causal

54 ms·

961.
by causal 2y ago
I agree with everyone saying this is not technically being fired, but PG's tweet omits the important context of YC partners losing faith in Sam as a leader.
962.
by causal 2y ago
I really want one too. This is the closest thing I could find, but doesn't claim the keystroke level of detail described above: https://www.zdnet.com/article/i-looked-at-all-the-ways-micro...
963.
by causal 2y ago
YES - and they hide their "syncing" spyware in so many places that tracking down all the settings to shut it off is such a chore. Including: - OneDrive - Windows Device Linking - Edge - Bing - Cortana - Copilot - Outlook / 36
964.
by causal 2y ago
I agree they don't sound the same. But, since it's a subjective test, OpenAI was pretty Twitter-foolish to push the "Her" angle after being explicitly rejected by SJ. It's just inviting controversy.
965.
by causal 2y ago
The ability to go observe this kind of history is one of the wonderful things about OSS
966.
by causal 2y ago
What's up with HN's general tone of hostility toward Wikimedia foundation? It's weird how the more free a service is the more entitled our attitudes seem to become.
967.
by causal 2y ago
Not to mention that if it's really needed, the LLMs this is made for also excel at converting CURL commands to whatever spec pretty trivially. I've used them this way on a few occasions.
968.
by causal 2y ago
Really impressive for the size. The 4.2B vision model apparently outperforms GPT-4V-Turbo on 5 out of 7 vision benchmarks.
969.
by causal 2y ago
Agreed. I'm not advocating for use of asbestos given its tendency to end up in places humans breathe.
970.
by causal 2y ago
Maritime regulation is hard when it operates in the space between states. Asbestos in particular is such a temptation in manufacturing because it's just so useful and cheap. An argument can probably be made that handled properly it sav
971.
by causal 2y ago
It's wild what hoops people jump through to stay on Chrome
972.
by causal 2y ago
So most of my understanding comes from this series, particularly the last two videos: https://www.3blue1brown.com/topics/neural-networks Essentially each token of a text occupies a point in a many-dimensional model tha
973.
by causal 2y ago
The token embedding is what ultimately gets nudged around by the heads though, right? The key vector just relates to the context size, not the token embedding size, afaik.
974.
by causal 2y ago
Are the same embeddings not used internally? I thought they were. Maybe I'm wrong about that. Mistral uses a 1024 dimension embedding for 8K context. I think the point about trying to capture that rich of a context into a smaller numbe
975.
by causal 2y ago
Has it? It's my understanding that GPT-3 was 12,288 and GPT-4 went down to 3072.
976.
by causal 2y ago
Your test is a good one but the point still stands that a novel dataset is the next step to being sure.
977.
by causal 2y ago
Yeah I also tried to get it to complete some limericks from the dataset. Curiously it believed it had heard of the limerick but would then recite a hallucination. So the good news is that the NIAN score might be real, bad news is you can&#x
978.
by causal 2y ago
Thanks for responding to that point - it's the one most on my mind. So OpenAI's large embedding model has 3072 dimensions, though in practice far fewer are probably used. Clearly you can't compress 1M tokens down to 3072. Yet
979.
by causal 2y ago
+1 on Haiku being oft overlooked.
980.
by causal 2y ago
1M token context by default is the big feature here IMO, but we need better benchmarks to measure what that really means. My intuition is that as contexts get longer we start hitting the limits of how much comprehension can be embedded in a
981.
by causal 2y ago
Even that just shows me GPT-3.5, don't see 4o in the dropdown. Guessing it's being gradually rolled out to accounts?
982.
by causal 2y ago
Clicking the "Try it on ChatGPT" link just takes me to GPT-4 chat window. Tried again in an incognito tab (supposing my account is the issue) and it just takes me to 3.5 chat. Anyone able to use it?
983.
by causal 2y ago
+1 on being more productive with chat than Copilot. I find copilot mostly to be a useful Google replacement, e.g. if I forget the syntax for a for loop in JavaScript I can just write a comment like `// For loop over X` and it will
984.
by causal 2y ago
LLMs are pretty good at anything that follows a pattern, even a really complex pattern. So unit tests often take a form similar to the n-shot testing we do with LLMs, a series of statements and their answers (or in the case of unit tests, a
985.
by causal 2y ago
Hinton is way off IMO. Amount of examples needed to teach language to an LLM is many orders of magnitude more than humans require. Not to mention power consumption and inelasticity.
986.
by causal 2y ago
Humans aren't able to project an image from their neurons onto a disk like ANNs can, if they could it would also be very fast. That 4 hour estimate includes all the mechanical problems of manipulating paint.
987.
by causal 2y ago
Okay that makes more sense. I was skeptical of the article's claim that the water was just being heat-vaporized into oblivion. That said, more water-sustainable power generation is clearly still a need.
988.
by causal 2y ago
Underrated comment
989.
by causal 2y ago
Cool idea! Curious how it handles auth. Like how do you collaborate on a shared URL without shared credentials
990.
by causal 2y ago
I am in favor of AI companies trying to source material responsibly, and if I'm reading this right OpenAI will actually be compensating the publishes to use their content. So this isn't adspace yet. That said, giving publishers &q

More ›