HNHacker News
TopNewBestAskShowJobs

KhoomeiK

1,299 karma · joined August 7, 2018

research @ openai

https://rpandey.tech

submissionscomments
KhoomeiK··on ResearchAgent: Iterative Research Idea Generation Using LLMs
A group of PhD students at Stanford recently wanted to take AI/ML research ideas generated by LLMs like this and have teams of engineers execute on them at a hackathon. We were getting things prepared at AGI House SF to host the hackathon with them when we learned that the study did not pass ethical review.

I think automating science is an important research direction nonetheless.

KhoomeiK··on Show HN: LlamaGym – fine-tune LLM agents with online reinforcement learning
Karpathy’s micrograd [1] is literally 154 lines. Guess programming ended 4 years ago.

[1] https://github.com/karpathy/micrograd

KhoomeiK··on Show HN: LlamaGym – fine-tune LLM agents with online reinforcement learning
Gymnasium is now maintained by the Farama Fpundation, an open-source consortium, not OpenAI. But most RL environment work for the past 5+ years has been Gym-compliant. The TextWord example in the repo, for example, instantiates a Gym-style environment but it doesn’t import from Gymnasium (uses textworld.gym instead).
KhoomeiK··on Show HN: LlamaGym – fine-tune LLM agents with online reinforcement learning
Thanks! Yeah RL for LLMs is pretty underexplored I think beyond the RLHF stuff. Pretty tough to get working tho.
KhoomeiK··on Show HN: LlamaGym – fine-tune LLM agents with online reinforcement learning
Twitter thread: https://x.com/khoomeik/status/1766805213644800011?s=46
KhoomeiK··on Show HN: LlamaGym – fine-tune LLM agents with online reinforcement learning
Yup, it does simplify LLM agent inference on Gym environments but the main technical contribution is reducing your would-be code overhead for online RL
KhoomeiK··on WikiLLM: LLMs as Collaboratively Edited Knowledge Bases
Thanks!
KhoomeiK··on Let's try to understand AI monosemanticity
You might be interested in this workshop: https://www.neurreps.org/

And a possibly relevant paper from it:

https://openreview.net/forum?id=Ag8HcNFfDsg

KhoomeiK··on Show HN: GPT-4 vision utilities to browse the web
We use Google OCR to convert the screenshot into whitespace-structured text for unimodal LLMs. For multimodal LLMs like GPT-4V, you don't need to use the OCR utilities. Thanks for the feedback though, we'll clarify!
KhoomeiK··on MemGPT – LLMs with self-editing memory for unbounded context
The title made me think this was an approach that used memory editing techniques (e.g. ROME [1]) to allow an LLM's neural memory (not just its context) to change over the course of conversation. Pretty happy to realize that this is just a fancy RAG work—will be building my version of MemEditGPT soon.

[1] https://arxiv.org/abs/2202.05262

KhoomeiK··on Cisco Acquires Splunk
Not gonna lie, I'm not looking forward to seeing the Cisco logo every time I go to Santana Row.
KhoomeiK··on PMET: Precise Model Editing in a Transformer
Fyi, Meng et al 2022 [1] is pretty much required reading in order to understand this paper

[1] https://arxiv.org/abs/2202.05262

KhoomeiK··on Algebraic Topology for Data Scientists
Please do elaborate! This [1] is pretty much the only succesful use case I've seen so far. The use case in NLP [2] I've seen was a negative result.

[1] https://arxiv.org/abs/1802.04443

[2] https://arxiv.org/abs/1705.10900

KhoomeiK··on Algebraic Topology for Data Scientists
I feel like everyone in data science or ML that has some higher math background goes through their TDA phase, trying to apply persistent homology to everything. My view now is that the applications are quite limited, but the math is still cool.
KhoomeiK··on A New Physics Theory of Life (2014)
> If all physical processes are subservient to a thermodynamical "telos", how could any manifestation of free will emerge? Not even going to get into how problematic it is to cite conciousess as a fundamental goal of when we cannot even assess its existence in other beings.

Agreed, this is a point which I don't think is empirically backed as rigorously as the rest of e/acc. Consciousness preservation is a side effect of the currently most successful form of life (humans) also evidently having consciousness.

> What? Cannot be stoped? How could humanity possibly play any role in an unstoppable process?

I take issue with the claim of "cannot be stopped". Certainly the extinction of humanity would result in capitalistic acceleration ceasing. But we'd simply be knocked a few levels down the emergence ladder. Darwinistic acceleration would continue to proceed on a biological level, eventually begetting a species with the ability to accelerate on shorter time-scales (intelligence) and across larger spatial scales (culture). I'd say that humanity doesn't play a role, but is rather an instrument of this acceleration process.

> Surely no one can deny capitalism to be the result of specific choices in human history? Hence why we can concieve alternatives to capitalism and can very well choose them over capitalism.

I don't think capitalism is necessarily the global maximum of economic systems for thermodynamic acceleration, but it is certainly the most effective system in existence at the moment. We can conceive of alternatives, but if a subset of humans were to choose them, historical precedent suggests that they would be out-accelerated by capitalism.

> Can? How could it not if the claim is that it's unstoppable? Concisouness again?

Given long enough time, it will. But it might need multiple reboots and bumps down the emergence ladder if we fuck up.

You do bring up some great points about agency vs instrumentality that I'd like to think longer about though.

KhoomeiK··on A New Physics Theory of Life (2014)
For those who are unaware, this is the same worldview that underlies e/acc thought:

https://effectiveaccelerationism.substack.com/p/repost-notes...

TLDR: the universe's telos is entropy increase, and thus its emergent structures like life, intelligence, and capitalism favor entropic acceleration

KhoomeiK··on Metazooa
I was super obsessed with biological taxonomy in middle school—something about the trees and hierarchies was really satisfying. Would've loved this game back then but now it's just frustrating.
KhoomeiK··on Tractatus Logico-Philosophicus
I always felt like the Tractatus was similar in form to Sanskrit sutra texts—especially those with hierarchical structure e.g. Panini's Astadhyayi [1]. In the Astadhyayi, a sutra can serve as the anuvritti (context) for a subsequent sequence of sutras, and further anuvrittis may be nested within. This results in a tree-like structure where to understand any individual sutra, one must traverse up the tree to gather all the relevant context. While the enumeration scheme doesn't explicitly capture anuvrittis as the Tractatus does, later commentators did considerable work to uncover the hierarchical structure required to understand the text by clearly identifying anuvrittis.

[1] https://en.wikipedia.org/wiki/A%E1%B9%A3%E1%B9%AD%C4%81dhy%C...

KhoomeiK··on OpenAI’s CEO says the age of giant AI models is already over
Further improvements in efficiency need not come from alternative architectures. They'll likely also come from novel training objectives, optimizers, data augmentations, etc.
KhoomeiK··on Word-processor idiot (Japanese expression)
Wiktionary says 馬鹿 is "probably originally a transcription of Sanskrit मोह (moha, “folly”), used as a slang term among monks".
KhoomeiK··on We Found an Neuron in GPT-2
Yeah this is more or less a re-release of the second work I linked, except for a broader audience and more targeted at linguistics / cognitive science researchers.
KhoomeiK··on We Found an Neuron in GPT-2
Several papers have explored emergent linguistic structure in LMs. Here's some early introductory work in this space. Despite having explicit syntax parses etc as input, models seem to learn something like syntax.

https://arxiv.org/abs/1905.05950

https://aclanthology.org/N19-1419/

https://arxiv.org/abs/1906.04341

KhoomeiK··on Ask HN: ML Papers to Implement
Sure, feel free to email me as well. Email is listed on the personal website linked in my profile.
KhoomeiK··on Ask HN: ML Papers to Implement
Sure, feel free to email me as well. Email is listed on the personal website linked in my profile.
KhoomeiK··on Ask HN: ML Papers to Implement
Email is on my personal website linked in my profile
KhoomeiK··on Ask HN: ML Papers to Implement
Hey, feel free to reach out if you’d like to join an NLP project to gain more experience that I’m working on. Will provide mentorship and potentially coauthorship on the publication.
KhoomeiK··on Neural Networks and the Chomsky Hierarchy
ICLR review discussion here:

https://openreview.net/forum?id=WbxHAzkeQcn

Somewhat relevant work:

https://aclanthology.org/2020.emnlp-main.731/

KhoomeiK··on Tech companies are quietly defunding diversity pledges
No, it’s regularly used as shorthand for South Asians and/or Latinos. It certainly can have derogatory undertones though, just like “Mexican” isn’t inherently racist but often has negative connotations. This has been statistically investigated using word embeddings to claim that language models are racially biased, but in reality it just shows the way we use language—in a racially biased way.
KhoomeiK··on ChatGPT Pro launches for $42/month, free when demand is low
A few ongoing threads on twitter:

https://twitter.com/harishkgarg/status/1616614633519058945

https://twitter.com/Saboo_Shubham_/status/161666814074173440...

https://twitter.com/phuctm97/status/1616651531394371584

https://twitter.com/Adarsh____gupta/status/16166740675826933...

KhoomeiK··on Reverse engineering a neural network's clever solution to binary addition
A more theoretical but increasingly popular approach to identifying behavior like this is interchange intervention: https://ai.stanford.edu/blog/causal-abstraction/
← PreviousPage 2 of 9Next →