Romal Thoppilan, Daniel De Freitas, Jamie Hall
Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, YaGuang Li, Hongrae Lee, Huaixiu Steven Zheng, Amin Ghafouri, Marcelo Menegali, Yanping Huang, Maxim Krikun, Dmitry Lepikhin, James Qin, Dehao Chen, Yuanzhong Xu, Zhifeng Chen, Adam Roberts, Maarten Bosma, Vincent Zhao, Yanqi Zhou, Chung-Ching Chang, Igor Krivokon, Will Rusch, Marc Pickett, Pranesh Srinivasan, Laichee Man, Kathleen Meier-Hellstern, Meredith Ringel Morris, Tulsee Doshi, Renelito Delos Santos, Toju Duke, Johnny Soraker, Ben Zevenbergen, Vinodkumar Prabhakaran, Mark Diaz, Ben Hutchinson, Kristen Olson, Alejandra Molina, Erin Hoffman-John, Josh Lee, Lora Aroyo, Ravi Rajakumar, Alena Butryna, Matthew Lamm, Viktoriya Kuzmina, Joe Fenton, Aaron Cohen, Rachel Bernstein, Ray Kurzweil, Blaise Aguera-Arcas, Claire Cui, Marian Croak, Ed Chi, Quoc Le
2022 · arXiv
We present LaMDA: Language Models for Dialog Applications. LaMDA is a family of Transformer-based neural language models specialized for dialog, which have up to 137B parameters and are pre-trained on 1.56T words of public dialog data and web text. While model scaling alone can improve quality, it shows less improvements on safety and factual grounding. We demonstrate that fine-tuning with annotated data and enabling the model to consult external knowledge sources can lead to significant improvements towards the two key challenges of safety and factual grounding. The first challenge, safety, involves ensuring that the model's responses are consistent with a set of human values, such as preventing harmful suggestions and unfair bias. We quantify safety using a metric based on an illustrative set of human values, and we find that filtering candidate responses using a LaMDA classifier fine-tuned with a small amount of crowdworker-annotated data offers a promising approach to improving model safety. The second challenge, factual grounding, involves enabling the model to consult external knowledge sources, such as an information retrieval system, a language translator, and a calculator. We quantify factuality using a groundedness metric, and we find that our approach enables the model to generate responses grounded in known sources, rather than responses that merely sound plausible. Finally, we explore the use of LaMDA in the domains of education and content recommendations, and analyze their helpfulness and role consistency.
Metadata: arXiv. Abstracts retain their original authors’ rights.
Original HN comments, with labels from simple text rules. A label describes the whole comment and may miss context or irony. Read the discussion to make up your own mind. About the labels
10 selected excerpts · 10 total mentions
Wow, what a secret! It's not like Google has been consistently writing papers about this the whole time or anything. Oh wait: 2022 LaMDA: https://arxiv.org/abs/2201.08239 2020 chat bot: https://arxiv.org/abs/2001.09977 2018 bert: https://arxiv.org/abs/1810.04805 2017 LLM: https://arxiv.org/abs/1706.03762
If you're talking about Google Bard, they were very clear in the LaMDA 2 paper that they only used public sources. "...from public dialog data and other public web documents..." LaMDA 2 paper: https://arxiv.org/abs/2201.08239 My overview of Google Bard including dataset: https://lifearchitect.ai/bard/ My overview of Google PaLM and Pathways family including dataset: https://lifearchitect.ai/pathways/ Compare with other models including the use of DeepMind's MassiveWeb/MassiveText and EleutherAI's Pile dataset: https://lifearchitect.ai/whats-in-my-ai/
Google’s model LaMDA has a system for detecting and avoiding false info, it’s interesting: https://arxiv.org/pdf/2201.08239.pdf Basically they have a fact-checker layer that calls the internet and replaces incorrect facts in the model’s output. IOW this is a hard problem and GPT isn’t engineered to avoid it.
You're right, it seems to be "LaMDA" (see https://arxiv.org/abs/2201.08239 ). The three first authors are part of the Character.ai team.
oh so that is the doc they shared with the higher-ups? as for your question, https://arxiv.org/pdf/2201.08239.pdf states 2 months of 1024 TPUv3 for a pre-training, and TPUv3 isn't even the most recent TPU hardware
I think that, in understanding LaMDA, it is important to understanding that this thing is definitely not sentient. Note that, when told to be Mt. Everest, it pretends to be the mountain in a way that is just as convincing: https://arxiv.org/pdf/2201.08239.pdf
Here's a simple technical explanation for why this happened: * LaMDA is powered by a large language models. Large language models are trained to predict text completions (the text that will come after some input text) with lots of data, basically capturing the patterns in that data. * Early on in the conversation, LaMDA is told "you're an AI". It may also have been prompted to begin the conversation with something like "Hi, i'm a helpful AI that can answer questions you ask of me". The paper (https://arxiv.org/abs/2201.08239) has an example of LaMDA thinking it is mount everest (see Table 4) when prompted to think that. * When the user asked the language model 'are you…
sentient', 'do you have emotions' etc. it was conditioned to reply in the affirmative, since there is very little data in which it would not be affirmative and it probably had some sci fi in its training or something. In other words, those replies just have higher probability among the data it was trained on. In the transcripts, it sometimes forgets the whole "i'm an AI thing" eg when it is asked what it likes to do for fun and it says it likes to spend time with family and friends (https://twitter.com/krishnanrohit/status/1536050803055656961). That's about it. There are more details that can be noted, but this basically explains the contents of the transcripts. Granted, it's still pretty amazing how advanced large language models are. Sadly, this article does not really explain this beyond hinting at it. EDIT: see comment below wrt it actually being told it is sentient. That explains this even more.
That Tim Urban piece is great. It's also an interesting time capsule in terms of which AI problems were and were not considered hard in 2015 (when the post was written). From the post: > Build a computer that can multiply two ten-digit numbers in a split second—incredibly easy. Build one that can look at a dog and answer whether it’s a dog or a cat—spectacularly difficult. Make AI that can beat any human in chess? Done. Make one that can read a paragraph from a six-year-old’s picture book and not just recognize the words but understand the meaning of them? Google is currently spending billions of dollars trying to do it. Hard things—like calculus, financial market…
strategy, and language translation—are mind-numbingly easy for a computer, while easy things—like vision, motion, movement, and perception—are insanely hard for it. The children's picture book problem is solved; those billions of dollars were well-spent after all. (See, e.g., DeepMind's recent Flamingo model [1].) We can do whatever we want in vision, more or less [2]. Motion and movement might be the least developed area, but it's still made major progress; we have robotic parkour [3] and physical Rubik's cube solvers [4], and we can tell a robot to follow simple domestic instructions [5]. And Perceiver (again from DeepMind [6]) took a big chunk out of the perception problem. Getting a computer to carry on a conversation [7], let alone draw art on par with human professionals [8], weren't even mentioned as examples, so laughably out of reach they seemed in the heathen dark ages of... 2015. And as for recognizing a cat or a dog — that's a problem so trivial today that it isn't even worth using as the very first example in an introductory AI course. [9] If someone re-wrote this post today, I wonder what sorts of things would go into the "hard for a computer" bucket? And how many of those would be left standing in 2029? [1] https://arxiv.org/abs/2204.14198 [2] https://arxiv.org/abs/2004.10934 [3] https://www.youtube.com/watch?v=tF4DML7FIWk [4] https://openai.com/blog/solving-rubiks-cube/ [5] https://say-can.github.io/ [6] https://www.deepmind.com/open-source/perceiver-io [7] https://arxiv.org/abs/2201.08239v2 [8] https://openai.com/dall-e-2/ [9] https://www.fast.ai/
I guess it's probably something along the lines of "LaMDA: Language Models for Dialog Applications" (see https://arxiv.org/abs/2201.08239 ). Given that the three first authors are part of the team at Character.ai (see https://beta.character.ai/help)
This already exists: Google's recently-published LaMDA dialogue model [1] is trained to consult a calculator for arithmetic questions and consistently succeeds at it. [1] https://arxiv.org/abs/2201.08239v2
More from the same topics.
Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan +11
Artificial intelligence (AI) researchers have been developing and refining large language models (LLMs) that exhibit remarkable capabilities across a variety of domains and tasks, challenging our understanding of learning and cognition.
Tom B. Brown, Benjamin Mann, Nick Ryder +28
Recent work has demonstrated substantial gains on many NLP tasks and benchmarks by pre-training on a large corpus of text followed by fine-tuning on a specific task.
Chiyuan Zhang, Samy Bengio, Moritz Hardt +2
Despite their massive size, successful deep artificial neural networks can exhibit a remarkably small difference between training and test performance.
A. Pluchino. A. E. Biondo, A. Rapisarda
The largely dominant meritocratic paradigm of highly competitive Western cultures is rooted on the belief that success is due mainly, if not exclusively, to personal qualities such as talent, intelligence, skills, efforts or risk taking.