![BuzzRobot](/img/default-banner.jpg)
- Видео 106
- Просмотров 810 877
BuzzRobot
США
Добавлен 23 авг 2020
Sophia, the founder of BuzzRobot, facilitates talks in which top AI researchers and engineers explore the cutting edge of AI research. We feature talks covering a wide range of topics - AI, Machine Learning, Generative AI, Large Language Models (LLMs), Reinforcement Learning, Computer Vision, Robotics, Autonomous Systems, Programming, and more.
Join the BuzzRobot Slack to connect with the community.
Join the BuzzRobot Slack to connect with the community.
Google's Med-Gemini: Game-Changing AI in Healthcare
@Google_DeepMind's Gemini models have very strong potential in the medical domain, which can improve the lives of billions and accelerate scientific progress.
During the @BuzzRobot event, researchers Khaled Saab and Wei-Hung Weng introduced Med-Gemini, a family of multimodal models specializing in medicine. Some of the cool capabilities of Med-Gemini - its long context window, the ability to perform internet search to enhance its reasoning, and its multimodality, which allows it to recognize complex surgical images.
This lecture covers some of the technical details behind the work.
#multimodal #multimodalai #largelanguagemodels #largelanguagemodel #healthcareinnovation #artificialintelligen...
During the @BuzzRobot event, researchers Khaled Saab and Wei-Hung Weng introduced Med-Gemini, a family of multimodal models specializing in medicine. Some of the cool capabilities of Med-Gemini - its long context window, the ability to perform internet search to enhance its reasoning, and its multimodality, which allows it to recognize complex surgical images.
This lecture covers some of the technical details behind the work.
#multimodal #multimodalai #largelanguagemodels #largelanguagemodel #healthcareinnovation #artificialintelligen...
Просмотров: 32
Видео
Using AI to Create 3D Worlds to Train #reinforcementlearning Agents
Просмотров 27919 часов назад
In this talk, the @BuzzRobot guest, Fan-Yun Sun from @stanfordUniversity, discusses how Compound AI systems can be used to generate environments and 3D scenes for #robotics and AI agent training. Timestamps: 0:00 Introduction 01:14 Contextual intelligence. Reasoning. How we can leverage them in AI systems. 11:42 Using Compound AI systems to generate 3D worlds. 12:50 Holodeck: Language Guided Ge...
The 18 Biggest Challenges in #llm Alignment and Safety: @BuzzRobot talk
Просмотров 19014 дней назад
Alignment and safety of Large Language Models (LLMs) are the cornerstones of modern artificial intelligence (AI) research, and prominent AI companies like Google DeepMind and OpenAI, among others, are dedicating research resources to AI safety. Despite these efforts, there are foundational challenges - 18 of them - that researchers have identified. In this lecture, the @BuzzRobot guest, Usman A...
Towards Guaranteed Safe AI: A Framework to Ensure Robust and Reliable AI Systems
Просмотров 1,1 тыс.28 дней назад
In this video, a researcher from Oxford University, Joar Skalse, proposes a framework that anticipates different stages of AI systems' development. The framework introduces world models, safety specifications, and verification methods for testing the safety of AI systems. It includes suggestions from prominent AI researchers and scientists, including Yoshua Bengio and Max Tegmark. The biggest a...
The Ethical and Societal Impact of AI Assistants by Google DeepMind
Просмотров 4,9 тыс.Месяц назад
Should a developer regulate human-AI love relationships? Can AI break the law in certain cases, such as in life-or-death situations? How do we balance conflicting relationships between AI, users, developers, and society? Is AI conscious? These and many more topics were discussed by Google DeepMind researcher Iason Gabriel with the @BuzzRobot community, laying out the foundation for an ethical f...
The Griffin architecture: A challenger to the Transformer
Просмотров 9 тыс.Месяц назад
The @BuzzRobot guest, Aleksandar Botev, a research scientist at Google DeepMind, presented an efficient alternative to the Transformer model architecture - Griffin. Griffin combines gated linear recurrences with local attention. It showed great performance on long sequences and does well during inference time compared to Transformers, having lower latency and higher throughput. Its biggest cave...
RLOO: A Cost-Efficient Optimization for Learning from Human Feedback in LLMs
Просмотров 3,4 тыс.2 месяца назад
RLOO: A Cost-Efficient Optimization for Learning from Human Feedback in LLMs
Exploring efficient alternatives to Transformer models
Просмотров 10 тыс.2 месяца назад
Exploring efficient alternatives to Transformer models
DiPaCo: Towards a New Paradigm of Distributed AI Training by Google DeepMind
Просмотров 12 тыс.2 месяца назад
DiPaCo: Towards a New Paradigm of Distributed AI Training by Google DeepMind
Lecture about Llama 3 with Thomas Scialom, an AGI researcher at Meta
Просмотров 23 тыс.3 месяца назад
Lecture about Llama 3 with Thomas Scialom, an AGI researcher at Meta
Aligning AI with Pluralistic Human Values
Просмотров 12 тыс.3 месяца назад
Aligning AI with Pluralistic Human Values
Aya, an Open-Source Dataset and Model for Over 100 Languages
Просмотров 11 тыс.3 месяца назад
Aya, an Open-Source Dataset and Model for Over 100 Languages
How Hardware Choices Impact Fairness in AI Systems
Просмотров 5 тыс.3 месяца назад
How Hardware Choices Impact Fairness in AI Systems
Databricks LLM, DBRX: Model design and challenges. The lecture for the @BuzzRobot community
Просмотров 5 тыс.3 месяца назад
Databricks LLM, DBRX: Model design and challenges. The lecture for the @BuzzRobot community
Large Language Models Might be Vulnerable to Trojan Virus. Demo
Просмотров 6 тыс.4 месяца назад
Large Language Models Might be Vulnerable to Trojan Virus. Demo
Large Language Models Vulnerability From GPU Local Memory Leak. AI Security
Просмотров 5 тыс.4 месяца назад
Large Language Models Vulnerability From GPU Local Memory Leak. AI Security
AI to Automatically Fix Security Bugs by Google
Просмотров 6 тыс.4 месяца назад
AI to Automatically Fix Security Bugs by Google
GPT-4 Based Language Model That Predicts The Future by UC Berkeley
Просмотров 4,9 тыс.4 месяца назад
GPT-4 Based Language Model That Predicts The Future by UC Berkeley
Towards Deployable Robotics Systems: Parkour Learning and Imitation Learning by Stanford AI Lab
Просмотров 7 тыс.5 месяцев назад
Towards Deployable Robotics Systems: Parkour Learning and Imitation Learning by Stanford AI Lab
Best strategies for distributed training of LLMs (Large Language Models)
Просмотров 10 тыс.5 месяцев назад
Best strategies for distributed training of LLMs (Large Language Models)
Testing the Quality of Synthetic Data With Self-Training by Google DeepMind
Просмотров 2,7 тыс.5 месяцев назад
Testing the Quality of Synthetic Data With Self-Training by Google DeepMind
Path to AGI: Massively multi-agent environment for RL research
Просмотров 3,9 тыс.5 месяцев назад
Path to AGI: Massively multi-agent environment for RL research
How to be hired as an AI researcher?
Просмотров 2,4 тыс.5 месяцев назад
How to be hired as an AI researcher?
The predictions of top researchers on future advancements of AI (Artificial Intelligence)
Просмотров 1,7 тыс.5 месяцев назад
The predictions of top researchers on future advancements of AI (Artificial Intelligence)
What are the important skills of an AI researcher?
Просмотров 2,8 тыс.5 месяцев назад
What are the important skills of an AI researcher?
Graph Neural Networks for Skillful Weather Forecasting by Google DeepMind
Просмотров 4,5 тыс.6 месяцев назад
Graph Neural Networks for Skillful Weather Forecasting by Google DeepMind
Safety Risks of LLMs (Large Language Models) by Princeton University
Просмотров 3,7 тыс.6 месяцев назад
Safety Risks of LLMs (Large Language Models) by Princeton University
A Co-Habitat for Humans, Avatars, and Robots by Meta
Просмотров 6736 месяцев назад
A Co-Habitat for Humans, Avatars, and Robots by Meta
How can robots develop visual imagination? By MIT
Просмотров 3307 месяцев назад
How can robots develop visual imagination? By MIT
New benchmarks in vision-language models for real-world use: Google Research
Просмотров 2297 месяцев назад
New benchmarks in vision-language models for real-world use: Google Research
Wish I could've attended this in person! ❤
Thanks for this super informative video! Med-Gemini is next level 🔥Big shoutout to Khaled Saab and Wei-Hung Weng for introducing us to this exciting tech - it was really helpful in understanding how it’s going to change healthcare 👍
Thank you for organizing this talk! :) Learning about these cutting-edge ML techniques has been very enriching.
Thank you for sharing!
Thanks for such an interesting presentation 👏Looking forward to seeing how compound AI systems evolve!
How does contextual intelligence deal with multi-modal data streams in real-time? Wouldn’t that require a crazy amount of Neural power? Considering the systems and chips we have right now, how long would it realistically take to get to the ‘comfortable’ stage? Great talk! The applications of contextual intelligence are actually crazy!
Also, I’m just thinking how useful contextual intelligence can be for game environment designers, especially when they hit a creative block
That's a great question. I asked our guest about it - will share his thoughts once I hear back from him.
Here is the response from the guest speaker: Contextual intelligence, in my view, is a means of optimizing computational resources rather than increasing them. Instead of running a complete forward pass through a large foundational model for every step of query, contextual intelligence allows us to only consider the relevant context. I have strong conviction that "contextual intelligence" will be a key aspect to the future of compound AI systems, including systems can seamlessly handle real-time multi-modal data streams.
Interesting use case of LLMs. I really enjoy when the focus of research is towards solving some real life applications. Thanks to the speaker for the cool results and the organizer for sharing this insightful talk.
Thank you for sharing! This was a very cool talk
Glad you enjoyed it!
These are the papers that Fan-Yun mentioned during the talk. Holodeck: yueyang1996.github.io/holodeck/ FactorSim cs.stanford.edu/~sunfanyun/factorsim/
Very cool, thank you for the talk!
Thanks!
After listening to this talk - I thought it would be cool if those AI systems could also generate 3D worlds for gaming. For what other use cases can we use AI-generated worlds?
I see a lot of scope in schools. Kids can be taught complex concepts visually using AI synthesised videos or images for better illustrations.
@@pranavagarwal779 I wish I had as a kid the things that kids in the future will have to learn about the world lol
@@BuzzRobot I guess everything has its trade-offs. A few years ago, with less technology and social media, people generally had a wider attention span, making it easier to focus and learn complex topics. However, with the increasing prevalence of technology, attention spans have been decreasing (according to some stats attention span in humans decreased from 12 seconds in 2000 to 8 seconds in 2013 primarily because of the use of social media). Therefore, it’s important to leverage external technologies to help the upcoming generation learn complex subjects effectively.
@@pranavagarwal779 Attention span is probably even lower now, and humans will become more and more reliant on technology. Already we basically can't live without our phones. At some point, we might physically merge with technology.
Timestamps: 0:00 Introduction 01:14 Contextual intelligence. Reasoning. How we can leverage them in AI systems. 11:42 Using Compound AI systems to generate 3D worlds. 12:50 Holodeck: Language Guided Generation of 3D Embodied AI Environments. 21:02 FactorSim: Generative Simulation via Factorized Representation. 29:30 Takeaways of the talk.
Very interesting talk, appreciated the insight, thank you very much!
Thank you Sophia! Great talk🎉
@@SophieGoPhi thanks for attending!
Thank you for the insightful talk!
@@MihaiTodor thanks a lot coming!
LLMs are in the same phase the internet was around 20 years ago, people back then thought it was just hype, now we can not even survive a minute without the internet. The same would happen with LLMs. A large part of our lives can be automated using LLMs saving time and cost. People avoiding it now may have to catch up in the future. This was a really interesting talk. Thanks to the speaker and the organizer.
@@pranavagarwal779 do you think llm will be a new form of interface? Let’s say we won’t do search by googling, won’t physically use laptop but rather ask our ai assistant via voice to find whatever info we need or accomplish any task?
@@BuzzRobot Definitely this is possible and is the future but ideally, I would like a scenario where a non-invasive approach can comprehend our feelings and make decisions accordingly. I don't think voice or text is an optimal approach to prompting an LLM as many feelings are hard to express using voice or text and can be easily misunderstood.
@@pranavagarwal779 but how then will it be able to comprehend our feelings? Measuring our heartbeat or other vitals?
@@BuzzRobot Something like this device ruclips.net/video/RuUSc53Xpeg/видео.html, this was really inspiring and I feel this can help us achieve automation at a rate no one has imagined.
@@BuzzRobot The device by MIT Media Lab -- AlterEgo: Interfacing with devices through silent speech seems a really good solution. Somehow I am unable to provide a link in the reply, but RUclips has an interesting video explaining that.
Thanks for the insightful talk! I loved the quote "AGI is the dream, and we're always just a few years away from it" 👏Fingers crossed we tackle the AI safety challenges soon so we can focus on the possibilities without worrying about catastrophic risks.
@@martinasardaryan2711 yeah if not all of the challenges-probably it’s impossible but at least many of them.
I feel LLMs will be very very mainstream in the near future, so I’m glad Usman touched on LLM safety. Great talk!
@@PH03NXHDFYeah yeah especially considering how much of VCs money goes into llm related startups
This is the research work covered in this lecture llm-safety-challenges.github.io/challenges_llms.pdf
Timestamps: 0:00 Introduction 0:44 The concept of Artificial General Intelligence (AGI) 2:36 The concept of this work 6:25 The list of the foundational challenges of LLMs 11:42 Who can contribute to LLM safety 20:08 Importance of interdisciplinary collaboration 26:46 If you are interested in digging more into LLM safety - here is a bit of guidance on how to best engage with this research work 28:46 Q&A
Insightful!
@@pranavagarwal779 thanks
It's here to stay anyway since even in its current form there are many products which can be built around it. Guess the "AI winter" discussion is more about investors who want to get a fat payoff in the future
according to Pitchbook, AI investment deals have started to drop but we still see VC whales investing hundreds of millions of dollars into freshly baked startups - probably signaling that even in investments we are unlikely to see an AI winter.
Great talk, thank you for sharing!
Thanks for coming to the talk!
Timestamps: 0:00 Introduction to AI safety issues 8:26 Designing "world models" for testing AI systems 15:28 Elaborating on safety specifications of AI systems 20:37 Verification methods for AI systems' safety 23:01 Q&A
...
It's a constant debate not just about whether AI can be conscious, but also what consciousness actually is, there's no widely accepted definition yet🙄. Maybe once people start seeing agency in AI's behavior, it'll make sense to start considering its consciousness🤔
Yeah, I just hope it won't be too late for humanity
This is a very comprehensive ethical framework. I am curious if it could be modified to account for players - nations, organisations, individuals - who operate outside such frameworks.
Thanks for attending the talk and for your question. Here is the answer that Iason provided: the answer probably varies a bit according to the actor involved. But in many ways this *is* the challenge -- to steer a whole range of actors, and the technology itself, towards better outcomes (given the varied but sometimes limited tools at our disposal e.g. incentives and penalties).
Timestamps: 0:00 Introduction: Iason's background 0:01:23 What are AI Agents? 0:03:37 What are Advanced Assistants? 0:10:13 Safety and Alignment Issues Associated with Advanced Assistants 0:15:27 Challenges Associated with Human-Assistant Interaction 0:22:15 AI Assistants and Society: Issues in This Dynamic 0:26:16 Evaluating Advanced Assistants 0:30:35 What We Might Ask for When It Comes to Advanced Assistants 0:33:26 Conclusion 0:34:50 Q&A
It's becoming clear that with all the brain and consciousness theories out there, the proof will be in the pudding. By this I mean, can any particular theory be used to create a human adult level conscious machine. My bet is on the late Gerald Edelman's Extended Theory of Neuronal Group Selection. The lead group in robotics based on this theory is the Neurorobotics Lab at UC at Irvine. Dr. Edelman distinguished between primary consciousness, which came first in evolution, and that humans share with other conscious animals, and higher order consciousness, which came to only humans with the acquisition of language. A machine with only primary consciousness will probably have to come first. What I find special about the TNGS is the Darwin series of automata created at the Neurosciences Institute by Dr. Edelman and his colleagues in the 1990's and 2000's. These machines perform in the real world, not in a restricted simulated world, and display convincing physical behavior indicative of higher psychological functions necessary for consciousness, such as perceptual categorization, memory, and learning. They are based on realistic models of the parts of the biological brain that the theory claims subserve these functions. The extended TNGS allows for the emergence of consciousness based only on further evolutionary development of the brain areas responsible for these functions, in a parsimonious way. No other research I've encountered is anywhere near as convincing. I post because on almost every video and article about the brain and consciousness that I encounter, the attitude seems to be that we still know next to nothing about how the brain and consciousness work; that there's lots of data but no unifying theory. I believe the extended TNGS is that theory. My motivation is to keep that theory in front of the public. And obviously, I consider it the route to a truly conscious machine, primary and higher-order. My advice to people who want to create a conscious machine is to seriously ground themselves in the extended TNGS and the Darwin automata first, and proceed from there, by applying to Jeff Krichmar's lab at UC Irvine, possibly. Dr. Edelman's roadmap to a conscious machine is at arxiv.org/abs/2105.10461
Join our community on Slack join.slack.com/t/buzzrobot/shared_invite/zt-1zsh7k8pd-iMu_M8bUxIK3pOJgqJgCRQ
Timestamps: 0:00 Introduction 0:54 Historical overview 2:49 What are State Space Models? 7:04 The design of the new alternative architecture 13:43 Why input gating is important 17:18 Scaling State Space Models 19:03 Model efficiency on device 20:39 A note on efficiency improvements 22:04 How do alternative models perform during inference 24:34 The caveat of the alternative models 26:11 Summary 27:13 Q&A
Timestamps: 0:00 Introduction 0:33 Background on RLHF (Reinforcement learning from Human Feedback) 5:45 Back to basics: REINFORCE 8:44 From PPO (Proximal Policy Optimization) to REINFORCE 23:27 Results of the new optimization method, RLOO 32:35 Conclusions 34:05 Q&A
Join BuzzRobot community on Slack join.slack.com/t/buzzrobot/shared_invite/zt-1zsh7k8pd-iMu_M8bUxIK3pOJgqJgCRQ
Timestamps: 0:00 Introduction 4:31 Sneak peek into the past to see how Transformers took off with language models, explanation of the basics of Transformers. 7:49 2022: The emergence of Transformer alternatives. Hungry Hungry Hippos: Towards language modeling with State Space Models. 19:32 2024: Modern times. Synthetics in modern efficient ML models. 24:22 Demos of what models based on new architectures can do today. 26:21 Q&A session
Join BuzzRobot community on Slack: join.slack.com/t/buzzrobot/shared_invite/zt-1zsh7k8pd-iMu_M8bUxIK3pOJgqJgCRQ
Timestamps: 0:00 Introduction 0:45 Google DeepMind's vision of distributed training 2:57 Building blocks towards DiPaCo 11:33 Building blocks towards DiPaCo: DiLoCo - a low communication distributed training optimization 18:41 Introducing DiPaCo: Distributed Paths Composition 22:37 Q&A session
🎉😮🎉
To me there is an understated and underaddressed risk to building powerful alignment tuning techniques Consider the potential threat to humanity posed by those tools in the wrong hands To what degree has this potential danger been considered, and should it be considered more or differently?
Humans have somehow figured out the protocol around nuclear weapons and so far humanity still exists. Most likely governments and the AI community will figure out how to manage AI systems. I am personally more concerned about what will happen when AI starts improving itself and humans won't be able to do anything about it
Join the BuzzRobot community of developers, ML engineers, and researchers join.slack.com/t/buzzrobot/shared_invite/zt-1zsh7k8pd-iMu_M8bUxIK3pOJgqJgCRQ
Timestamps: 0:00 Introduction 0:29 Open access models vs. private models 6:39 Training data for Llama 3 10:36 LLMs as AI agents 16:26 Superalignment 20:29 Interpretability of large-scale models 22:19 Is compute the only answer when it comes to intelligent AI systems? 27:14 Can small research companies without enormous funding bring value to the AI community? 28:46 What are the future developments for the Llama family models? 30:07 AGI and how humans and intelligent AI systems could co-exist 32:09 Career advice: How to get into AI-first companies
How to implement the paper
check out this github repo: github.com/llm-attacks/llm-attacks
@@BuzzRobot I know about this but if someone can make a tutorial for setting up this repo then it would be cool
@@noadsensehere9195 Gotcha. Let me check this with the researcher
@@BuzzRobot thanks
@@BuzzRobot There are lot of dependencies error I am facing!
🎉🎉
Join the BuzzRobot community on Slack join.slack.com/t/buzzrobot/shared_invite/zt-1zsh7k8pd-iMu_M8bUxIK3pOJgqJgCRQ
Timestamps: 00:00 Introduction 04:00 ValuePrism: A large-scale dataset of 218k values, rights, and duties connected to 31k human-written situations. 07:50 Introducing the Kaleido System. 16:23 A roadmap to pluralistic alignment. 28:13 Q&A
super!
Thanks for taking the time to watch
Join the BuzzRobot community on Slack join.slack.com/t/buzzrobot/shared_invite/zt-1zsh7k8pd-iMu_M8bUxIK3pOJgqJgCRQ
Timestamps: 0:00 Introduction to the Aya project 1:53 Why is it important to enable multilingual LLMs, and what is the current state of the art? 6:28 How the Aya dataset and collection were built 13:58 Details about the Aya language model that can follow instructions in over 100 languages 24:47 Conclusions 26:07 The Q&A session
Is there some paper link?
hi! please check out this paper arxiv.org/abs/2110.14168
Join the BuzzRobot community on Slack to connect with fellow developers, ML engineers, researchers, and practitioners join.slack.com/t/buzzrobot/shared_invite/zt-1zsh7k8pd-iMu_M8bUxIK3pOJgqJgCRQ
Timestamps 0:00 Introduction 8:45 Impact of hardware selection on fairness 12:30 Why hardware tooling induces unfairness 20:05 Experiments to showcase the issue 24:15 Exploring mitigation solutions to address the problem 27:42 The Q&A