aichannel

is there any interest in cast summaries like these of new ML papers from big companies? 

Apple’s approach: https://arxiv.org/abs/2412.06329 [2412.06329] Normalizing Flows are Capable Generative Models

Meta’s approach: https://arxiv.org/abs/2210.02747 [2210.02747] Flow Matching for Generative Modeling

𝚐𝔪𝟾𝚡𝚡𝟾

Do it for him, our under appreciated AI hero

Do it for him, our under appreciated AI hero 

https://farcaster.xyz/gm8xx8/0xda629828

not promising you’re sending me a cast with no likes

He was a prolific poster of /ai research and sometimes called out things that would happen weeks/months later, but his academic way of posting didn't get a lot of engagement. So I think he went back to X where much of his crowd is!

ahhh got it, i’m not much interested in the academic stuff i’m trying to ensure i am up to date on:

1. text to speech and speech to text (people pay me to apply this for them)
2. image labeling and classification (same as above)
3. time series prediction + anomaly detection (i love projecting revenue)

Dia is great for text to soeech small open model

planting causal loops; harvesting reflexivity

TimesFM interesting transformers for time series

i haven’t found any transformers or seq2seq model that works better for time series than math based models

but you can pass many types of data for a date/tick whatever

so it could extrapolate both from numbers and sentiment analysis or whatever data u feed it per day

idk its all stories we tell ourselves at the end of the day

brother even if it was 10x better i can’t go to a stakeholder and say “the magic thing says we will hit 1.2 next year” — they’re way more impressed if i go in there saying “if the fall sale and back to school season perform similarly to last time and traffic grows 7%, we can hit 1.4 next year”

so dont think how can beat math (it cant)

think ehat can it do that math cant

yeah i have 0 real world experience on timeseries so i believe u

i have never found a complicated model or an ensemble or etc to perform better than identifying what the most recurring sales spikes are and then predicting out based on the growth of those YoY, like huge drastic performance difference