Falcon Dai

Don't assume, observe. Formerly @TTIC_Connect @Strikingly and @UChicago'12.

[Falcon Dai @falcondai]

Built a demo for a fun algorithm that can hide encrypted messages in LLM-generated natural looking text. https://t.co/wElIGrByrh #llm #privacy #ml

[Falcon Dai @falcondai]

@SebastienBubeck @nytimes @bing Been waiting for this reply. To be fair, their critique was wrt GPT-3.5 and GPT-4 does seem much more capable. What troubles me about the original article is their categorical criticism of the LM paradigm without a fair assessment of limitations of GOFAI.

[Falcon Dai @falcondai]

@natfriedman love it. was thinking to build something similar y-day and I found this today. #zeitgeist

[Falcon Dai @falcondai]

@3blue1brown I imagine that you have seen this excellent WP article https://t.co/CaM8PfN8FP One obvious effect missing in that is the limited medical capacity and death.

[Falcon Dai @falcondai]

@UChicago Perhaps some instructors will find this useful. A quick guide for setting up a virtual blackboard on iPad via Zoom/Google Meet https://t.co/BwgpRQygyn

[Falcon Dai @falcondai]

A review of solutions of using iPad as a virtual blackboard/whiteboard + tips for remote teaching. #COVID19 #SocialDistancing https://t.co/AYzqIcdkPZ

[Falcon Dai @falcondai]

A review of solutions of using iPad as a virtual blackboard/whiteboard + tips for remote teaching. #COVID19 #SocialDistancing https://t.co/AYzqIcdkPZ

[Falcon Dai @falcondai]

We need clear and easy-to-understand communication of timely information to calm the public. Best explainer of the spread and countermeasures of #COVID19!

[Harry Stevens @Harry_Stevens]

If you like bouncing balls explaining how to slow down #coronavirus, my latest story in the @washingtonpost is for you: Why outbreaks like coronavirus spread exponentially, and how to “flatten the curve” https://t.co/28CQLp0Xfc

[Harry Stevens @Harry_Stevens]

If you like bouncing balls explaining how to slow down #coronavirus, my latest story in the @washingtonpost is for you: Why outbreaks like coronavirus spread exponentially, and how to “flatten the curve” https://t.co/28CQLp0Xfc

[Falcon Dai @falcondai]

@johnsorrentino liking your https://t.co/awM8asbDb6. found a minor bug: non-default background options (square or circle) won't get saved (the generated url will have rounded).

[Falcon Dai @falcondai]

@JJJ92296341 @iclr_conf Not a change of rating (discussed and finalized). But wish to reply to authors like via post-rebuttal review updates (in other conferences).

[Falcon Dai @falcondai]

@iclr_conf How should a reviewer add a post-feedback reply? By appending to the original official review? Or will author-readable comments be enabled again before the decision notification deadline?

[Falcon Dai @falcondai]

@WeCNLP Algorithmic safeguards relevant to the issues of communication privacy, misinformation, etc. Both positive and negative theoretical results assuming the availability of strong ML models.

[Falcon Dai @falcondai]

@srush_nlp Cool demo. Congrats! I cannot shake the feeling that we are instruments to manifest the zeitgeist. Thanks for citing our concurrent work :) https://t.co/SYShOTDQ7U

[Falcon Dai @falcondai]

@beenwrekt Thoroughly enjoyed your talk. I wish I could talk to you more about the dynamical sys view of recommendations you mentioned near the end. In particular, it might even be in the interest of *long-term* profit for platforms to consider non-immediate impact of their sys.

[Falcon Dai @falcondai]

@beenwrekt Would the selected model be similar to posterior sampling (Thompson sampling)? Maybe we can be more colorful with names. I prefer to think of optimistic methods (OFU) as Leibnizian after his moral philosophy of "best of all possible worlds"

[Falcon Dai @falcondai]

@hugo_larochelle What if the large binary files are digitally signed (or just hash) and only the signatures are submitted as a proof of their existence at the deadline? The file upload can happen post-deadline.

[Falcon Dai @falcondai]

@harvardnlp @gwern @MITIBMLab A lower bound on samples can be given by total variation distance (TVD) between the two distributions. We need at least omega(1/TVD^2) but I don't know a useful upper bound myself which requires an algorithm that can estimate a distribution over an extended alphabet well.

[Falcon Dai @falcondai]

@harvardnlp @gwern @MITIBMLab The interface is very neat! Indeed. I miss your inspirations. I recently worked on something tangentially related, could be seen as a form of biased sampling in this context. BTW a minor suggestion, maybe putting units on the entropy plot? Nats?

[Falcon Dai @falcondai]

@harvardnlp @gwern @MITIBMLab Let me clarify: I think @gwern pointed out there are two issues, generating from an LM (not LM*) and biased sampling from LM. I am commenting on the former issue. Even having a better LM will not necessarily help with distinguishing LM vs human (LM*) based on samples.

[Falcon Dai @falcondai]

@harvardnlp @gwern @MITIBMLab Furthermore, if this margin TVD(LM*, LM) < TVD(GPT2, LM) is small, you will need to see a lot of texts to achieve high confidence in the judgment asserting that samples come from GPT2 and not LM*.

[Falcon Dai @falcondai]

@harvardnlp @gwern @MITIBMLab I am not sure. Suppose LM* is the true LM. A model just better than GPT2 might not be able to distinguish between GPT2 and LM* because it should not falsely reject texts from LM*. This better LM has to be closer (in terms of TVD or KL) to LM* than to GPT2 to work.