joaquind

@luiscapobianco It’s open source Luis:-)

Too much fuzz about DeepSeek. But how good is an AI so controlled by the Chinese government censorship? What else does the Chinese government controls about this AI?

Check out Maestro, our AI playlist generator! aboutamazon.com/news/entertain…

@burkov I agree that human intelligence is much more than language but we *humans* always tout spoken and written language as the vehicle of our advancements. How do you known our internal neural nets are not doing exactly that: predicting next word in stringing things together?

Don't let them fool you: AGI today is no nearer to us than it was two years ago. While ChatGPT might appear to be a step closer to AGI, from a scientific standpoint, it's not: training a neural network to predict the next word is not groundbreaking science. Achieving AGI would necessitate multiple significant scientific breakthroughs. I'm talking about genuine breakthroughs, where real scientists engage in real science, not just enlarging the scale of the autocomplete.

I just published My Silicon Valley Journey (Part 3 of 3): It’s a Small World After All link.medium.com/DsVuKHZK1Fb

Hour of code. I explain about AI and invite kids to program a dance party linkedin.com/posts/joaquind…

@ylecun Once auto-regressive models master planning do you still think we are very far from AGI?

Please ignore the deluge of complete nonsense about Q*. One of the main challenges to improve LLM reliability is to replace Auto-Regressive token prediction with planning. Pretty much every top lab (FAIR, DeepMind, OpenAI etc) is working on that and some have already published ideas and results. It is likely that Q* is OpenAI attempts at planning. They pretty much hired Noam Brown (of Libratus/poker and Cicero/Diplomacy fame) to work on that. [Note: I've been advocating for deep learning architecture capable of planning since 2016].

@ZhiyingJ @LukeGessler Doesn’t surprise me. DNNs are a just a sophisticated form of compression!

For anyone who’s interested, here is the code github.com/bazingagin/npc…. btw, I’m the author of the paper and thanks @LukeGessler for digging my paper out of that many ACL papers lol 😂

@karpathy I think this would be even more useful for people that are vision impaired

I wish I could ask questions of GPT about things that I’m randomly looking at or working with. An omnipresent assistant. Feels tractable, current constraint I think is the ease of I/O, mostly on the embedded side. (prompted by wanting to ask a Q about a paragraph in a book)

A cool use of reinforcement learning to improve information retrieval through user feedback by @joaquind and Paul Greyson @AmazonScience amazon.science/blog/from-stru…

@ylecun @ylecun You only refer to the pitfalls of LLMs in the context of Generative Question-Answering (e.g. ChatGPT), but what about Extractive Question-Answering where answers are extracted verbatim from the document text itself, thus is more reliable (no hallucination)?

Good article on LLMs at Forbes. The media are starting to agree with my much-criticized statements about LLMs. "LLMs as they exist today will never replace Google Search. Why not? In short, because today’s LLMs make stuff up." forbes.com/sites/robtoews…

@ylecun What ads would you make today @ylecun ?

AT&T could make those wild claims with confidence because the underlying technology for all of these wonders was actually being developed at Bell Labs. (I was working there at the time). [The tablet-like "fax from the beach" thingy was an actual AT&T product].

@ylecun Right, I don’t think anyone is denying what @fraser or @ylecun said. The problem is that META did not package it right in their first attempt to expose it to the public, while OpenAI did. Just acknowledge this and let’s move on.

From the head of product at OpenAI who just left OpenAI.

@xamat @JustinBasilico I also feel we should talk about the convergence of RecSys, AdvSys and Search (IR) which are fundamentally converging to the same base architecture!!!

Ten years ago @JustinBasilico & me published a blog post describing an architectural blueprint for Recommender Systems. I'm now revisiting it by including several alternatives published since, and a new one that in some ways includes all the previous ones: amatriain.net/blog/RecsysArc…

@xamat @JustinBasilico Great recap!! Not sure your new blueprint introduced anything novel @xamat. For example previous blueprints already have separate models for retrieval and ranking.

@ylecun Hey, humans make stuff up… does that make ChatGPT less useful? I think it should defenitly be improved by clarifying fiction from fact through citing sources. The reason ChatGPT is so different (and better) to what Meta has done is the task and the data it was trained on.

Excellent WaPo article about large language models and chatbots that corroborates what I've been posting recently: they are useful but they make stuff up. They detail the reasons why large tech cos have been hesitant to release such things for public use. washingtonpost.com/technology/202…

@alansaid No wonder we see so much overlap!!

34% of submitted papers to #sigir2023 deal with something related to #recsys

@alansaid And advertisements and e-commerce search/recs are contextualized recommendations that try to satisfy multiple objectives (e.g. customer, publisher, advertiser and platform objectives)

@alansaid Well, search is just a form of. Contextualizad recommmedations (the context here is given by the query that represents user intent)

@ylecun @DataChaz @Meta @facebook Because it works better!

@DataChaz @Meta @facebook You are making my point. Ask yourself why Galactica was pilloried and crucified while chatGPT (which has similar flaws) was welcome as the second coming of the Messiah?

🚨 Correcting @ylecun here: @Meta/@facebook *did* release a #ChatGPT-like thing, the highly-anticipated #Galactica, but it was withdrawn just three days after its release, following a deluge of trolling and accusations of bias! The internet remembers: techunwrapped.com/galactica-meta…