Helin
168 posts
All Epstein files can be found here. They were just uploaded by House Oversight:
drive.google.com/drive/folders/…
English
I used #notebooklm to generated a 35min podcast discussing 4 top stories on hacker news. The flow: pick top stories, stitch their audio overviews together, make smooth transitions between topics, add intro and outro music (generated by @suno_ai_ ). youtu.be/NsOnGt8d9fA?si…
YouTube
English
Saying modern art today is crap, is just lazy. There's amazing new artists, just takes a lot of sifting through noise to find.
Finding classic art that's incredible is *really easy* because a hundred+ years of filtering the best stuff to the top has already happened by museums.
Paul Graham@paulg
The people who think modern art and architecture are crap are often right. But then they undermine their own case by picking terrible examples of "good" old stuff. They can't resist a Victorian pastiche.
English
ideogram.ai is now publicly available and free! come join us and try it out Today!
Ideogram@ideogram_ai
Today, we’re opening Ideogram to everyone on the planet! Sign up at ideogram.ai and have fun! Ideogram enables you to turn your creative ideas into delightful images, in a matter of seconds. It’s free and has no limits, and it can render text! ideogram.ai/publicly-avail…
English
One of the biggest points of feedback I have seen is ensuring that the JSON output does not hallucinate the function parameters. What else?
English
Developers using @OpenAI: how can we improve function calling for you?
Please share your use cases and feature requests!
🧵👇
English
@jimkxa Totally! We need more computing, not “smaller transistors+better instruction set+branch prediction”. 1000x can come from: Once a software is developed, it’s compiled to an ASIC design. Then foundry build the chip and host my software. “Serious software peeps build own hardware.”
English
Moore's law is still not dead!! My guess 1000x to go on known physics. Probably not EUV
English
Little AI is a good development for founders: you can treat massively expensive foundational models like GPT-4 as a prototyping tool, then cheaply train your own models to higher accuracy
If it were all Big AI, all power goes to the biggest and that’s not what I’m hoping for
Ali Ghodsi@alighodsi
We are open sourcing Dolly, a ChatGPT-like model that can do instruction following, created for $30, trained 3 hours on 1 server. The secret in magical human-like interactivity probably lies in a small dataset. databricks.com/blog/2023/03/2…
English
@natfriedman This is great! We had internal scripts to run the comparison but the problem is the time spending on getting access to all the APIs. Now I can run a small test on your tool and decide models of interest. As a result, I just submit an access application to Anthropic. Thank you!
English
@megankao_ Agreed. That’s the future we are enabling with a gpt-powered data analysis platform, to help non-sql-speakers to talk to data.
English
@OfficialLoganK @OpenAI A data security feature that promises prompts and results will not be used for model training.
English
If you are using @OpenAI to build your startup or product, how can we help you? 🧵⬇️
English
@sawyermidddd @petergyang How much time ChatGPT saves you roughly in each category?
English
@petergyang 1) writing SQL scripts
2) writing marketing/sales copy
3) structuring form emails
English
@EdeyJulia I spent 6 years in West Lafayette for my PhD study. You, Zach, and the team make the town magical!
English
Purdue On BTN@PurdueOnBTN
"The team loves that she's here almost as much as I do." @zach_edey's mother, Julia, spends basketball season in West Lafayette to support her son. She also enjoys the chance to cook @BoilerBall the occasional meal. 😋 Catch the season premiere of @BTNJourney at 6 ET 1/11.
QME
@erik_nijkamp @nathanbenaich If LLM API vendors are willing to provide read/customization access to the parameters of the last 2-3 layers of their models, would it be sufficient for RLHF?
English
@nathanbenaich I’m afraid the RLHF opportunity is only viable to players, who own the remaining vertical stack (model, serving, training, etc.), making those even stronger.
English
Here’s a version of the LLM future:
- Few large models maintained and improved by deep pocketed AI cos
- Others use RLHF with their own use case + data to build end products.
If true, there’s an oppy in RLHF SaaS…
…but as an on-ramp to using more LLMs, big cos might do it.
English
@kevinyang @OpenAI Great work! If you fine tune models for each user, do you see that actually gives better personalization compared with in-context learning? Great to see llm is disrupting how we work. I’m experimenting it for analytics, and got amazed by how much it can do in many different ways.
English
OMG it finally works!! 🚀🚀🚀
I got @OpenAI to draft emails 📧 for me in the background 🤯
This christmas I built an email assistant (EmailTriager.com) that automatically drafts email replies behind the scenes
no chrome extension necessary, here's the story 👇
English
I’m pro-Elon when engineering topics are discussed.
Elon Musk@elonmusk
@Rainmaker1973 If you know the solution is a whole number, doing mental cube approximations & guessing works great
English
@_paulshen Natto is a wonderful piece of software that I used to teach my kid programming. Whatever you decide on how much energy to spend on future Natto development, I would appreciate, even it’s 0.
English
i'll keep working and building on natto but i'm going to wander a little more. it's my work that i'm most proud of (natto and maaybe drawbattle.io ✏️). please keep using it and spreading the word! thanks again 🙏
English