chris chiu

PhD Student @Cambridge_Uni | M.D. | thinking about Mind, Machine, Medicine

[chris chiu @CHYChiu]

@zoomdong2 100x engineer things

[chris chiu @CHYChiu]

@LiuZiyin10 pretty cool wanna try test this on LLMs tgt

[chris chiu @CHYChiu]

@oneill_c took aws 24 years to implement this and they still made it

[chris chiu @CHYChiu]

after prodding around the playground, I marvel at how nice mech interp for machine intelligence is compared to biological intelligence. it took us 53 years to demonstrate a Jennifer Aniston neuron in human brains, and here we easily (and scalably) did the same in machine minds

[Goodfire @GoodfireAI]

In Neural Networks are Already Clean, @nickcammarata argued that neurons were far more legible than previously thought – and that this was great news for interpretability and safety.

Now, we’re releasing the first technical companion to Nick’s essay. It shares our findings from studying thousands of neurons using our novel automated interpretability pipeline, and outlines why we now think of neurons by their computational function, instead of whether or not they are polysemantic. Here’s what we found: 🧵

[chris chiu @CHYChiu]

@jayden_teoh_ @phillip_isola @SimulatedAnneal how one might think about alien optimizers:

[Khoa Vu @KhoaVuUmn]

The fact that two Huangs are involved in an affair, but only one takes the blame is commonly known as Jensen's Inequality.

[chris chiu @CHYChiu]

@Jacklu_me cool work! wondering how combining multiple deletion masks would affect performance, v.s. applying multiple loras?

[chris chiu @CHYChiu]

@nalidoust @tahoe_ai For these envs that you built, what if the reward for a task is beyond the available experimental results?

[chris chiu @CHYChiu]

@swann_bessa @rosinality great work! just played with it. I note that unlike traditional k-means the cluster never reaches equilibrium point and is jittery (e.g. at epoch 15+). I am assuming this is because of the nature of the algo with ema?

[chris chiu @CHYChiu]

@sarahwooders I think the concept of harness needs to be disentangled between workflow / actions (e.g. stream framing) and knowledge (e.g. all skills.md). The latter can be amortized into model weights while the former should be separate from model

on a diff note, long time!

[chris chiu @CHYChiu]

@benjamin_bolte if you can't tell, does it matter?

[chris chiu @CHYChiu]

Makes sense since Nobert Wiener invented the field of cybernetics and I am a radical transhumanist

https://t.co/ZfGcL1hQko

[chris chiu @CHYChiu]

@sheriyuo I've long contended that when LLM does the writeup + reviews, and then eventually some of it drips down into developing better LLMs, it is essentially RSI (with meat proxies in the loop)

[chris chiu @CHYChiu]

@Chonghan_Liu 2, 2.5

[chris chiu @CHYChiu]

"I learned little or nothing new from this paper"
thanks ICLR reviews

[chris chiu @CHYChiu]

@mathyouf when we all live on computers does it still matter what race we are?

[chris chiu @CHYChiu]

@LuluYZeng Dems bad for economy!!
FR though recorrection probably good + consider it a discount (inb4 bigger crash)

[chris chiu @CHYChiu]

@LuluYZeng /skill ASD-STE100

[chris chiu @CHYChiu]

@LuluYZeng @geoffreyhinton Yah he was giving a talk the other day! Super cool that i can cycle for 10 minutes and listen to the godfather rant about LLMs being too complex

[chris chiu @CHYChiu]

I asked @geoffreyhinton: Can computational intelligence replace biological ones one day?
Him: Yes what’s what I am worried about