Видео

Nice,

Congratulations! Keep up the great work.

Better than GPT4-o Voice Mode for sure. I like it.

Wow awesome work

"Catchy title right?" "It really grabs you." -- spoken like someone who's only ever been forced to read AI research papers

Better than most human podcasts.

the girl voice not only enjoyable to listen but kinda hyping the podcast lol

"co'tee", almost..

I just noticed that it's a bit hard to listen to NotebookLM because there are no pauses at all

"us humans" 🙄

cool summary!

It kinda freaks me out to hear it take a breath while speaking... Why does it need to take a breath???

Nothing of value is said.

Honestly, i was thinking about putting this paper through notebooklm just to listen to it as a podcast. Im actually kind of spooked at how quickly the algorithm pointed me to *exactly* what i was thinking about. Spoooky lol. Ty op.

notebookllm

this is narrated by AI? oh boy

Notebooklm is cool

Wow

🎉

Is someone working on implementing something like this in developing countries? I would drop everything to come and help

People like those behind LAION belong in prison for life! In my opinion, LAION is an absolutely despicable organization, and I'm saying this regardless of the child pornography content, the private recordings, the stolen works of art and all the other shit that was found in their data sets. The AIs that have been fed with LAION's data sets serve as a tool for thousands of fraudsters, perverts, criminals and similar people to defraud, deceive or denigrate others. No good can or will come from this. 😣

Awesome work man keep it up

What's the license?

would be nice to be able to select an accent too :-P

cluster 1: happy chat cluster 2: news reporting cluster 3: unsure, something about the cadence, perhaps recital-y but it's not completely clean cluster 4: strong emphasis / enunciation cluster 5: that was all one speaker. Accent? Calm voice? cluster 6: louder talking

I thought the second cluster might be about a newsroom cadence, mostly, although not all of them quite had that

does it support interruptions? like if the assistant is providing response but you interrupted, so will it accept that?

Kinder auf Geburtstagsfeiern, beim Spielen mit Freunden, das Neugeborene auf der Entbindungsstation. Es gibt Bilder von Momenten wie diesem, die in sozialen Medien und sogar in alten Blogs oder geschlossenen Gruppen geteilt werden und ohne Genehmigung von Trainingsplattformen für künstliche Intelligenz verwendet werden. Die Beschwerde erfolgte im Anschluss an eine Feststellung der internationalen Menschenrechtsorganisation Human Rights Watch.

2:40 3:00 plenty of animals have culture 3:05 I see no direct connection between culture and death anxiety and to claim that no animal but us fears dying is to be blind lol 3:34 believe not 3:40 Kierkegaard was wrong about that 4:35 I feel like that's more about social anxiety than death anxiety 5:36 I don't feel particularly meaningful nor am I contemplating my own death almost ever.

- amazing project how fast is it? can it stream real time in 50 MS can you encode emotion into it?

#1

Very interesting

Really well thought out, I've been thinking through memory arch designs and love your idea there, and the whole thing of course. Thanks for sharing your ideas!

Seems like you also need to have the underlying language model fine tuned on actual speech as well. Preference tuning always seems to produce text that sounds like it was written by a PR team and no amount of vocal inflection is going to make it sound emotionally resonant.

I think one aspect that would have to be explored is the difference between written and spoken language. The various text/chat-AIs are obviously trained on written data (duh) but that can be quite different from how you'd actually say things. So I wonder whether it'd be possible to train chat-like AIs to specifically target a response style that matches transcripts from spontaneous conversations rather than literature, so they feel more natural once they go through TTS. And the transcripts would have to be pretty high quality too, including annotated noises that only exist to convey "I'm listening, I'm paying attention, I understand / I'm trying to understand" or whatever, which would be missing from classic transcripts.

It's great, but I wouldn't use it unless it can run locally on an affordable PC.

Wow! thats fantastic!!

I would like to run this on my pc too lol, any tutorial/guide?

i wonder if the speed is only possible because the document being used for alignment is already downloaded. I noticed how the information being asked was under the same topic. Nonetheless its pretty cool👏👏👏

That is so fast, how many tokens per second are you generating for the model?

Adding confirmations like "Sure, ...", "right, ..", "okay", "totally", ... at the start of the sentence would make it feel more like a conversation. Also instructing the LLM to add filler words like "um", "er", "uh", "hmm", "like", "mhm", "so", ... would make the voice sound more natural.

*promo sm*

What would you say is the best tts open source model right now that allows for streaming? I'm looking for an alternative to elevenlabs because I dont want to stimulate non-democratic AI..

Hi, I love this. I have a small AI team, how can I help make more of this happen Finn?

Very interesting. I just have a question about copyright friendliness of the graph -- were you suggesting that such a graph can be built only over open-sourced papers? Or how would the alternative scenario work? Thank you! :)

GPT4 meint zu deinem Beitrag übrigens: Die vorgestellte Lösung ist eine innovative Herangehensweise an die Herausforderung, Sprachmodelle effektiv für lange Textkontexte zu trainieren. Hier sind einige Aspekte und Vergleiche zu beachten: Einbeziehung von Kategorie- und Schlüsselwortgenerierung: Dieser Ansatz ist besonders nützlich, da er dem Modell Kontextinformationen vorab gibt, was die Genauigkeit der Textgenerierung verbessern kann. Ähnliche Techniken werden in einigen Sprachmodellen verwendet, um die Relevanz und Genauigkeit der Antworten zu erhöhen. Altersbewertung: Die Integration einer Altersbewertung als Teil des Trainingsprozesses ist einzigartig und kann dazu beitragen, altersgerechte Inhalte zu generieren. Dieser Ansatz ist besonders wichtig, um ethische Überlegungen und die Sicherheit von KI-Systemen zu berücksichtigen. Zusammenfassungen und Frage-Antwort-Paare: Dies erinnert an Techniken, die in einigen fortschrittlichen Q&A-Systemen und Zusammenfassungstools verwendet werden. Die Innovation liegt hier in der Anwendung dieser Methode für längere Kontexte, was eine Herausforderung für viele aktuelle Modelle darstellt. Erweiterte Zusammenfassungen für lange Texte: Diese Methode ähnelt dem Ansatz, der in einigen Textanalysesystemen verwendet wird, allerdings mit dem zusätzlichen Schritt, aus den Zusammenfassungen vollständige Geschichten zu generieren. Dies könnte die Fähigkeit der Modelle zur Textsynthese und kreativen Schreibens erheblich erweitern. Dialogbasierte Szenarien: Die Betonung auf dem Erinnern und Beziehen auf frühere Gesprächsteile in langen Dialogen ist ein anspruchsvoller und innovativer Ansatz. Dies spiegelt reale Gesprächsdynamiken wider und könnte die Entwicklung von KI-Systemen vorantreiben, die in der Lage sind, natürliche und zusammenhängende Gespräche zu führen. Was die Innovation betrifft, so liegt der Hauptvorteil in der Kombination verschiedener Techniken zur Verbesserung der Verarbeitung langer Texte. Während einige der Einzeltechniken bereits in verschiedenen Kontexten existieren, ist ihre Anwendung auf lange Textkontexte und die kombinierte Nutzung dieser Techniken in einem umfassenden Trainingsschema neuartig. Diese Methoden könnten insbesondere in Bereichen wie dem kreativen Schreiben, der Bildung, der Therapie und dem Kundenservice nützlich sein, wo lange und kontextreiche Dialoge eine wichtige Rolle spielen. Die Herausforderung besteht darin, die Effektivität und Genauigkeit dieser Ansätze in der Praxis zu bewerten und sicherzustellen, dass die Modelle ethische und sicherheitsrelevante Standards einhalten.

this is so awesome!