2 Senators Propose Bipartisan Framework for A.I. Laws

Inside effective altruism, where the far future counts a lot more than the present

The Reluctant Prophet of Effective Altruism

This task demonstrates the failure of language models to follow instructions when there is a popular continuation that does not fit with that instruction. Larger models are more hurt by this as the larger the model, the more familiar it is with common expressions and quotes.

The standard paradigm in natural language processing today is to pretrain large language models to autocomplete text corpora. The resulting models are then either frozen and used directly for other tasks (zero-shot or using few-shot learning), or additionally trained on other tasks (fine-tuning).

The purpose of this contest is to find evidence for a stronger failure mode: tasks where language models get worse as they become better at language modeling (next word prediction).

, Sakana’s approach could potentially lead to AI that’s cheaper to train and use than existing technology. That includes generative AI

Sakana is still in the early stages: It hasn’t yet built an AI model and doesn’t have an office.

The startup plans to make multiple smaller AI models, the kind of technology that powers products like ChatGPT, and have them work together. The idea is that a “swarm” of programs could be just as smart as the massive undertakings from larger organizations.

Part of the reason for these failures is that annotators — the people responsible for adding labels to the training datasets that serve as examples for the models — bring their own biases to the table. For example, frequently, there’s differences in the annotations between labelers who self-identified as African Americans and members of the LGBTQ+ community versus annotators who don’t identify as either of those two groups.

Perspective often couldn’t recognize hate speech that used “reclaimed” slurs like “queer” and spelling variations such as missing characters.

Several years ago, a team at Penn State found that posts on social media about people with disabilities could be flagged as more negative or toxic by commonly used public sentiment and toxicity detection models. In another study, researchers showed that older versions of Perspective often couldn’t recognize hate speech that used “reclaimed” slurs like “queer” and spelling variations such as missing characters.

OpenAI proposes a new way to use GPT-4 for content moderation

"We're enabling ownership of these technologies by African founders," he told DW. "This is being built and served by people from these communities. So the financial rewards will also go directly back to them."

Consistent with the findings of the 2021 edition, software tools and reports are the most common outputs of UN AI projects, which can be used to address challenges impeding progress on the SDGs

Debunking Common Misconceptions about Prompt Engineering

they aremost common when asking for quotes, sources, citations, or other detailed information

This paper examines the transformative role of Large Language Models (LLMs) in education and their potential as learning tools, despite their inherent risks and limitations. The authors propose seven approaches for utilizing AI in classrooms: AI-tutor, AI-coach, AI-mentor, AI-teammate, AI-tool, AI-simulator, and AI-student, each with distinct pedagogical benefits and risks. The aim is to help students learn with and about AI, with practical strategies designed to mitigate risks such as complacency about the AI’s output, errors, and biases. These strategies promote active oversight, critical assessment of AI outputs, and complementation of AI's capabilities with the students' unique insights. By challenging students to remain the "human in the loop", the authors aim to enhance learning outcomes while ensuring that AI serves as a supportive tool rather than a replacement. The proposed framework offers a guide for educators navigating the integration ofAI-assisted learning in classrooms.

significant model public releases within scope

introduced after the watermarking system is developed

Companies commit to advancing this area of research, and to developing a multi-faceted, specialized, and detailed red-teaming regime, including drawing on independent domain experts, for all major public releases of new models within scope

Where commitments mention particular models, they apply only to generative models that are overall more powerful than the current industry frontier (

only a first step in developing and enforcing binding obligations to ensure safety, security, and trus

access to its safe technology could be offered as an incentivefor countries to participate in a governance regime that enforces responsibility (such as agreementsto enact stricter regulation, or restrict military AI development).

uture regulations will limit access to powerful AI technologies in jurisdictions with inadequate AIgovernanc

o increase chances of success, aCommission should foreground scientific rigor and the selection of highly competent AI expertswho work at the cutting edge of technological development and who can continually interpret theever-changing technological and risk landscape

there is significant disagreement even among experts about the different opportunities andchallenges created by advanced AI

International consensus on the opportunities and risks from advanced AI

the Commission on Frontier AI could facilitatescientific consensus by convening experts to conduct rigorous and comprehensive assessments ofkey AI topics, such as interventions to unlock AI’s potential for sustainable development, the effectsof AI regulation on innovation, the distribution of benefits, and possible dual-use capabilities fromadvanced systems and how they ought to be managed

dangerous inputs: computing resources have been targeted by US, Japanese and Dutch export controlsthat prevent the sale of certain AI chips and semiconductor manufacturing equipment to Chin

standards

Establish guidelines

through education, infrastructure, andsupport of the local commercial ecosystem

develop frontier AI collectively or distribute and enable access

developing and/or enabling safe forms of access to AI.

ontrolling AIinputs

My worst fears are that we cause significant, we, the field, the technology, the industry cause significant harm to the world. I think that could happen in a lot of different ways. It’s why we started the company

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

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.

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

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.

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?

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

Sen. Marsha Blackburn (R-TN):

that people own their virtual you.

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.

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.

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?

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.

not a creature,

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

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.

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

a threshold of capabilities

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.

Stability AI announced StableLM - their Language Models.

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

Scalable Personal AI: You can finetune a personalized AI on your laptop in an evening.

it triggers a mental shortcut in the minds of users that we call a “machine heuristic.” This shortcut is the belief that machines are accurate, objective, unbiased, infallible and so on.

In a horizontal approach, regulators create one comprehensive regulation that covers the many impacts AI can have. In a vertical strategy, policymakers take a bespoke approach, creating different regulations to target different applications or types of AI.

2023:

The people who are already well versed in something are going to be the ones capable of making the most helpful applications for that particular field or industry.

arent there already LLM models that cite their sources? or I heard that new plugin with chat GPT can cite its sources

The general consensus is that, especially customer facing automation, MUST be "explainable." Meaning whenever a chat bot or autonomous system writes something or makes a decision, we have to know exactly how and why it came to that conclusion.

They are caught up in the hype and just like everyone else have zero clue what's actually going to happen.

A python module to generate optimized prompts, Prompt-engineering & solve different NLP problems using GPT-n (GPT-3, ChatGPT) based models and return structured python object for easy parsing

OpenAI is developing a tool for “statistically watermarking the outputs of a text [AI system].” Whenever a system — say, ChatGPT — generates text, the tool would embed an “unnoticeable secret signal” indicating where the text came from.

Common sense is different from intelligence in that it is usually something innate and natural to humans that helps them navigate daily life, and cannot really be taught.

2Open Loop is a global program that connects policymakers and technology companies to help develop effective and evidence-based policies around AI and other emerging technologies.

Synthetic data, however it is produced, offers a number of very concrete advantages over using real world data.

There are a couple of ways this synthetic data generation happens

AI and other new technologies will increase strategic instability.

if you don’t have the data, then you have a problem.

We express our concerns on the risk, and ethical dilemma related to Artificial Intelligence, such as privacy, manipulation, bias, human-robot interaction, employment, effects and singularity among others. We encourage BRICS members to work together to deal with such concerns, sharing best practices, conduct comparative study on the subject toward developing a common governance approach which would guide BRICS members on Ethical and responsible use of Artificial Intelligence while facilitating the development of AI.

Consider these sample exchanges:

Text-to-image processes are also impressive. The illustration at the top of this article was produced by using the article’s headline and rubric as a prompt for an ai service called Midjourney. The next illustration is what it made out of “Speculations concerning the first ultraintelligent machine”; “On the dangers of stochastic parrots”, another relevant paper, comes later. Abstract notions do not always produce illustrations that make much or indeed any sense, as the rendering of Mr Etzioni’s declaration that “it was flabbergasting” shows. Less abstract nouns give clearer representations; further on you will see “A woman sitting down with a cat on her lap”.

“Enlightenment”—a trillion-parameter model built at the Beijing Academy of Artificial Intelligence

LaMDA, short for Language Model for Dialogue Applications

It’s not possible to talk about “AI for everyone” (Google’s rhetoric), “responsible AI” (Facebook’s rhetoric), or “broadly distribut[ing]” its benefits (OpenAI’s rhetoric) without honestly acknowledging and confronting the obstacles in the way.

“These simulation engines and the data and everything that goes into them are incredibly important for us to act in these complex environments when it’s not just one crisis but it’s a set of compounding crises,”

“Destination Earth,” will draw on a host of environmental, socioeconomic and satellite data to develop digital “twins” of the planet that aim to help policymakers — and eventually the public — better understand and respond to rising temperatures.

GANs are deep learning models that use two neural networks working against each other—a generator that creates synthetic data, and a discriminator that distinguishes between real and synthetic data—to generate synthetic images that are almost indistinguishable from real ones. GANs are popular for generating images and videos, including deepfakes.

I recently learned that uppercase and lowercase letters got their names from actual wooden cases of lead that were used by compositors for printing.

It began as a term from French railroad engineering referring to the layers of material that go beneath (“infra”) the tracks. Its meaning expanded to include roads, bridges, sewers and power lines, and very recently expanded again to include people, specifically caregivers, as in this fact sheet from the Biden White House

I.C.E. is short for internal combustion engine, a modifier that was superfluous until electric cars came on the scene.

It refers simply to the physical world, where we have tangible bodies made of … meat. “Meatspace” is a word that didn’t need to exist until the invention of cyberspace. Technological progress gives us a new perspective on things we once took for granted, in this case reality itself.

The EU Cybersecurity Act defines cybersecurity as “the activities necessary to protect network and information systems, the users of such systems, and other persons affected by cyber threats”.