shoni.eth

@alexpaden

564 Following

43439 Followers

shoni.eth pfp

Weekly Farcaster AI Research Group Call: open to the public every Tuesday-- no expertise required. 1–2 PM EDT → 11 AM–12 PM MDT, 10–11 AM PDT, 6–7 PM BST, 7–8 PM CEST, 1–2 AM CST/SGT (next day), 2–3 AM JST (next day) Link to join: meet.google.com/aue-jdvu-rgb https://events.xyz/f7c740

1 reply

5 recasts

28 reactions

shoni.eth pfp

0 reply

1 recast

2 reactions

shoni.eth pfp

hi my name's alex and i have table bloat in my database

0 reply

0 recast

3 reactions

shoni.eth pfp

lets pay $100 each to the top 20 walk and talk video makers of June

8 replies

1 recast

30 reactions

shoni.eth pfp

https://nouns.build/dao/base/0x8de71d80ee2c4700bc9d4f8031a2504ca93f7088/vote/35

5 replies

1 recast

8 reactions

shoni.eth pfp

lets spend all the money already jeez

0 reply

1 recast

7 reactions

shoni.eth pfp

I can’t run enough gemma locally my MacBook Pro was only pushing 25 rows per second (25m rows) so I’ll have to use Gemini for backfill then probably fine tune gemma later but at this point I need to switch to the new Neynar tables but should I run media summaries with multimodal gemma I’m not sure that can even do videos and what about websites? Anyway I don’t even know if I’ll use those embeddings anymore or embeddings on the new intent and keyword classifications. What am I doing this for again? Oh right to build a profile dataset.. but wait do I need media embeds from farcaster to get started on that or it’s just a future upgrade.. right right a future upgrade. Okay so remake all the procedures for the new Neynar tables then rerun everything without embeds then extract the new classifications then see what all this data is even for in the first place by making a data profile for each user

0 reply

1 recast

4 reactions

shoni.eth pfp

it's okay, i forgive you - ajp

0 reply

0 recast

11 reactions

shoni.eth pfp

jit ai workflow(s) ?

0 reply

0 recast

1 reaction

shoni.eth pfp

onto v4 (yes that’s mid)

3 replies

0 recast

9 reactions

shoni.eth pfp

making a pfp project in 2025 is crazy

1 reply

0 recast

4 reactions

shoni.eth pfp

added ashwagandha rhodiola to the pill stack and also the pea protein took me like maybe 2 weeks to adjust to fart wise— idk where people find macadamia nut milk but I do love the macadamia nuts from the blueprint shop. Using coconut milk as almond milk was also making me gassy.

4 replies

0 recast

1 reaction

shoni.eth pfp

i bought a don’t die rope chain and low key i think it’s kinda fire

2 replies

0 recast

15 reactions

shoni.eth pfp

crazy how taking an eye off the ball for a second can make a whole week disappear—realizing that kind of sucks, but it’s also invigorating. /ball-is-life

0 reply

0 recast

6 reactions

shoni.eth pfp

I’ll do a video crash course on fine tuning at some point— been using llms for classification a lot lately and they’re usually mid/inconsistent performance out of the box (comparatively)

0 reply

1 recast

6 reactions

shoni.eth pfp

something deeply changed in me over the past year and it’s not finished— a new hope

1 reply

0 recast

7 reactions

shoni.eth pfp

i think the best thing for me is time prioritization and not wasting any more

6 replies

0 recast

19 reactions

shoni.eth pfp

feels like i’m getting old fast— farcaster is in the golden years (imo)

3 replies

0 recast

21 reactions

shoni.eth pfp

was about to start prompt evaling to max returns on small model classifiers but i'm just gonna ship with less right (left) classifications instead. i think a simple eval framework/dash is an easy github nextjs project though. i just want to see how small changes effect results compared to manually labeled ones on a variety of small models i.e. this prompt is 48% right on gemma3:1b this one is 64% right on gemma3:1b this one is 87% right on gemma3:1b:ft

0 reply

0 recast

4 reactions

shoni.eth pfp

does anyone know an eval provider like promptlayer that support small models like gemma3 or anything on hf

2 replies

0 recast

9 reactions