Frederic Barthelemy / is on Nostr

Software Engineer, Tech Enthusiast, aspiring Entrepreneur, AR Dreamer

Last Notes

2025-01-05 00:59:54

- reply

@nprofile…dhpg @nprofile…q54u @nprofile…mhl7 @nprofile…3zdw @nprofile…0af3
Perhaps we’re getting too philosophical 😉. As a dev, I could agree that the “source code” is the model.
But if you cannot run the source code (without the weights, without the proprietary training data) & get the same final result (that consumers are paying for), then consumers are getting a combined product, and they/others currently call *that* the Model.
If weights are compiled input data, input data needs to be equally open source… right?

2025-01-04 20:56:46

- reply

@nprofile…dhpg @nprofile…q54u @nprofile…mhl7 @nprofile…3zdw @nprofile…0af3 Doesn’t that mean that at the minimum this is proprietary-data-required?
Perhaps that’s my misunderstanding: Does the Model encompass *just the tools and algorithms and how they come together*? Because if that’s the case, then I’d agree with you.
But also: users are buying/being sold the model *with* the derived-from-proprietary data, and that’s not open-source if you have to buy some of the sources to build the whole assembly. — Literally TiVo, no?

2025-01-04 20:41:23

- reply

@nprofile…dhpg @nprofile…q54u @nprofile…mhl7 @nprofile…3zdw @nprofile…0af3
I'm a surprised that we're quibbling at all. Surely: If a 3rd party actor cannot train an identical model using the code & public datasets (not the weights), then the model isn't "open source” -- *any* secret / non-public datasets (or even copyright datasets like ebooks#.torrent) are not open.
The weights are a derived / compiled artifact. They can slap an OSI approved license on that artifact, but that's not the whole model being Open-source.

2024-12-09 08:22:13

- reply

@nprofile…48vs Can Facebook impersonate me (say in an ad)? Hopefully click-through agreements don’t count as valid consent.

2024-08-19 16:11:37

@npub1rky…0xq2 Feature Request: Relative Dates on older content shouldn’t just always be in “days”.
eg. the relative date on the following linked article is 31k days. Perhaps convert it to years when it’s > 700 days ago?
https://time.com/archive/6757157/france-no-more-tipping/

2024-06-20 00:35:44

- reply

@npub1u24…32pl I know I'm certainly interested!

2024-05-21 04:24:56

- reply

@npub1eee…e209 you should sharpen it into a pocket knife.

2024-05-21 04:24:24

- reply

@npub1eee…e209 I thought Apple had a trade-in envelope too?

2024-05-09 04:49:35

- reply

@npub1src…fms4 you get one layer of tab bars. Other stuff should be modes/dropdowns.

2024-04-23 23:47:55

- reply

@npub1src…fms4 http://markdowncss.github.io/splendor/

2024-03-23 02:19:13

- reply

@npub1u24…32pl I've filed a few feature requests! Happy to contribute implementations if you agree with them!

2024-03-22 20:38:08

- reply

@npub1u24…32pl I'm super interested in this project! Burnt out on package.json and friends after 10 years of it.
If we have questions about the project, how would you recommend one asks? Is Mastodon the right place, or a github issue? Somewhere else?

2024-02-03 08:34:13

- reply

@npub1sy6…5dxr foveated rendering messes with that!