sprobertson

12 ms·

91.
by sprobertson 3y ago
Ah good point, I have "Please do not use web search unless asked." as part of my system prompt.
92.
by sprobertson 3y ago
England/London is a bad example, but this one (based on the paper) is easy to replicate with GPT4 and other top models: > Who is Mary Lee Pfeiffer's son? >> Mary Lee Pfieffer's son is Sean Pfieffer. > Who is Tom C
93.
by sprobertson 3y ago
> one of the more pedestrian friendly cities in the country, San Francisco Sarcasm?
94.
by sprobertson 3y ago
Sorry wrote that too hastily - directly searching a 10Tb dataset of the internet for open source models
95.
by sprobertson 3y ago
They mention that they are Google searching for closed source models, and directly searching the internet for open source models.
96.
by sprobertson 3y ago
The links I posted are from the original Twitter guy but a simplified version he made to open source it
97.
by sprobertson 3y ago
Might as well link to the original / "stolen from" author's equivalent demo with Three.js (open this in multiple overlapping windows) - https://bgstaal.github.io/multipleWindow3dScene/ - and repo -
98.
by sprobertson 3y ago
I'm intrigued but your site gives me little to go on, I feel like I'm missing a big "what this is" page
99.
by sprobertson 3y ago
I think "was banned" is correct according to that link - > future events ... are referred to in the past tense as if they had already happened Lots of wordy theologist writing elsewhere in the article but it sounds like in prac
100.
by sprobertson 3y ago
It's learned, not programmed. It would take more effort to remove every breath.
101.
by sprobertson 3y ago
A more blatant example: it will start responding in French if you start speaking to it in French. These things are pattern machines and it makes sense that it will pick up on the lingo they fed (just like we will)
102.
by sprobertson 3y ago
It'd be nice if there was some sense of scale between the points, looks like everything is equally spaced.
103.
by sprobertson 3y ago
P|Q != Q|P
104.
by sprobertson 3y ago
Your anecdata isn't in the evaluation set - this is summarization
105.
by sprobertson 3y ago
Nice, working for me now
106.
by sprobertson 3y ago
Most of your demo pages are returning 500s
107.
by sprobertson 3y ago
Is there any way to debug requests?
108.
by sprobertson 3y ago
What's the use case?
109.
by sprobertson 3y ago
Eating a bunch of the berries can make you nauseous but the poisonousness is widely overstated.
110.
by sprobertson 3y ago
Put something in your system prompt along the lines of "don't waffle"
111.
by sprobertson 3y ago
Have you tried MemGPT or RAG in general? I'm not convinced long context alone will solve the issue (at least not very soon), some kind of retrieval and especially self-retrieval is good.
112.
by sprobertson 3y ago
I skimmed the first 2 chapters and didn't see "YouTube" but it does seem like his style to mention a bunch of brands. He went as far as calling a key a "Yale key". The last page of chapter two says "Ralph Laure
113.
by sprobertson 3y ago
On the other end of the price/quality spectrum, I have Hues that are still going strong after 6+ years of daily use. You can use them hub-free in HA using a ZigBee dongle.
114.
by sprobertson 3y ago
I see what you mean - not sure about the usage of ambiguous but some people do have an overly narrow definition in mind
115.
by sprobertson 3y ago
> all it really means is that you put some extra information in your prompt (from somewhere) that didn't come from the user Yes, and if you are getting (retrieving) some extra information to add to (augment) your prompt for further
116.
by sprobertson 3y ago
Because "RAG" is shorter than "use a vector db/embeddings coupled with LLM to generate better answers" But better answer: "retrieval augmented" has been used in papers since at least 2018 [1], the full &qu
117.
by sprobertson 3y ago
They could have used a better metaphor for the headline. Hanging up their helmet?
118.
by sprobertson 3y ago
Funny enough I once built a metallic USB volume knob in a box with multiple modes, and one mode was as a controller for a WiFi bulb (along with other modes like screen brightness and of course volume) https://iams.pro/projec
119.
by sprobertson 3y ago
I don't mind the menu bar approach but another way would be something that can be triggered by CLI or a launcher like Alfred. ("Typed" with the menu bar app, by the way.)
120.
by sprobertson 3y ago
Enjoyed the [flagged] + dang response easter egg

More ›