wskish

10 ms·

61.
by wskish 4y ago
Based on my experience with vision models (my previous company has several thousand models in production), basic CNNs are great at detecting camouflage patterns that they have seen before in training. This type of strategy would have the ex
62.
by wskish 4y ago
isn't it more about "bifurcated globalization" versus "singleton globalization"?
63.
by wskish 4y ago
Delivery logistics have advanced dramatically and is trending towards near-realtime. For everything that is delivered, some non-trivial percentage of that needs to be taken away as trash. But there is no incentive to improve trash pickup l
64.
by wskish 4y ago
Any talking points on how CometlML compares to HF and wandb for this purpose?
65.
by wskish 4y ago
I noticed a lot of dups on the HN Summary bot ( https://github.com/jiggy-ai/hn_summary ) so was wondering if we needed an embedding similarity search to filter them. So I checked the database of recent stories and found
66.
Ask HN: How do you like to version datasets for production ML systems?
4 points by wskish 4y ago|2 comments
67.
Show HN: Compare GPT DaVinci-002 and 003 summaries of current top HN stories(news.jiggy.ai)
4 points by wskish 4y ago|1 comments
68.
by wskish 4y ago
You can see comparisons of 002 and 003 here for summaries of current top HN stories: https://news.jiggy.ai (via open source summary bot https://github.com/jiggy-ai/hn_summary )
69.
by wskish 4y ago
You can seem comparisons of 002 and 003 here for current top HN stories: https://news.jiggy.ai
70.
by wskish 4y ago
Just switched to newly released davinci-003 model.
71.
by wskish 4y ago
For anyone interested in seeing summaries without telegram you can now view the summaries for the corresponding front page here: https://news.jiggy.ai
72.
by wskish 4y ago
I have a lot of respect for Tim Bray as a technologist and have followed him closely over the years. I have also been thinking about Paul Graham's "The Four Quadrants of Conformism" a lot lately. I am curious to hear where o
73.
by wskish 4y ago
I tweaked the GPT prompt to include the domain name from the story url as additional context. That fixes this one and a few other problematic ones. Here is a typical summary of this story now: "This is an essay written by Paul Graham
74.
by wskish 4y ago
We prepend the prompt prefix to the page text to create the actual model prompt. GPT doesn't have any such abstraction.
75.
by wskish 4y ago
we are strategically prompting it as such: https://news.ycombinator.com/item?id=33750013 open to all ideas here
76.
by wskish 4y ago
Here is the output of GPT reading and summarizing this (see https://news.ycombinator.com/item?id=33748363 ). Model seems to be hallucinating the author. Any thoughts on how to improve it? "The Need to Read is an essay
77.
by wskish 4y ago
Here are a few recent ouputs from the service: Editing Binaries in DOS (2002) https://news.ycombinator.com/item?id=33737207 This is a technical report on editing binaries in DOS. The author describes how to use the debugger
78.
by wskish 4y ago
haha yeah thats pretty funny. We need to figure out how to derate these models when the have no actual clue. And also how to scrape the source content even better...
79.
by wskish 4y ago
yeah news articles really tend to highlight the "controversial" aspect. Unfortunately they are also the hardest ones to scrape...
80.
by wskish 4y ago
yeah, one of the biggest issues on HN is that folks comment without reading the story article. Perhaps this sort of summary of the story article can help drive a higher quality engagement.
81.
by wskish 4y ago
Definitely, it would be great to figure out how to summarize the discussions and work them into the UX. Anyone please send PRs with ideas around how this might work!
82.
by wskish 4y ago
wow, good stuff; would be super entertaining to try to duplicate that with models
83.
by wskish 4y ago
We use the GPT2 tokenizer (which is apparently the same as the GPT3 tokenizer) to count the number of tokens in the extracted input text and brutally truncate the input text to keep it under the limit. There is a lot of room for more fines
84.
by wskish 4y ago
It is perhaps unnecessary, and I don't have any data that would support that it makes things better, but I am basically guessing that in the training set of {mostly the entire internet} the responses to questions that include "ple
85.
by wskish 4y ago
what was n-gate? I have a pet theory that humor will be the actual litmus test for AGI.
86.
by wskish 4y ago
FYI, here is the current prompt prefix we prepend to the story text before sending to GPT-3. Any suggested alternatives? Will try them based on points here. "Provide a detailed summary of the following web page, including what type o
87.
Show HN: Open Source Bot That Summarizes Top Hacker News Stories Using GPT-3(github.com)
71 points by wskish 4y ago|32 comments
88.
by wskish 4y ago
>> shared suffering cements lifelong bonds. > I'm sorry, is this a cult or a workplace? Why is this a good thing? There are many contexts where some amount of suffering is a prerequisite for achieving great things. For exampl
89.
by wskish 4y ago
can you share the datapoints you are referencing that lead you to this conclusion? My experience over 30 years spent in many startups as well as other domains such as mountaineering where shared suffering cements strong bonds leads me to a
90.
by wskish 4y ago
given the strength of the network effects a subscription model is worth exploring, which seems to be what he is pursuing.

More ›