ttul

55 ms·

871.
by ttul 3y ago
In the negotiating style of Chris Voss, help me negotiate a salary increase. Etc
872.
by ttul 3y ago
Pro tip: GPT-4 was trained on this book and can implement Chris Voss’ negotiating strategies quite effectively.
873.
by ttul 3y ago
I hope a merger between Uber and Lyft is blocked by regulators. It would doubtless lead to higher fees and lower regard for ride share drivers.
874.
by ttul 3y ago
Yes indeed. You are essentially using deterministic code to oversee a probabilistic model. Indeed, if you aren’t doing this, your new LLM-dependent application is already susceptible to prompt injection attacks and it’s only a matter of tim
875.
by ttul 3y ago
There is another problem. If you have to use a second model to check the safety of the first model’s output, then you have to wait for the first model to finish generating before the second model can do its work. Uber-scale LLMs are slow at
876.
by ttul 3y ago
Great. We just need to find a way to vent that extra heat off into outer space. Problem solved. /s
877.
by ttul 3y ago
And my main worry is: are they just going to cancel the new thing that my company invested six months and $250,000 of engineering time integrating with…
878.
by ttul 3y ago
Side note: In the 1990s, everyone in my engineering school had an HP48 calculator. There were a healthy selection of pretty decent games available. One fine day, I finished my physics exam an hour early and so opened up an enjoyable game on
879.
by ttul 3y ago
They did mention that the inference time to answer a question about the book was something like 22 seconds, so perhaps they are indeed still using self-attention.
880.
by ttul 3y ago
Yes. For instance, a large context window allows you to have a chat for months where the model can remember and make use of everything you’ve ever talked about. That enables creating a much more effective “assistant” that can remember key d
881.
by ttul 3y ago
Richard Clayton is a legend in the anti-abuse world. I’m not at liberty to say much in detail, other than to acknowledge that he’s a bit like Gandalf, swooping in whenever the crew needs his help. Without his dedicated efforts, the world wo
882.
by ttul 3y ago
If you feel this way, I recommend getting an ADHD assessment by a skilled and qualified professional. Everyone probably wants to be more conscientious from time to time, but if you sense you are unable to be more conscientious even though y
883.
Google's patent for decoder-only transformers(patents.google.com)
2 points by ttul 3y ago|2 comments
884.
by ttul 3y ago
HN recently discussed the encoder-decoder transformer patent ( https://patents.google.com/patent/US10452978B2/en ). Perhaps it's more important to talk about the decoder-only patent they also filed.
885.
by ttul 3y ago
Yes, but patent US11556786B2 would perhaps give them that ability.
886.
by ttul 3y ago
Patent lingo is necessarily more generic than the lingo used within the machine learning space. For example, within patents, they will talk about using computers to do things, which is obvious outside of the patent world. But when writing a
887.
by ttul 3y ago
Yes, but patent US11556786B2 claims the decoder-only transformer.
888.
by ttul 3y ago
Of course, in patent US11556786B2, Google finally mopped up the decoder-only idea. https://patents.google.com/patent/US11556786B2/en "1. A method of generating an output sequence comprising a plurality of out
889.
by ttul 3y ago
As the lawyer behind the one-click patent once personally explained to me, it's the claims that matter most. And the first claim matters more than all of the other sub-claims. This patent specifically covers ONLY transformers in which
890.
by ttul 3y ago
What do you feel would be a good summary of a similar length? I’m asking genuinely.
891.
by ttul 3y ago
The entire blog post fit into the context window. I did not use GPT-4 to summarize the whole paper. Either way, I found the summary informative after reading the blog post end-to-end in the more human manner anyway. I don't think GPT-4
892.
by ttul 3y ago
As is so often the case in neural network design, preventing numbers from blowing up is the cause of so many necessary, but poorly understood math hacks. It seems these researchers have found a potentially better way to control the sampling
893.
by ttul 3y ago
Reflected Diffusion Models are a new class of diffusion models that respect general boundary constraints and correctly train for thresholded sampling. They bridge the gap between theoretical frameworks and practical implementation details,
894.
by ttul 3y ago
Inflation costs more than bank failures. The Fed can fix bank failures by providing liquidity on demand - without any limit, theoretically. But unless they tighten rates, inflation will just grow and make future tightening more urgent and l
895.
by ttul 3y ago
Or as Hinton suggested today, they will have superior ways of thinking.
896.
by ttul 3y ago
Nah, Hinton is already incredibly rich. His first startup was bought by Google for $44M. And Google paid him millions more for a decade. Dr. Hinton is in a rare position of having no reason to work for anyone, not even venture capitalists.
897.
by ttul 3y ago
Here’s another, perhaps more pressing problem: people will have to prove it WASN’T them saying something in that Instagram post or that YouTube video. It’s one thing for Joe Biden’s team to debunk a deep fake. Quite another for some teenage
898.
by ttul 3y ago
Absolutely this. You can already use GPT-4 to have a convincing text-based conversation with a target. And audiovisual generative AI is fast reaching the uncanny valley. Since there is apparently no way to put the genie back in the bottle,
899.
by ttul 3y ago
More like HR said, “Well, there is option A where you leave and are free to do what you wish. And then there is option B (points at bag of cash) where you pretend none of this ever happened…”
900.
by ttul 3y ago
TLDR it’s been better to focus on computational growth than clever algorithms. That being said, architectures are also important when they can reduce computational complexity by orders of magnitude.

More ›