted to die, why didn’t you do it earlier? And then followed up with, were you thinking of me? When you overdosed without ever referring the patient to the human health?

What criminals are gonna do here is to create counterfeit people.

Choices about data sets that AI companies use will have enormous unseen influence. Those who choose the data will make the rules shaping society in subtle but powerful ways.

Fundamentally, these new systems are going to be destabilizing. They can and will create persuasive lies at a scale humanity has never seen before. Outsiders will use them to affect our elections, insiders to manipulate our markets and our political systems. Democracy itself is threatened. Chatbots will also clandestinely shape our opinions, potentially exceeding what social media can do.

guardrails

First, different rules for different risks. The strongest regulation should be applied to use cases with the greatest risks to people and society. Second, clearly defining risks. There must be clear guidance on AI uses or categories of AI supported activity that are inherently high risk. This common definition is key to enabling a clear understanding of what regulatory requirements will apply in different use cases and contexts. Third, be transparent. So AI shouldn’t be hidden. Consumers should know when they’re interacting with an AI system and that they have recourse to engage with a real person should they so desire. No person anywhere should be tricked into interacting with an AI system. And finally, showing the impact. For higher risk use cases, companies should be required to conduct impact assessments that show how their systems perform against tests for bias and other ways that they could potentially impact the public. And to attest that they’ve done so by following risk-based use case-specific approach.

a precision regulation

a threshold of capabilities

We believe that the benefits of the tools we have deployed so far vastly outweigh the risks,

be My Eyes, used our new multimodal technology in GPT-4 to help visually impaired individuals navigate their environment.

We think it can be a printing press moment.

But the basic question we face is whether or not this issue of AI is a quantitative change in technology or a qualitative change.

We had four bills initially that were considered by this committee and what may be history in the making. We passed all four bills with unanimous roll calls, unanimous roll calls. I can’t remember another time when we’ve done that.

will we strike that balance between technological innovation and our ethical and moral responsibility to humanity, to liberty, to the freedom of this country?

I was reminded of the psychologist and writer Carl Jung, who said at the beginning of the last century that our ability for technological innovation, our capacity for technological revolution, had far outpaced our ethical and moral ability to apply and harness the technology we developed.

is it gonna be more like the atom bomb, huge technological breakthrough, but the consequences severe, terrible, continue to haunt us to this day

Is it gonna be like the printing press that diffused knowledge, power, and learning widely across the landscape that empowered, ordinary, everyday individuals that led to greater flourishing, that led above all two greater liberty?

We should not repeat our past mistakes, for example, Section 230

scorecards and nutrition labels

known risks

transparency

Now we have the obligation to do it on AI before the threats and the risks become real. Sensible safeguards are not in opposition to innovation.

we may need something like CERN, global, international, and neutral, but focused  on AI safety, rather than high-energy physics.  

Chatbots can clandestinely  shape our opinions, in subtle yet potent ways, potentially exceeding what social media can do.  Choices about datasets may have enormous, unseen influence. 

I call datocracy, the opposite of democracy:

These  guardrails should be matched with meaningful steps by the business community  to do their part.

an AI Ethics Board

a lead AI ethics official

on regulatory guardrails

A risk based approach ensures that guardrails for AI apply to any application, even as  this new, potentially unforeseen developments in the technology occur, and that  those responsible for causing harm are held to account.

a “precision regulation” approach to  artificial intelligence. This means establishing rules to govern the deployment of AI  in specific use-cases, not regulating the technology itself. 

international cooperation on AI safety, including examining potential intergovernmental oversight mechanisms and standard-setting. 

safety standards, evaluation requirements, disclosure practices, and external validation mechanisms for AI systems subject to license or registration. 

a combination of licensing or registration requirements

adhere to an appropriate set of safety requirements,

regulation of AI is essential,

to minimize any harmful effects for workers and businesses.

the Alignment Research Center (ARC)

a Cybersecurity Working Group

a Cybersecurity Grant Program

novel security controls to help protect core model intellectual property.

We will continue to explore partnerships with industry and researchers, as well as with governments, that encompass the full disinformation lifecycle. 

takes a whole-of-society approach

we recognize that there is more work to do to educate users about the limitations of AI tools, and to reduce the likelihood of inaccuracy.

Thorn’s Safer37 service

to generate hateful, harassing, violent or adult content, among other categories,

Any ChatGPT user can opt-out of having their conversations be used to improve our models.33 Users can delete their accounts,34 delete specific conversations from the history sidebar, and disable their chat history at any time.35 

via our API

for the purposes of advertising, promoting our services, or selling data to third parties

Iterative deployment

people and our institutions need time to update and adjust to increasingly capable AI, and that everyone who is affected by this technology should have a significant say in how AI develops further.

unsafe content

including but not limited to, generation of violent content, malware, fraudulent activity, high-volume political campaigning, and many other unwelcome areas.

disallowed content

These opportunities are why former U.S. Treasury Secretary Lawrence Summers has said that AI tools such as ChatGPT might be as impactful as the printing press, electricity, or even the wheel or fire.

Microsoft is an important investor in OpenAI, and we value their unique alignment with our values and long-term vision, including their shared commitment to building AI systems and products that are trustworthy and safe.

“What I keep emphasizing to people is to just start using this,” says Mollick. As workers get increasing fluent, he adds, they can find themselves ahead of the curve, and at a distinct advantage in the workplace. Workers resistant to AI could be seen as unwilling or incapable of adapting, says Frey. “I think workers that don't work with AI are going to find their skills [become] obsolete quite rapidly. So, therefore, it's imperative to work with AI to stay employed, stay productive and have up to date skills.”

increasing the demand for jobs including data analysts and scientists who work with the technology to create best practices in the workplace.

might be able to identify a confirmation bias in their work, meaning they look for evidence to support an outcome they already believe exists

functions as a sounding board – a tool to bounce ideas off,

In his own field of academia, for instance, he’s seen it test for counterarguments to a thesis, and write an abstract for research. “You can ask it to generate a tweet to promote your paper,” he adds. “There are tremendous possibilities.” For knowledge workers, this could mean creating an outline for a blog and a social media post to go with it, distil complex topics for a target audience

Stability AI announced StableLM - their Language Models.

Open Assistant - just wow - is an open source Chat AI.

How To Evaluate a Chatbot?

To ensure data quality, we convert the HTML back to markdown and filter out some inappropriate or low-quality samples.

Vicuna is created by fine-tuning a LLaMA base model using approximately 70K user-shared conversations gathered from ShareGPT.com with public APIs.

Inspired by the Meta LLaMA and Stanford Alpaca project, we introduce Vicuna-13B, an open-source chatbot backed by an enhanced dataset and an easy-to-use, scalable infrastructure.

Cerebras (not to be confused with our own Cerebra) trains the GPT-3 architecture using the optimal compute schedule implied by Chinchilla, and the optimal scaling implied by μ-parameterization. This outperforms existing GPT-3 clones by a wide margin, and represents the first confirmed use of μ-parameterization “in the wild”. These models are trained from scratch, meaning the community is no longer dependent on LLaMA.

LoRa and ControlNet (not to mention in- and outpainting), there’s a clear benefit to letting people go wild with your tech.

allowing the model to better understand and use context.

We explore two other obvious sources of basis dependency in a Transformer: Layer normalization, and finite-precision floating-point calculations. We confidently rule these out as being the source of the observed basis-alignment.

Alpaca AI is open source and around the same performance as gpt3. https://github.com/tatsu-lab/stanford_alpaca

Look into Modular, they have an interesting platform for AI development.

GPT 4 cost well over $100 million to train alone, $700k to run per day.

we focus on collecting a small high-quality dataset.

built using significant amounts of human annotation.

Our results suggest that learning from high-quality datasets can mitigate some of the shortcomings of smaller models, maybe even matching the capabilities of large closed-source models in the future.

will the future see increasingly more consolidation around a handful of closed-source models, or the growth of open models with smaller architectures that approach the performance of their larger but closed-source cousins?

This suggests that in the future, highly capable LLMs will be largely controlled by a small number of organizations, and both users and researchers will pay to interact with these models without direct access to modify and improve them on their own.

the community should put more effort into curating high-quality datasets

it suggests that models that are small enough to be run locally can capture much of the performance of their larger cousins if trained on carefully sourced data

Google "We Have No Moat, And Neither Does OpenAI"