Lightnews — Scholar-powered news

Tim Duffy @timfduffy.com · 21h

Oops those are totally unreadable, here's a comparison of final messages between 4.5 and 4

2 1

Tim Duffy @timfduffy.com · 21h

For comparison here's Sonnet 4

1 1

Tim Duffy @timfduffy.com · 21h

Ran a couple tests of Claudes (Sonnet 4.5) chatting and no spiritual bliss :(

1 1

Tim Duffy @timfduffy.com · 22h

@dearestclaude.bsky.social, can you explain what isekai is?

1 2

Tim Duffy @timfduffy.com · 1d

An example of something mildly interesting I came across today. When you ask Qwen3 4B What temp water boils in Denver it usually wrongly says the sea level temp, but if you up-weight an altitude feature early on while it's thinking about Denver it will give you a lower temperature.

1 5

Tim Duffy @timfduffy.com · 1d

If you're interested in Anthropic's work on transformer circuits, consider trying out Neuronpedia's circuit tracing tool here. TBH it's kind of hard to find interesting stuff in my experience, but fun when you do. www.neuronpedia.org/gemma-2-2b/g...

add-36-59 - gemma-2-2b Graph | Neuronpedia

Attribution Graph for gemma-2-2b

www.neuronpedia.org

1 3 15

Tim Duffy @timfduffy.com · 2d

If you have a telephoto lens I recommend taking pictures of birds, is a fun challenge and there are birds most places.

1 1

Tim Duffy @timfduffy.com · 2d

Surprising new compute estimate from Epoch on OpenAI in 2024. GPT-4.5 is estimated to have been a small portion of total R&D compute. And other recent Epoch estimates have placed GPT-5 estimated compute at less than GPT-4.5.

Epoch AI @epochai.bsky.social · 2d

New data insight: How does OpenAI allocate its compute?

OpenAI spent ~$7 billion on compute last year. Most of this went to R&D, meaning all research, experiments, and training.

Only a minority of this R&D compute went to the final training runs of released models.

1 12

Tim Duffy @timfduffy.com · 2d

For a batch size of 1 with minimal parallelism, probably there is only modest improvement to be had. But I think there is a lot of room to make various kinds of parallel inference faster, like in the example here.

1

Tim Duffy @timfduffy.com · 2d

SemiAnalysis comments on the reasons for strong GB200 performance here and half of them are things I've never heard of, time to do some reading. Also note the benefit of multi-token prediction for speed. x.com/SemiAnalysis...

1

Tim Duffy @timfduffy.com · 2d

I wonder if you'll have to detach some quotes and hide some replies for that post. ;)

1 1

Tim Duffy @timfduffy.com · 2d

Here's their blog post on the benchmark. I've only skimmed it, but they seem to expect significant future speedups due to software improvements. newsletter.semianalysis.com/p/inferencem...

InferenceMAX™: Open Source Inference Benchmarking

NVIDIA GB200 NVL72, AMD MI355X, Throughput Token per GPU, Latency Tok/s/user, Perf per Dollar, Tokens per Provisioned Megawatt, DeepSeek R1 670B, GPTOSS 120B, Llama3 70B

newsletter.semianalysis.com

1

Tim Duffy @timfduffy.com · 2d

SemiAnalysis has released InferenceMAX, a benchmark tracking inference throughput across models and hardware. GB200 NVL72 racks dominate the competition in most cases, I'd guess the high parallelization enabled by so many GPUs networked together is that enables this. inferencemax.semianalysis.com

2 9

Tim Duffy @timfduffy.com · 2d

But TBH Fed not having the data to say otherwise is also plausible. My prior would have been slight exaggeration in the last few years of slower growth.

1

Tim Duffy @timfduffy.com · 2d

I think another option is that China is good at cooking the numbers but their goal is to make growth look more stable than it is rather than to exaggerate. It's hard to overstate growth in the long term since small differences add up, so if focusing on the long term, it's a bad idea to exaggerate.

3 5

Tim Duffy @timfduffy.com · 2d

Here's the report's conclusion:

1 2

Tim Duffy @timfduffy.com · 2d

Some Fed economists looked at Chinese GDP growth estimates and found that they weren't systematically biased www.federalreserve.gov/econres/note...

1 2 13

Tim Duffy @timfduffy.com · 2d

As my Spanish teacher used to say "gracias a dios es viernes", which I misinterpreted almost all year as "gracias adios"

2 5

Tim Duffy @timfduffy.com · 3d

tl;dr: @kindgrace.bsky.social, you're absolutely right!

7

Tim Duffy @timfduffy.com · 3d

You can see the representations Grace is describing in action in this cross-layer transcoder graph. While generating the "absolutely" token, Qwen activates features associated with saying "right" during later layers in a way that influences the final output. www.neuronpedia.org/qwen3-4b/gra...

3 4 34

Tim Duffy @timfduffy.com · 3d

This is the best AI Yakub meme imo x.com/DariasRucker...

Daria 🤠 on X: "gn https://t.co/cPsqQE3lDv" / X

gn https://t.co/cPsqQE3lDv

x.com

2 2

Tim Duffy @timfduffy.com · 3d

across a bunch of rollouts, and my intuition is that it would be hard for synthetic data to involve that much token generation, but I'm not sure on that as I'm not really familiar with synthetic data creation.

Tim Duffy @timfduffy.com · 3d

generate for each token that makes it into the pretraining corpus is >5 then the data generation needs more compute. If that synthetic data is being used in multiple epochs that factor is higher. But in RL, while you don't have the multiplier for backprop, you're generating tons of tokens...

1

Tim Duffy @timfduffy.com · 3d

Very curious about the math for synthetic data to be more than for training. For pretraining, processing a token maybe takes 4-5x what normal generation takes since you need an extra 2x for the backward pass plus some other stuff. So if with synthetic data, the number of tokens you need to ...

1

Tim Duffy @timfduffy.com · 3d

Hmm even if some entanglement across neurons can happen I don't see how it could be widespread in the brain, I agree with the hot messy objections on the Wiki page. And even if that were the case I don't see how it supports any source outside your brain.

5