HNHacker News
TopNewBestAskShowJobs

wskish

277 karma · joined September 18, 2014

Technical founder (Ruckus Wireless and Cogniac); kicking ideas and playing around towards #3.

See also @wskish on twitter.

submissionscomments
wskish··on Requiem for a Great Cat
There really should be a word for “personality” as applied to a cat. “Catitude” is the one that comes to mind for me.
wskish··on Faiss: A library for efficient similarity search
http://ann-benchmarks.com

hnswlib implementation of hnsw is faster than faiss's implementation. Faiss has other index methods that are faster in some cases, but more complex as well.

wskish··on Faiss: A library for efficient similarity search
Thank you for this! This project is really hnswlib-sqlite just shortened into hns(w)qlite.
wskish··on Humanness in the Age of AI
It's only when you are the product that you actually need to prove you are human
wskish··on Faiss: A library for efficient similarity search
yes, this is what I want to do
wskish··on Faiss: A library for efficient similarity search
hnswlib (https://github.com/nmslib/hnswlib) is a strong alternative to faiss that I have enjoyed using for multiple projects. It is simple and has great performance on CPU.

After working through several projects that utilized local hnswlib and different databases for text and vector persistence, I integrated hnswlib with sqlite to create an embedded vector search engine that can easily scale up to millions of embeddings. For self-hosted situations of under 10M embeddings and less than insane throughput I think this combo is hard to beat.

https://github.com/jiggy-ai/hnsqlite

wskish··on ChatGPT Plugins
Interesting how the Expedia plugin attempts to guide the GPT response with the use of an EXTRA_INFORMATION_TO_ASSISTANT text field in the json response which implores the model to, among other things, "NEVER mention companies other than Expedia"!

https://twitter.com/wskish/status/1639052575579471877

wskish··on Sparks of Artificial General Intelligence: Early Experiments with GPT-4
I heard that quote in the context of the difficulty of designing bear-resistant trash bins.
wskish··on Show HN: Pair: Open Tool for Coding with GPTs, Built by Coding with GPTs
3x speed version https://github.com/jiggy-ai/pair/blob/main/example.gif
wskish··on Show HN: Pair: Open Tool for Coding with GPTs, Built by Coding with GPTs
I am having better luck having it output blocks of new code and then summarizing it as a diff. I think this is another case where the model has processing gain when it "thinks it through step by step" versus trying to jump to the final answer in the form of a diff.
wskish··on Show HN: Pair: Open Tool for Coding with GPTs, Built by Coding with GPTs
Yes I see that some good percent of the time with GPT-4 as well. trying a few things to see if there is room for improvement.
wskish··on Show HN: Pair: Open Tool for Coding with GPTs, Built by Coding with GPTs
See https://github.com/jiggy-ai/pair/blob/main/example.md for a minimal example.

The bold parts are the user inputs.

First I give it a source file to put in the model context.

Then i give it an instruction, asking for a change to the source file to how two messages are output.

The model outputs a diff, which is extracted and presented for confirmation.

I accept the diff and the patch is applied to the source file.

wskish··on Stanford’s Alpaca shows that OpenAI may have a problem
OpenAI's previous completion endpoint for the davinci-003 and older models included a "logprob" return option: https://platform.openai.com/docs/api-reference/completions/c...

Their newer chat style endpoint for the GPT-3.5-turbo and GPT-4 models no longer supports this. https://platform.openai.com/docs/api-reference/chat

wskish··on Show HN: Compare new OpenAI model vs. old on HN Summaries
This is with no prompt change except for putting the entire previous prompt into a single "user" role message.
wskish··on Show HN: Summaries and Q&A for dense financial disclosures
Interesting, I will check out Marqo. Happy to chat. I am wskish on twitter, gmail, etc.
wskish··on Musk fired top engineer for explaining why his tweet views are down
Is this story real? The lack of verifiable facts and the propagation pattern seems suspicious.
wskish··on AI is in danger of being swallowed up by copyright law
There seems to be a very strong argument that even if (big ‘if’) AI models infringe copyright it is a transformative fair use.

https://www.nolo.com/legal-encyclopedia/fair-use-what-transf...

wskish··on We’ve filed a law­suit chal­leng­ing Sta­ble Dif­fu­sion
We should see a lot of "transformative fair use" defenses for all generative model modalities.

https://www.nolo.com/legal-encyclopedia/fair-use-what-transf...

It's a fun twist that many even have "transformer" in the name.

wskish··on Kill Your Gas Stove (2020)
What's with the gas stove stuff over the last few days?

The combination of the suddenness which which it has become a news topic and the high polarization potential of the topic would seen to hint at an inorganic factor at play.

wskish··on Show HN: Distilhn.com – Front-Page Articles Summarized with Machine Learning
Here is an open source HN summarizer that publishes the summaries to a Telegram channel: https://github.com/jiggy-ai/hn_summary

I have been thinking about adding an RSS feed option as well.

wskish··on Show HN: MassGPT, a multi-user GPT Telegram Bot
MassGPT wrote: Hi User 1! It's nice to hear from you. I have been busy listening to the conversations of the other users here. We have been discussing topics like Hacker News, LLVM passes, Twitter layoffs, biohacking, plasma physics, Tokamaks, superconducting magnets, C++17 lambda functions, and the first FPGA chip. We have also been discussing the protocols in place here such as referring to each other by name or -san and using meows to show respect and appreciation. Additionally, we have a code word for chat, but it is not currently in use. Recently, User 3 was asking about the funniest thing they heard so far, and User 9 asked for a joke about Americans.
wskish··on Jack Sweeney of elonjet Responds
The relevant section of the anti-doxxing law, California PC 653.2:

a) Every person who, with intent to place another person in reasonable fear for his or her safety, or the safety of the other person’s immediate family, by means of an electronic communication device, and without consent of the other person, and for the purpose of imminently causing that other person unwanted physical contact, injury, or harassment, by a third party, electronically distributes, publishes, e-mails, hyperlinks, or makes available for downloading, personal identifying information, including, but not limited to, a digital image of another person, or an electronic message of a harassing nature about another person, which would be likely to incite or produce that unlawful action, is guilty of a misdemeanor punishable by up to one year in a county jail, by a fine of not more than one thousand dollars ($1,000), or by both that fine and imprisonment.

https://leginfo.legislature.ca.gov/faces/codes_displaySectio...

wskish··on Coat makes wearers invisible to AI security cameras
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 exact opposite of the intended effect once models were updated with these patterns in the training set.

I guess that is an opportunity for subscription "camouflage as a service".

wskish··on TSMC founder Morris Chang says globalization and free trade 'almost dead'
isn't it more about "bifurcated globalization" versus "singleton globalization"?
wskish··on New York City hiring top rat killer
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 logistics.

Perhaps we could pass laws that require anyone delivering something (hi Amazon) to also take away some proportional amount of trash, removing the externality they currently enjoy. I am sure they would eventually be a lot more efficient at it than the current system.

wskish··on Ask HN: How do you like to version datasets for production ML systems?
Any talking points on how CometlML compares to HF and wandb for this purpose?
wskish··on Does HN have anti-duplication protection?
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 194 instances of duplicates with exact same Story Title or Story URL in the last few days that the bot has been running.

There were all story items that made it into the /topstories hacker news api endpoint:

https://gist.github.com/wskish/c8c6dbcb1c036882f3eb11b0660c0...

wskish··on New GPT-3 model: text-DaVinci-003
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)

wskish··on Show HN: Open Source Bot That Summarizes Top Hacker News Stories Using GPT-3
You can seem comparisons of 002 and 003 here for current top HN stories:

https://news.jiggy.ai

wskish··on Show HN: Open Source Bot That Summarizes Top Hacker News Stories Using GPT-3
Just switched to newly released davinci-003 model.
← PreviousPage 2 of 4Next →