We are a way for the cosmos to know itself.

Joined April 2013
15 Photos and videos
I would like to share some work we've been doing at cascadetech.ai: Predicted Outputs in vLLM. If you aren't familiar with PO, it allows you to dramatically speed up generation when you know something about the contents of the output (think: code modification).
3
1
8
594
@OfficialLoganK I think a lot of people are sleeping on the potential of Predicted Outputs to optimize coding agent workflows, even OpenAI, the first implementers (although with a broken implementation). Shouldn't this be a feature in Google AI Studio?
1
86
It's rare to have such an easy win.
61
The change is in our experimental vLLM fork vLLMx for the moment, but we will be submitting a PR to vLLM main shortly.
1
78
I don't do social media so nobody is going to read this, so I'll just @ some of my favorite LLM twitter accounts begging for some retweets. @swyx @teknium @maharshii @Tim_Dettmers
1
83
Replying to @maximelabonne
@maximelabonne Curious if you have looked at all into using SAE's to suppress induction head activation and reduce the "repetition curse" in (small) LLMs?
1
41
I have a theory that the strong repetition tendencies in small models comes from the fact that the datasets used to train them are trying to maximize coding benchmarks which naturally require strong repetition tendencies.
1
37
I'm going to try the technique from the paper but it seems a little bit of a blunt instrument to me, and it seems like SAEs might be a better tool for this.
28
Carl A. Sagan retweeted
25 Feb 2025
Guess what... Old Friend is finally on Quest! meta.com/en-gb/experiences/o… Do you have a Quest? Perhaps you'd like to download and enjoy it! Or tell your friends who have a Quest. Or just bask in the knowledge that it's there in case you want it.

1
1
13
283
I find it interesting that everyone rushed to corporate chatbots as the most practical near term use for LLMs, when actually the most practical use case is allowing people to interact with their software in English rather than pokes and clicks.
While GenAI is fun, I think its economic value is grossly over estimated, because it’s unreliable, risky and expensive to make and serve. It’s fine for creative tasks, but not (yet) autonomous agents
1
152
I should never again have to scroll through dozens of cryptic menus to figure out how to change some setting in my word processor or operating system if I don't already know the hotkey or menu location.
1
77
Google and Microsoft are correctly honing in on this, but there is so much more that can be done by every other company, and none of it requires ANY new LLM technology.
66
Carl A. Sagan retweeted
There's a lot of speculation about whether OpenAI's video generation model Sora has a 'physics engine' (bolstered by OAI's own claims about 'world simulation'). Like the debate about world models in LLMs, this question is both genuinely interesting and somewhat ill-defined. 🧵1/
27
150
895
617,431
Carl A. Sagan retweeted
When you talk to young folks, they think that the belief in imminent AGI was caused by the rise of LLMs. In reality, this belief is axiomatic and long predates LLMs. DeepMind, Vicarious were founded in 2010 based on this belief. OpenAI in 2015. Interest in LLMs only began in 2019. In 2015, when deep learning was in its infancy and LLMs were years away, many folks were just as convinced that AGI was around the corner as they are today. The only thing that changed is that people are now anchoring these beliefs on LLMs, whereas in 2015 they were looking to Deep RL, LSTMs, and Neural Turing Machines.
81
217
1,494
378,004
Carl A. Sagan retweeted
12 Dec 2023
I asked a google engineer what it would take to build GPT-4 today. I will never forget his answer: "We can’t, we don’t know how to do it."
110
132
2,727
852,295
I am totally heartbroken that I have sunk so much time into a product like #Unity3d that is so incredibly hostile to their developers. Like everyone else with sense, it's time to move on to a new platform. Goodbye Unity.
4
228