#ACL2024
was great. But very few papers were asking hard scientific questions. A lot of “LLM engineering”. In particular, university labs should really defocus on this type of engineering as industry labs, though not publishing every minute detail, are doing way more advanced
I told President
@EmmanuelMacron
that I started my AI research career from a small project on automatic speech recognition at
#EcoleCentraleParis
when I was a student.
I have long hoped to bring top US and China AI researchers together for a scientific exchange. This finally happened with
@ylecun
, Harry Shum, Feng Junlan (chief AI scientist, China Mobile), Li Hang (Head of Research Bytedance) and Zhou Jingren (CTO of Alibaba Cloud)
Do humans need language for
#reasoning
and
#planning
? It really depends. We “reasons and plan” a lot of physical and artistic activities without thinking in language (eg dance, swimming, painting, playing an instrument…) but language helps us teach/learn how to do them.
The keynote speaker talked for about an hour but cut me off after 10 seconds of my question. If you can’t take a question don’t give a keynote!
#ACL2024
Excellent
#acl2024
keynote by Barbara Plank on how to work with the inherent (and beautiful) variation in languages that leads to uncertainty in LLMs and other AI models. Exactly what NLP needs to hear.
So this paper got a 1 from a reviewer from
#EMNLP
- never gotten a 1 before ever, and is rejected. So I am sharing its arxiv version for feedback. I did get a bunch when I shared it with people in person before. TIA.
It’s 2024, NLP people need to work on the phenomenon of language as manifested in LLMs and other AI models - so that we can make scientific progress - instead of *still* being on the bandwagon of shitting on AGI/LLM endlessly. It’s boring.
#acl2024
We got two awards at
#AACL
2023 - for the ChatGPT benchmarking paper and the Nusawrites Indonesian language resources paper. Congratulations to all co-authors.
For PhD students to avoid becoming just GPT app developers, you need to ask not just « what » you are building but « why » you are doing it. Ask « why » at least every 3 months.
Natural language processing became natural language engineering in the 90s. Now that LLMs can engineer any NLP task by scaling, natural language processing needs to become a science of modelling cognition and metacognition via language.
Got attacked by a man for not citing enough women authors in my keynote on hallucination at
#aaai2024
#Realaaai
showing that alignment tax is not just a problem for GenAI but humans as well. Photo by
@FrancescaRossi_
LLMs have learned human values. Ever wonder how value distributions differ between LLMs? Do ChatGPT and Llama have the same values in English, French, etc.? Are Asian values closer to each other than with Western values? Look for the "value maps".
A number of us from
#hkust
are among the AI top 100 contributors of 2023 by an automatic metric by Bench Council. HKUST and Tsinghua are the top 2 from China. (This is a metric based on publications and citations, not social media or media influence.)
Our research team has published a question answering engine with abstractive summarization from the
dataset of CoVID 19 publications. It shows two versions of abstracts generated by deep learning models and a top 10 snippets that answer the question.
To eliminate any human intention, I generated this from an AI model by prompting with a random string of keyboard strokes. So this is pure
#AIart
it is breathtakingly beautiful, reminiscent of 60s psychedelic pop art and Vasarely but not from any artist in particular.
GenAI tools trained on content cannot generate anything without a human creator prompting them to do so. So AI art is the fruit of human creation as much as digital art or photography art and for that matter, human artists who learned their craft from prior work.
The problem with these images is due to “alignment tax” - when a model has been trained to align with multiple objectives it becomes hard to balance between them. For example here is an apparent conflict between “diversity” and “faithfulness”. It is a hard optimisation problem.
I really love the active discussion abt the role of ethics in AI, spurred by Google Gemini's text-to-image launch & its relative lack of white representation. As one of the most experienced AI ethics people in the world (>4 years! ha), let me help explain what's going on a bit.
@ednewtonrex
AI does not create anything without human prompts. The creation of an artist using AI is not different in essence from that of digital art or photographic art which are also scalable. There is nothing “copied” by AI art as there is nothing “copied” by photographic or digital art.
@ednewtonrex
Regulations need to be designed by those humble enough to study the current technology. You cannot regulate what you don’t understand with any wisdom.
Yoshua Bengio’s talk and Q and A on the necessity of AI safety via governance, Bayesian estimation of uncertain and risks, and quantitative harm prevention at
#HKUST
“techbro” is a sexist term invented by lazy journalists. Now they keep using this term to describe the AI community and basically deleting all the women in AI. If you are not ignorant do not use this term.
#responsibleAI
should be practiced by every research group big or small. We absolutely need to hold people accountable for the accessibility of their datasets and code and ensure reproducibility before we give out any paper award, if not every publication.
Need a semantic benchmark for Sora and the like. A beautiful and dynamic video of stuff counter to the laws of physics (like that of the shattering glass) is no use. It’s just beautiful hallucination. They must be doing RLHF on Sora now?
Most image generation tools are built by engineers who are accustomed to tasks like video captioning and object recognition. This is NOT what artists need for creative work. In fact artists are currently creating work within the constraints of these very limiting tools.
@ylecun
C obviously and depending on the audience size it is disseminated to. Just follow current copyright laws. Art students routinely copy the grand masters in their studies, nobody calls that copyright infringement.
@kaushik_himself
@ylecun
@sama
@gdb
OpenAI couldn’t have done it without the open-sourced Transformer and many models from Google Meta etc before them. It’s like pulling up the ladder after you have climbed up.
Google, as well as Meta and OpenAI, among others, have large teams of people dedicated to ethics and responsible AI. There is currently no tractable solution to avoid all errors (can any human claim to be ethically perfect?) but the research community will keep trying.
The problem with these images is due to “alignment tax” - when a model has been trained to align with multiple objectives it becomes hard to balance between them. For example here is an apparent conflict between “diversity” and “faithfulness”. It is a hard optimisation problem.
Has anyone noticed that
#Sora
has problems with
#perspectives
in some of the videos it generated? For example in this one. Also it seems to be trained from video game data???
Do humans need language for
#reasoning
and
#planning
? It really depends. We “reasons and plan” a lot of physical and artistic activities without thinking in language (eg dance, swimming, painting, playing an instrument…) but language helps us learn to do them faster.
To people who claim that "thinking and reasoning require language", here is a problem:
Imagine standing at the North Pole of the Earth.
Walk in any direction, in a straight line, for 1 km.
Now turn 90 degrees to the left.
Walk for as long as it takes to pass your starting point.
@ylecun
And anyone “infringing copyrights” in private going through a platform cannot be sued although the platform can detect and prevent dissemnination.
@deliprao
NN does not learn rules. Before NN, there were statistical
AI that learned probabilistic rules automatically. •Rule based systems” are those that are based on manual rules by people. People are very good at pattern recognition but absolutely terrible at coming up with rules!
@DrJimFan
@ylecun
I agree with
@DrJimFan
. However I also with the design of this imperfect benchmark with all the caveats. (1) It is impossible to evaluate extrinsic hallucinations so summarisation is a good task as a proxy for intrinsic hallucination evaluation.
Today we're announcing a new privacy-preserving approach to improve fairness & robustness of automatic speech recognition systems. This unique approach lets researchers improve ASR performance without relying on demographic data.
More info ⬇️
There are many greater achievements than our papers in 2023. Google and Meta are the top 2. (OpenAI doesn’t publish much.) US is number 1 with China at number 2.
A number of us from
#hkust
are among the AI top 100 contributors of 2023 by an automatic metric by Bench Council. HKUST and Tsinghua are the top 2 from China. (This is a metric based on publications and citations, not social media or media influence.)
@ylecun
@cwizprod1
@harryshum
This is because the apocalypse and doomsday scenario is very much a Christien religious thing influencing western thinking,
What is the meaning of “a world class prompt engineer”? Who certifies one as “world class” and what qualifies one as a “prompt engineer”? (I am not dismissing the importance of
#prompts
. )
Representation by pixels in images never made much sense. What about subobject representation like subwords in NLP? They contains more semantic information than pixels.
@DrJimFan
@ylecun
2) I would not get too attached to the absolute performance scores but the relative performance is probably credible. Leaderboards and benchmarks are good for comparing models.
@Liv_Boeree
That’s a weird classification. Did you come up with that yourself? Most people working on AI safety are motivated by AI ethics. Without ethics there is no principled way of implementing AI safety.
@ednewtonrex
GenAI tools trained on content cannot generate anything without a human creator prompting them to do so. So AI art is the fruit of human creation as much as digital art or photography art and for that matter, human artists who learned their craft from prior work.
@neilturkewitz
@ednewtonrex
The issue is that the people sitting at the top of the bureaucracy are probably not the people who have the wisdom and knowledge to provide such oversight for a technology that will have this much impact.
@gerardsans
If he didn’t cut me off I would have finished my question in 1 minute. I would have asked him to elaborate on his thesis on why LLM cannot plan well. Instead he went on a tangent about paper reviews etc like a defensive student. His talk was disrespectful of a scientific audience
@ziqi_huang_
This is a video quality evaluation. What’s important is semantic evaluation. A beautiful video portraying things counter to physics is no good.
@gjzhang1
There are many scientific questions to ask. If you want to work on LLM then ask why do LLMs work or does not work in X or Y? How can we benchmark and improve different aspects of machine intelligence - reasoning, planning, world modeling etc? In NLP how is language being
Professional cinematographers are looking at new videos from Sora and breathing a sign of relief - there is only so much you can do prompting when you are untrained. Same story with image generators. The beauty and the creativity still need to come from humans prompting.
@lateinteraction
RAG alone cannot solve the hallucination problem and is not even a guarantee to improve the outcome. Multiple retrieved facts together can still be generated in an incoherent way. Just try Bing. 2/2
@adawan919
I strongly that we need better benchmarking and better understanding of how they work. And how to do that would have been a scientific discussion.