Cool stuff. From my understanding, this is less valuable at inference time and more useful when running offline pre-training data prep.
When tokenizing terabytes of text for your training corpus, the speedup here is probably doing real work in saving you time (and money?). You get a faster iteration cycle when figuring out and adjusting your datasets.
I started working on Alki because wrangling and deploying AI models on edge devices felt clunky. Going from a Hugging Face model to something production-ready required too many manual steps. Alki aims to simplify that by offering more conversion options, packaging models into reproducible bundles, and making them easy to publish or run on edge hardware.
Llama.cpp was a great starting point, but it led me down the rabbit hole of building a toolchain that could handle different quantization methods, export formats, and deployment bundles for orchestration.
It’s still early and I’d love feedback from infra engineers and startups running LLMs outside the cloud, especially around packaging formats, performance, and what would make the biggest quality-of-life improvements.
Congrats Recall team, I've been a customer for the past 1.5 years and the entire infra layer behind meetings has let our company really focus on "what makes our beer taste better" instead of having to worry about building universal support for different (and tedious) platforms like Teams and Zoom. Eager to give the desktop recording sdk a try soon.
Funny to think about now with hindsight - but a prototype version of the Magic Leap One.
It was an incredibly novel feeling of seeing the room I was in get mapped out, furniture and all, to see interfaces and 3d models rendered in a way that felt interactable in real space. Moving renders of objects around to see how things could be - well, it was just a lot of fun; it felt like playing with a new video game console as a child.
Sure, the overall experience was clunky, with the motion tracking and sub-par lenses constantly reminding you that it was only a prototype, but that first interaction excites you about all the things that could be possible if it actually is refined into a final product. Perhaps this was the same experience that led many of the early investors to put so much faith in Magic Leap while there was still a long road ahead of them.
The book "Lost Connections" by Johann Hari (which I discovered through HN) has helped me a lot in understanding so much of the context of what makes so many of us feel depression and anxiety at many points in our lives. A big part of the narrative is that we are too quick to diagnose and provide ourselves with medication without really understanding environmental variables that have impacted our overall well-being.
While I do believe our understanding of how the brain functions still has a long way to go, which includes treatments for mental illness, there is something to be said about acknowledging more the external factors of our modern day lives that fill us with so many problems.
Personally, realizing where I have "disconnected" from people, values, and truly human experiences has been a large step forward in improving my mental health. I hope it's something that others who feel so many emotional challenges can try as well to see if they can make progress with their own mental health.
Thanks for sharing your experience. As a 25 year old right in the thick of it, trying to find my own personal balance, I'm curious to know your thoughts some more:
1. What are some of the trade-offs you're still feeling the impact of?
2. How did your "adult stuff" transition play out? Was it something smooth that occurred over a long period of time or a short immediate change?
I'm currently living in Houston, working as a SWE, and can attest to the pros of living here. Regarding your concerns, the city is quite demographically diverse from people of all walks of life. Summers do get hot and humid, but nothing unbearable. The winters are more than tolerable (its currently 68 degrees outside) as well.
If access to clean data was what set Renaissance apart, we'd most likely see other funds/firms do just as well or Renaissance's returns be much lower.
Market data services such as Reuters/Morning Star provide incredibly clean data at great convenience for anything you can imagine, even including astronomical data.
Its kind of like the Black-Scholes model. If it really was that simple and straightforward, everybody would be using it to make a ton of money.
Founded by Jim Simons, the "Father" of quantitative research. The firm is famous for their significant year-over-year returns and notorious for only hiring PhDs from mathematics/physics/computer science.
There is an interesting book on how Jim Simons created the company and built his team of academics from the ground up, "The Man Who Solved the Market"
The generated pages definitely are very structured, with each page belonging to a certain area/subject (ie. "Blockchain for Kids" is then presented as "Not just a Blockchain"). A lot seems to be cut-and-paste together, and not a purist implementation of GANs. With that being said, it still is very impressive that these generic page designs can feel like real world examples of what you might see in a pre-seed startup site. The satire hits right on the mark.
Thats understandable. Where "This Waifu Does Not Exist" serves as a demonstration of GANs, "This Startup Does Not Exist" is more of a satirical jab on how "disruptive" startup endeavors are really generic efforts that don't amount to anything
When tokenizing terabytes of text for your training corpus, the speedup here is probably doing real work in saving you time (and money?). You get a faster iteration cycle when figuring out and adjusting your datasets.