leobg

18 ms·

301.
by leobg 2y ago
You have users? If TTS is your bottleneck, I might be able to help. Email in bio.
302.
by leobg 2y ago
Interesting. I didn’t think you could get anything close to realtime with Wav2Lip.
303.
by leobg 2y ago
Makes me want to clone sheep and cows and chickens and release them in Montana… just in protest for sending an old man who admits his guilt and pays the price to prison.
304.
by leobg 2y ago
This looks awesome. Didn’t seem to hear me, but the video looks great. Can you share what models you are using? You say these are all open models.
305.
by leobg 2y ago
From the AI generated “best of” pages: > Elon Musk's Bold Support for Trump Amid Health Concerns This is after the assassination attempt.
306.
by leobg 2y ago
You worked there at the time?
307.
by leobg 2y ago
Depends on whose viewpoint you read. He was the biggest shareholder and received $176 million when PayPal was sold to eBay. He was there from the beginning. And he wasn’t the marketing guy. Whether you like him or not, that’s all very diffe
308.
by leobg 2y ago
I imagine that means that he observes a lot and thinks a lot before he acts. Which is great. I just wish that he had some higher God than fame and power.
309.
by leobg 2y ago
I’m sure he’d love that comparison. But Elon is responsible for PayPal, Tesla, SpaceX, OpenAI and now Grok. I have a Starlink antenna on my roof because of him, and I take my kids to school safely and quietly in a car that, 15 years ago, so
310.
by leobg 2y ago
Yeah, in my case it’s not about .com domains. Last year they also removed their free included email hosting and asked some ridiculous price to not delete my mailboxes. So it’s more that I’m no longer expecting good things to come from them.
311.
by leobg 2y ago
If training data is the constraint for LLMs, and scraping from web and books is exhausted, what is a new source that you could tap into to leave Meta, Grok, Anthropic behind…?
312.
by leobg 2y ago
Moved all my stuff to porkbun. Gandi was looking great a few years ago. Now they just milk their customers, calculating that many will be too lazy to leave. I was surprised how easy it is these days to move a domain. Hope porkbun won’t play
313.
by leobg 2y ago
Another (simpler?) approach is to also split your chunks into sentences. So you’ll end up with chunk embeddings and sentence embeddings. Now you can do sentence level search. And also distill chunks down to their most relevant sentences at
314.
by leobg 2y ago
These are gorgeous. What a great idea.
315.
by leobg 2y ago
I’m surprised nobody has produced an audio version of this using Stephen’s cloned voice yet.
316.
by leobg 2y ago
Isn’t hallucination just the result of speaking out loud the first possible answer to the question you’ve been asked? A human does not do this. First of all, most questions we have been asked before. We have made mistakes in answering them
317.
by leobg 2y ago
It does make sense. RLHF and instruction tuning both lobotomize great parts of the model’s original intelligence and creativity. It turns a tiger into a kitten, so to speak. So it makes sense that, when you’re using CoT, you’d want the “bra
318.
by leobg 2y ago
Because it may be in another language or aimed at another audience beyond my kid’s reading level.
319.
by leobg 2y ago
Old books are on Gutenberg, archive.org etc. Physical ones, I scan. Cutting the spine is easiest. But today you can also just take pics with your phone. Many retailers also sell EPUB. Which is just HTML. Obviously, that’s all for private co
320.
by leobg 2y ago
I made something like this for my kids: 1. Take a science book. I used one Einstein loved as a kid, in German. But I can also use Asimov in English. Or anything else. We’ll handle language and outdated information on the LLM level. 2. Extra
321.
by leobg 2y ago
Has been on my list for a long time. Would love to build it. Email in profile.
322.
by leobg 2y ago
You may be right. And it’s odd. Everyone is talking about AI and the future and all of that. But really, we live as if there is no tomorrow. Nobody would start a construction project anymore that would pay off a generation later. Compare th
323.
by leobg 2y ago
Curious: How does that make you feel about JD Vance as potential vice president? From what I read (I’m German), he “was there”.
324.
by leobg 2y ago
If your tub is leaking, it would be wiser to focus on fixing the leak before you focus on tapping new sources.
325.
by leobg 2y ago
Relevant ads is functional search, is it not? Ideally, search without the need to query.
326.
by leobg 2y ago
Video compression algorithms use key frames. So can’t you do the same thing? Essentially, generate five seconds. Then pull out the last frame. Use some other AI model to enhance it (upscale, consistency with the original character, etc.). T
327.
by leobg 2y ago
Right. Just let OpenAI do it. Then it’s allowed. And internally, let the “model” be a 7zip compression algo. Just call it an “LLM”. Courts won’t know the difference. Haha
328.
by leobg 2y ago
One more reason why this is B.S.: They can obviously license that content. No rights to publish. Just the right to use the content as part of their training data for the AI.
329.
by leobg 2y ago
People think of the Karpathy tutorials which do indeed build LLMs from the ground up, starting with Python dictionaries.
330.
by leobg 2y ago
What is Pinecone used for? I would assume that an average repo yields only a few hundred or thousand chunks. Even with brite force similarity search that is just 2-digit milliseconds on CPU. Faster than any API call. And even if you got int

More ›