We gave ChatGPT access to everything you know.
Every text that has been on your screen.
Every word spoken in a meeting.
(Powered by
@RewindAI
)
Now you can ask questions like:
- How do I know this person?
- What should be our OKRs this quarter?
Work with
@itsalfredw
While building our AI survey tool,
@itsalfredw
and I developed the fastest GPT4 audio pipeline on the planet*🤯
🗣️ Talk to the survey and it talks back!
💯 Try it out for yourself!
📝 Technical breakdown below.
We built an AI form 📋 that asks follow-up questions and makes form-filling fun.
No more feedback like “it’s broken” without any context 💁🏻♂️
🔍 Our AI digs deeper and learns why, so you don’t need to follow up.
Try Merlin 🧙🏼♂️
Breakdown of the latencies:
- 500ms transcription of the last audio chunk
- 500ms OpenAI latency
- 300ms until the first half-sentence is streamed
- 500ms text-to-speech latency
→ 1.8s total pause
Compare that to Pi...
Merlin is an AI-led user interview that asks follow-up questions and analyzes responses
👥 Instead of speaking to a few users, you can talk to everyone – instantly!
📞 Interview over phone, text or web
📊 Extract themes, sentiments, and personas
Merlin is hiring founding engineers! We give $2K to any recommendation that leads to a hire.
Why recommend/join Merlin?
- Our user numbers are growing like crazy!
- The best investors in Silicon Valley back us.
- We pay top-of-market SF salaries, with high equity + cash, and
Very soon LLMs will be able to answer any reasonable request.
What is missing?
Your knowledge. Your context.
With that you can answer questions like:
- How do I know this person?
- What should be our OKRs this quarter?
Limitless uses
@RewindAI
to get everything you’ve seen/heard and uses it to autocomplete any question.
They won 2nd place overall!
Built by
@florian_jue
,
@itsalfredw
(apologies - vid only last min of demo)
Introducing ChatGPT Enterprise: enterprise-grade security, unlimited high-speed GPT-4 access, extended context windows, and much more. We’ll be onboarding as many enterprises as possible over the next few weeks. Learn more:
How we did it:
1. voice recording is streamed and transcribed in real-time
2. as soon as you stop, we send the transcript to GPT4
3. GPT4 streams the response back, token by token
Just got a masterclass in leadership from Nvidia CEO Jensen Huang. Mind blown. Takeaways:
He has 55 direct reports, no 1:1s, just group talks = max transparency.
Org chart is modeled like a neural network, breaking the old pyramid model.
Nvidia bets on new grads to lead huge
4. once there is a somewhat complete sentence, we send the first text chunk to a text-to-speech server
5. server streams synthetic voice back and we play it back as the bytes arrive
Winners 🏆
-
@marckohlbrugge
for marcbot, using Whisper API specifically. I hate telegram but just voice-noting with the bot was an A+ experience
-
@florian_jue
,
@itsalfredw
won the ChatGPT API challenge, they 'hacked' rewindai to build this mega handy assistant (check the
@AviSchiffmann
@itsalfredw
1. use streaming and start playing the sound as soon as the first bytes arrive
2. Elevenlabs has a "optimize_streaming_latency" parameter
3. (Most important) do not wait for the entire text but start sending the first words as soon as they come from GPT
I’m creating a search engine that remembers everything on your screen and lets you ask abstract questions over all your data.
Not only that, I’ve developed AI that re-opens any indexed content in its original app.
Watch Boundless retrieve an Apple note with its virtual cursor:
Did anyone investigate if Falcon 40B has some biases eg regarding political views of the UAE?
If Falcon becomes the new open source standard, this will be crucial to know.
AI is the ultimate disruptor. At
@itsMerlinAI
, they use Azure OpenAI Service to create GPT-4 powered surveys that engage like a human interviewer, while ingesting data securely on Azure.
Head to the blog to learn why Merlin chose Azure as their AI partner
A bit disappointed to see that the Gemini demo was not an actual video but some handpicked few-shot example with 3 images.
Now you just need to learn which 2 cups the hands touch, not really any temporal relation between the frames.
Have been playing around with fine-tuning a llama variant on a large document to get better hypothetical answers for retrieval.
Does not seem to beat one-shot GPT yet but it's a work in progress.
Let me know if you have thoughts on this.
I'm sure GPT *could* perform actions on websites like filling out forms, sending emails, etc.
The reason for the API approach for plugins is probably to have tighter control over which actions it is allowed to do.
Power outage in our office in SF since yesterday evening. Will be out till Saturday.
Almost 3 days without power in the financial district in one of the most advanced cities in the world?!
Limitless uses
@RewindAI
to get everything you’ve seen/heard and uses it to autocomplete any question.
They won 2nd place overall!
Built by
@florian_jue
,
@itsalfredw
(apologies - vid only last min of demo)
"Gemini was created from the ground up to be multimodal, highly efficient at tool and API integrations and built to enable future innovations, like memory and planning."
Will Google make agents work?
#GoogleIO2023
Ozone raised $7.1m in seed funding and is launching its beta today!
Ozone is building the first AI-powered and collaborative video editing platform, designed to supercharge content creation for video creators and their teams.
Try Ozone now at .
@pranavrajpurkar
Me and my co-founder
@itsalfredw
(both from Harvard) are building
@itsMerlinAI
: collecting and analyzing qualitative feedback at scale.
Building our founding engineering team right now:
Merlin is hiring founding engineers! We give $2K to any recommendation that leads to a hire.
Why recommend/join Merlin?
- Our user numbers are growing like crazy!
- The best investors in Silicon Valley back us.
- We pay top-of-market SF salaries, with high equity + cash, and
I never understood why people used VPNs, but when I was back in Europe I set it up so I wouldn't have to see the GDPR cookie banners on every website..
Google is stopping all of its coding competitions including CodeJam and HashCode.
Is this to focus on their core business as competition is heating up?
20 years, more than a million participants and billions of lines of code later, our coding competitions are coming to a close. It's been an honor to learn and enjoy coding with you. Thank you.
Join us 4/15 at 2 p.m. UTC for farewell rounds of competition:
Did you know that there's already an AGI out there? It can (basically):
* Order your coffee at Starbucks
* Perform market analysis
* Find and negotiate a lease
Easter project inspired by
#AutoGPT
&
#BabyAGI
Try Godmode
Bard thinks it's better than GPT-4 at generating creative and engaging text but worse at factual and informative text.
What are your experiences so far?
⚡Introducing Bland Turbo. The world's fastest conversational AI:
- Send or receive up to 500,000+ phone calls simultaneously.
- Responds at human level speed. In anyone's voice.
- Program it to do anything.
Call and talk to Bland Turbo now:
🧵 (1/3)
Has anyone tried out streaming with Azure GPT-4?
Seems like the chunk size is much larger. Anyone knows a way around this?
I want it to have the smallest possible delay so it should send the first chunk immediately and not only after the first sentence.
@contentstrats
@RewindAI
@itsalfredw
Data storage and search are offline. However, we send the most relevant context info to ChatGPT once you trigger the shortcut to generate a response.
from sama YESTERDAY: "4 times now in the history of OpenAI - the most recent time was just in the last couple of weeks - I’ve gotten to be in the room when we push the veil of ignorance back and the frontier of discovery forward”
@karpathy
Makes sense. The SVM finds how your query is unique and then gives results that also have this unique property.
Will try it out on my data sets!
What does the next social media look like?
@_alfredw
,
@SilasAlberti
and me just launched our
@scale_AI
hackathon project:
🪄 ✨
Remix images with your friends
What’s the funniest thing you can create?
: Remix images with your friends ✨
@fjuengermann
,
@SilasAlberti
, and I built a new social network based on generative AI in 5 hours for the
@scale
hackathon.
Upload images and make funny edits with text.
At a glance, these pictures look just like me. But I never posed for them — they were entirely AI-generated, using cutting-edge technology from Google scientists called DreamBooth.
What could this mean for the future of deepfakes? My latest. (1/)
We gave ChatGPT access to everything you know.
Every text that has been on your screen.
Every word spoken in a meeting.
(Powered by
@RewindAI
)
Now you can ask questions like:
- How do I know this person?
- What should be our OKRs this quarter?
Work with
@itsalfredw
the scary thing about the xz backdoor is that it was only discovered because they rushed the release and made mistakes (noticeable slow down and valgrind errors).
the valgrind errors were already fixed in the next release!
while this was already a 2 year effort, we will never
@joshm
@browsercompany
Better ordering of my recent tabs with command+T.
If I want to go back to the second most recently opened tab, command+T does not show it so I have to search through all of the tabs I have open on the left -- impossible!
I want to cycle through tabs by recency.
We gave ChatGPT access to everything you know.
Every text that has been on your screen.
Every word spoken in a meeting.
(Powered by
@RewindAI
)
Now you can ask questions like:
- How do I know this person?
- What should be our OKRs this quarter?
Work with
@itsalfredw
: Remix images with your friends ✨
@fjuengermann
,
@SilasAlberti
, and I built a new social network based on generative AI in 5 hours for the
@scale
hackathon.
Upload images and make funny edits with text.
Unit test LLMs with LLMs
Tracking if your prompt or fine-tuned model is improving can be hard. During a hackathon,
@florian_jue
,
@fekstroem
, and I built “Have you been a good bot?”.
It allows you to ask another LLM to judge the output of your model based on requirements.