Daniel Tan Journalist Profile

Articles

In the future, language models will be our interface to the world — LessWrong

Jan 24, 2025 | lesswrong.com | Daniel Tan

Language models are really powerful and will continue to get more powerful. So what does the future of language model usage look like? Here are some things people likely already do. Instead of reading a long piece of text, you might give a book to an LLM and ask it questions about things you want to know. Instead of writing an essay, you might draft some points and ask an LLM to format it into prose. Here's some other things that seem doable in the near future, given current rates of progress.
The entrepreneur’s dilemma: Fundraising or taking a loan?

Nov 29, 2024 | technode.global | Daniel Tan

Fundraising investments or taking a loan is a question, I often see being asked on Facebook groups for entrepreneurs or aspiring entrepreneurs. This article will break down all the pros and cons so that any entrepreneur deciding between the two, may make a more informed decision. Pros:No repayment obligations: Investors provide capital in exchange for equity, so there are no monthly repayments or interest.
In the age of AI, which human skills increasingly stand out?

Nov 15, 2024 | technode.global | Daniel Tan

According to a 2020 report by the World Economic Forum, it is estimated that by 2025, automation and AI could displace 85 million jobs worldwide. Tools like ChatGPT have become so ubiquitous in a professional’s life that in fact, I had a lot of help in crafting this article and saved so much of my time. Take for example the above figures.
Evolutionary prompt optimization for SAE feature visualization — LessWrong

Nov 14, 2024 | lesswrong.com | Daniel Tan |Dmitrii Kharlapenko |Neel Nanda |Arthur Conmy

TLDR:Fluent dreaming for language models is an algorithm based on the GCG method that can reliably find plain-text readable prompts for LLMs that maximize certain logits or residual stream directions by using gradients and genetic algorithms. Authors showed its use for visualizing MLP neurons. We show this method can also help interpret SAE features.
How to build a tech startup without a CTO – From a founder without a CTO & no programming skills | e27

Aug 13, 2024 | e27.co | Daniel Tan

From fear of failure, fear of losing a stable income, work-life balance concerns and more, many factors can hold someone back from starting their own business. However, one recurring reason I noticed was the lack of CTO or programming skills holding those with a business idea back. This is probably why so many Facebook groups and community events exist to help people meet their co-founders, and even incubators like Antler have popped up to help with the problem.
The Residual Expansion: A Framework for thinking about Transformer Circuits — LessWrong

Aug 2, 2024 | lesswrong.com | Daniel Tan

This is an informal note describing my current approach for thinking about transformer circuits. I've not spent a lot of time thinking deeply about this but I believe the overall claims here are correct. Note: A lot of the high-level ideas I include here are not really original, but I haven't seen the specific framing here applied to transformers, and I would like more people to think about this / tell me whether this is obviously flawed in some way.
An Interpretability Illusion from Population Statistics in Causal Analysis — LessWrong

Jul 29, 2024 | lesswrong.com | Daniel Tan

This is an informal note on an interpretability illusion I've personally encountered, twice, in two different settingsLet's say we have a function f, (e.g. our model), an intervention I, (e.g. causally ablating one component), a downstream metric M, (e.g. the logit difference), and a dataset D (e.g. the IOI dataset). We have a hypothesis that the intervention I will affect the metric in some predictable way; e.g. that it causes m to increase.
Mech Interp Lacks Good Paradigms — LessWrong

Jul 16, 2024 | lesswrong.com | Daniel Tan

Note: I wrote this post rather quickly as an exercise in sharing rough / unpolished thoughts. I am also not an expert on some of the things I've written about. If you spot mistakes or would like to point out missed work / perspectives, please feel free! Note 2: I originally sent this link to some people for feedback, but I'm having trouble viewing the comments on the draft. The post is also in a reasonably complete state, so I've decided to just publish it.
A startup founder lives on the ‘Edge Of Tomorrow’

Jun 30, 2024 | e27.co | Daniel Tan

‘Edge of Tomorrow’, released in 2014, is a science fiction action film starring Tom Cruise and Emily Blunt. The movie is set in a future where Earth is under attack by an alien race known as Mimics, where William Cage (Tom Cruise) finds himself caught in a time loop, reliving the same brutal battle each day.
Activation Pattern SVD: A proposal for SAE Interpretability — LessWrong

Jun 28, 2024 | lesswrong.com | Daniel Tan

Epistemic status: This is a rough-draft write-up about a thought experiment I did. Reasonably confident about the broad arguments being made here. That said, I haven't spent a lot of time rigorously polishing or reviewing my writing, so minor inaccuracies may be present. When interpreting an SAE feature, one common technique is to look at the max-activating examples, and try to extrapolate a pattern from there.