zxp

PhD

[zxp @zxp567]

𝗔𝗻 𝗮𝗻𝘀𝘄𝗲𝗿-𝗼𝗻𝗹𝘆 𝗺𝘂𝗹𝘁𝗶-𝗮𝗴𝗲𝗻𝘁 𝗽𝗿𝗼𝘁𝗼𝗰𝗼𝗹 𝗹𝗼𝗼𝗸𝘀 𝗹𝗶𝗸𝗲 𝗰𝗼𝗹𝗹𝗮𝗯𝗼𝗿𝗮𝘁𝗶𝗼𝗻, 𝗯𝘂𝘁 𝗯𝗲𝗵𝗮𝘃𝗲𝘀 𝗹𝗶𝗸𝗲 𝗮 𝗽𝗼𝗹𝗹.

Write-up + code:
https://t.co/6S0c7qVYZP

[zxp @zxp567]

𝗗𝗼𝗲𝘀 𝗿𝗲𝗱𝘂𝗰𝗶𝗻𝗴 𝘀𝘆𝗰𝗼𝗽𝗵𝗮𝗻𝗰𝘆 𝗮𝗰𝘁𝘂𝗮𝗹𝗹𝘆 𝗶𝗺𝗽𝗿𝗼𝘃𝗲 𝗮𝗰𝗰𝘂𝗿𝗮𝗰𝘆? 𝗔𝗻𝗱 𝗶𝗳 𝗻𝗼𝘁, 𝘄𝗵𝗮𝘁 𝗱𝗼𝗲𝘀?

[zxp @zxp567]

https://t.co/gpa3Ew8ijv

[zxp @zxp567]

“Advanced RAG Techniques: an Illustrated Overview” by IVAN ILIN
https://t.co/yAp4KeyyJe

[Chomba Bupe @ChombaBupe]

Link to paper: https://t.co/XUTtDCqFap

[zxp @zxp567]

https://t.co/8twm6wSwC9

[zxp @zxp567]

@omarsar0 How do they handle knowledge cutoff, given the timeliness of financial information is critical?

[zxp @zxp567]

How Uber Optimizes the Timing of Push Notifications using ML and Linear Programming https://t.co/DGIJdNweo5

[zxp @zxp567]

“Building Airbnb Categories with ML & Human in the Loop” by Mihajlo Grbovic
https://t.co/kfblx7zvtf

[zxp @zxp567]

@svpino Well, they’re not comparable at all. Except for similar technical flows, the generated responses are far less helpful and accurate than ChatGPT (GPT-3.5). Try it first yourself before claiming a “clone” of ChatGPT.

[zxp @zxp567]

@omarsar0 Tried this question “list some approaches of offline reinforcement learning and limitations of each”, ColossalChat’s answer isn’t even close to what ChatGPT provides, and in fact can’t even correct after further prompts.. Stop saying cloning ChatGPT! Not even close!

[zxp @zxp567]

asking for area to improve or potential issues,
and then learning possibly another new approach, implementing it again!

It’s so much fun and so efficient to learn with GPT-4!

[zxp @zxp567]

How my experience of using ChatGPT looks like these days:
firing up a GPT-4 instance,
asking for ideas how to solve a problem, asking for an explanation of a concept I am not familiar with,
verifying some thoughts of steps of applying it to solve a problem, …

[zxp @zxp567]

asking for code to generate a simulated dataset and prototype to implement the approach,
diving deeper to verify the details or just asking it to explain,
copying the code and testing run it if possible,
pasting error message back to ask for debug help, …

[zxp @zxp567]

@ylecun How different is this framework of world prediction from model-based RL? Also would like to hear view from @AndrewYNg @svlevine on this framework.

[zxp @zxp567]

@AndrewYNg Research brainstorming; quickly learning new things with summary and example code generation

[zxp @zxp567]

Why was GPT4 becoming stupid in this question? It’s been like this except the first time it showed some correct reasoning. Disappointing that it seems to stick to this wrong answer @OpenAI

[zxp @zxp567]

Undoubtedly, #GPT-4 is much better than #ChatGPT! It’s phenomenal most of the time as a tool for research brainstorming, drafting plans/docs, and coding buddy!!

Well it still makes shits up sometimes without enough challenge on statements given (so be careful)！

[zxp @zxp567]

https://t.co/ShBDXielYs

[Sergey Levine @svlevine]

Can conservative Q-learning be used to pretrain followed by online finetuning? Turns out that naive offline RL pretraining leads to a "dip" when finetuning online, but we can fix this with a 1-line change! That's the idea in Cal-QL: https://t.co/iZ1TKkaAqi
A thread👇