turnsout

52 ms·

691.
by turnsout 3y ago
Out of curiosity, what leads you to train models from the ground up rather than fine tuning existing models?
692.
by turnsout 3y ago
This is the frontier—tiny, specialized models like this and ReALM [0], coupled to the application logic and able to run on-device. Eventually devices will be powerful enough to run more general purpose models locally, but for high-frequency
693.
by turnsout 3y ago
Whoa, Bark got a major update recently. Thanks for the link as a reminder to check in on that project!
694.
by turnsout 3y ago
Oh absolutely—there was a meeting of the minds, and both sides are getting value.
695.
by turnsout 3y ago
Maybe, but there are very real costs associated with running those servers, so it’s not totally imaginary money.
696.
by turnsout 3y ago
Before that, they actually ran film projectors in some planes, using a variety of half-baked schemes.[0] Wild times! [0]: https://en.wikipedia.org/wiki/In-flight_entertainment#Histor...
697.
by turnsout 3y ago
Absolutely the worst scroll jacking I’ve ever seen (just trying to go back or back up to the top on mobile)
698.
by turnsout 3y ago
To the parent: “AI” in 2024 is just using REST APIs. Don’t bother inventing new models, just get up to speed on what the various vendors are good at and the basic techniques (particularly Agents/ReAct and RAG) and you’ve learned enough
699.
by turnsout 3y ago
Pi is nice to talk to because unlike other chatbots, it always, always, always asks a follow-up question. I'd love to see the system prompt. That's what's driving their claim of high engagement—the bot just keeps drilling you
700.
by turnsout 3y ago
My dude, all it takes is for a 27yo Product Manager to fall in love with a 28yo Software Engineer. Or two lawyers. Etc. At least half the couples at my daughter’s school are dual income.
701.
by turnsout 3y ago
At that level I think the only possible answer is generational wealth (must be nice)
702.
by turnsout 3y ago
As much as it pains me to say, people with good salaries can sometimes meet other people with good salaries in their city and fall in love.
703.
by turnsout 3y ago
I guess it's all relative, but when I say "luxury condo" I'm thinking of $800k–$1M new construction with fancy appliances. Honestly not my particular taste. But when I see 30-somethings coming out of these buildings walk
704.
by turnsout 3y ago
The number of luxury condos in big cities has always confused me. I guess the market is dual-income couples with high salaries? Even coming from the software world, I wonder what these people do. Finance?
705.
by turnsout 3y ago
Makes me excited for the next model though!
706.
by turnsout 3y ago
If it makes you feel any better, news articles make almost no impact on a product’s revenue or success.
707.
by turnsout 3y ago
Now imagine that instead of a local restaurant, it was a global conglomerate with restaurants in nearly every country
708.
by turnsout 3y ago
In no way is it a part time job. You analyze the reports when you change where your mail is sent from (adding in Sendgrid or what have you), and once you confirm it’s working, you ignore the reports. If your email stops going through, you’l
709.
by turnsout 3y ago
In 2024, if you can only take cash, you might as well close down until you sort out CC processing.
710.
by turnsout 3y ago
The Pile is pretty old—is this an updated version?
711.
by turnsout 3y ago
Yeah, it's not clear who this is for. If you can handle ownership, this doesn't seem to have many benefits over Rust. If you can't handle ownership, and don't mind a runtime, just use Swift, which seems to be the main in
712.
by turnsout 3y ago
Yeah I mean if you don't notice an extra 3/4 pound on a 3 pound laptop, then more power to you—get the bigger one
713.
by turnsout 3y ago
Could not agree more. As an industry, we're stuck in the "Magic Button" paradigm, where AI does something cool, but only if you tap the Magic Button. AI can obviously generate automatic summaries from meeting transcripts, s
714.
by turnsout 3y ago
This is incredibly good—definitely good enough for videoconferencing. To me this is a more promising path than the classical 3D approach Apple is taking with Personas on the Vision Pro. Just need to make it realtime!
715.
by turnsout 3y ago
Quantized 7B LLMs should work fine on your machine, though maybe you’re talking about speed?
716.
by turnsout 3y ago
They do, but unlike a known license, these terms are custom and non-standard. Which means I would guide my commercial clients away from this particular model.
717.
by turnsout 3y ago
What is the license? I couldn’t find it on the 1P site or Kaggle.
718.
by turnsout 3y ago
Yes, aka a webserver
719.
by turnsout 3y ago
That’s not the installation script, it’s the entire “framework!”
720.
by turnsout 3y ago
I really don’t meant to be a hater, but when your blog post has a 26 item TOC, it might be time to take a brief break from typing to do some editing. Substack really has pioneered the “write-only” blog format.

More ›