Joseph Chu

I came up with a dumb D&D metaphor for myself. I'm basically a paladin with the stats of a wizard. Except not really because my INT isn't actually that high either, I just have points in lots of random and mostly useless Knowledge skills for some reason.

What does it matter? Despite everything, it seems like I still can't escape the nonsense of this world. Indecisiveness strikes again. Many directions, but nothing promising yet. Who am I to dream? Reality is regression to the mean. What is wisdom? Socrates was right (probably).

The “AI 2027” Scenario: How realistic is it? Everyone’s been talking about it. JD Vance read it. What does it actually tell us?

Also, to be fair, here's a criticism of AI 2027 by actual AI expert Gary Marcus: garymarcus.substack.com/p/the-ai-202...

YouTube video by AI In Context We’re Not Ready For Superintelligence

A very slick, well presented video on AI risk and a particular scenario that's been gamed out by some experts. You may not agree with the AI hype or the exact timeline, but if you think ASI is even in theory possible someday, this is an interesting video to watch.

www.youtube.com/watch?v=5KVD...

Watching the boss make tough decisions and seeing things from the inside make me doubt I could do it myself.

Thanks! Yeah, my research into what creative industries are actually like does not make me hopeful. I also spent the last few years at an indie game dev company that actually made a game that sold 10,000+ copies but that's still struggling to stay afloat.

Thanks!

Thinking again about abandoning the world of jobs and dealing with people to become a full-time novel writer.

While I'm dreaming, I'd also consider being an indie game dev making my own games, but I'd need an artist, as I can't draw for the life of me.

Despite being a former AI research scientist, I've resisted actually using GenAI for my writing and coding. As a former AI person who still follows developments, I know how unreliable it can be, and I also take a certain amount of pride in my ability to perform these skills on my own.

I started off on the outside criticizing the old student government and claiming I could do better. When I got elected, I discovered that juggling all the responsibilities and a full course load while also being sick for a while was absurdly hard and I ended up resigning to try to save my GPA.

It's much easier to judge and criticize than it is to actually work at difficult challenges and face complex tradeoffs and try to actually make things work in a way that satisfies everyone. I learned this in my days in student government back in undergrad.

Are Sparse Autoencoders Useful? A Case Study in Sparse Probing Sparse autoencoders (SAEs) are a popular method for interpreting concepts represented in large language model (LLM) activations. However, there is a lack of evidence regarding the validity of their in...

A paper that claims that linear probes make a strong baseline compared to SAEs:

arxiv.org/abs/2502.166...

Truth is Universal: Robust Detection of Lies in LLMs Large Language Models (LLMs) have revolutionised natural language processing, exhibiting impressive human-like capabilities. In particular, LLMs are capable of "lying", knowingly outputting false stat...

A paper that claims that truth is two directions/vectors instead of one:

arxiv.org/abs/2407.12831

How well do truth probes generalise? — LessWrong Representation engineering (RepEng) has emerged as a promising research avenue for model interpretability and control. Recent papers have proposed me…

A post from Less Wrong that follows on Marks and Tegmark's paper:

www.lesswrong.com/posts/cmicXA...

I'm not sure because I haven't been doing much actual research since early 2024, but last I heard, Neel Nanda claimed that simple linear truth probes were considered a very good baseline compared to things like SAEs and were still an active area of research.

One of the biggest differences between most fiction and real life is that in real life there are many possible reasons for things that happen, and it's almost impossible to know which one is actually the reason. In fiction, it's usually spelled out, but in reality, there's a lot more uncertainty.

This is amazing! I was playing around with the code from that Geometry of Truth from Marks and Tegmark paper a while back where they figured out a "truth vector" existed. Was thinking about building a "mind reader", but now I want to see if I can find a "morality vector" and solve alignment with it.

Life is feeling nervous about the toddler going to daycare.

Take things one step at a time, one day at a time. Focus on what you can reasonably do given the circumstances. Life is a million little moments put together. Cherish what you can. Let go of what you can't. The world will continue to spin regardless of what you do. Live your values. Don't look back.

I'm more of an indoor cat, but the toddler really likes the outdoors, particularly playgrounds, parks, dogs, cars, leaves, and the moon. I've gotten more fresh air just strolling around the neighbourhood almost every day in the last year and a half than probably the prior decade.

In my weaker moments, I wish things could be different. But then I remember that I chose this path for reasons that still hold.

The toddler gets super upset if I don't play Fur Elise correctly and stop too early.

My birthday month and day are two single digits, so the probability of seeing them on a license plate with three numbers at random is about 5% or 1/20. So, I -should- expect to see them about once every twenty cars.

As much as I might want for something to be meaningful, it usually is just chance.

I'm not sure where I'm going with this. I guess I'm just venting my frustrations now. Words that reflect my impasse, my doubts, my worries. It probably isn't wise to post like this, but whatever. It's not like anyone actually reads these anyway.

I don't know. It would be nice if doing the "right" thing were easy. It would be nice if it were clear what that even was. I want to do the right thing. But obligations and responsibilities and everything. It's like, complicated by the ties that bind, the duties we have, etc etc.

It looks awkward to go back as well. I staked a narrative that I left the field in part for ethical reasons. But it's a kind of privilege that even allows me to make that choice. Practical considerations are still there. Daycare is expensive, and I'd be just another cog in the machine.

Though, realistically, I'd still have to pass the interviews and do the freaking leetcode problems that I despise with a passion. The opportunities may not exist for me regardless.

An alternative I wish was sustainable would be AI Safety work, which in theory is much more purposefully good for the world, but which seems hard to get into and also lacks the kind of lucrative stability that working in industry provides.

Still, I've been thinking about going back into the AI field. Game development sucks for compensation and stability, both of which are important for having a family. But in addition to being ethically questionable, AI tech work can be stressful and potentially soul-destroying.

I have mixed feelings about AI. With things like AlphaFold figuring out Protein Folding, it has tremendous potential for good. But the big corpos are mostly using it for evil and also releasing half-baked AI slop models in search of a use case. That's before we get into the existential risk stuff.