[ML News] OpenAI's Whisper | Meta Reads Brain Waves | AI Wins Art Fair, Annoys Humans

#mlnews #openai #ai

Everything important going on in the ML world right here!

Sponsor: Paperspace

OUTLINE:
0:00 - Introduction
0:20 - Whisper: Open-Source Speech Transcription
6:30 - Sponsor: Paperspace
9:30 - Meta: How the brain hears audio
11:25 - PyTorch moves to Linux Foundation
12:15 - French Government uses AI to find unlicensed swimming pools
13:35 - AlphaFold extends database
14:10 - John Carmack raises 20M to build AGI0729970510422016
16:10 - Cerebras achieves model size record
17:40 - Andrej Karpathy on YouTube
18:35 - ColabPro changes pricing
19:15 - Huggingface runs evaluation on the hub
20:35 - AI wins art fair
22:50 - PaLI: Multilingual Language-Image Learning
23:40 - Operationalizing Machine Learning: An Interview Study
24:35 - LAION OpenCLIP: New Models
25:10 - BlenderBot 3 175B Released
25:45 - OWL-ViT on the Hub
26:10 - GLM-130B
26:35 - Ernie-ViLG
27:10 - Digitizing Smell using Molecular Maps
28:00 - AlexaTM 20B
29:00 - Audio-LM
29:45 - Useful Things
37:20 - Raycasting in JAX
38:00 - GPT-3 Prompt Injection
39:20 - GPT-3 plus Python
40:45 - Game Emulation via DNN

References here (external bc too long for YT):

Links:

If you want to support me, the best thing to do is to share out the content :)

If you want to support me financially (completely optional and voluntary, but a lot of people have asked for this):
Bitcoin (BTC): bc1q49lsw3q325tr58ygf8sudx2dqfguclvngvy2cq
Ethereum (ETH): 0x7ad3513E3B8f66799f507Aa7874b1B0eBC7F85e2
Litecoin (LTC): LQW2TRyKYetVC8WjFkhpPhtpbDM4Vw7r9m
Monero (XMR): 4ACL8AGrEo5hAir8A9CeVrW8pEauWvnp1WnSDZxW7tziCDLhZAGsgzhRQABDnFy8yuM9fWJDviJPHKRjV4FWt19CJZN9D4n

[YannicKilcher]

OUTLINE:
0:00 - Introduction
0:20 - Whisper: Open-Source Speech Transcription
6:30 - Sponsor: Paperspace
9:30 - Meta: How the brain hears audio
11:25 - PyTorch moves to Linux Foundation
12:15 - French Government uses AI to find unlicensed swimming pools
13:35 - AlphaFold extends database
14:10 - John Carmack raises 20M to build AGI0729970510422016
16:10 - Cerebras achieves model size record
17:40 - Andrej Karpathy on YouTube
18:35 - ColabPro changes pricing
19:15 - Huggingface runs evaluation on the hub
20:35 - AI wins art fair
22:50 - PaLI: Multilingual Language-Image Learning
23:40 - Operationalizing Machine Learning: An Interview Study
24:35 - LAION OpenCLIP: New Models
25:10 - BlenderBot 3 175B Released
25:45 - OWL-ViT on the Hub
26:10 - GLM-130B
26:35 - Ernie-ViLG
27:10 - Digitizing Smell using Molecular Maps
28:00 - AlexaTM 20B
29:00 - Audio-LM
29:45 - Useful Things
37:20 - Raycasting in JAX
38:00 - GPT-3 Prompt Injection
39:20 - GPT-3 plus Python
40:45 - Game Emulation via DNN

[fastfitc]

Whisper is just too good, apparently they are going to use it for collecting data for training gpt 4

[dr.mikeybee]

Good work, Yannic. Yes. This release is a result of the release of free Stable Diffusion but also Bloom.

[BrainSlugs]

Art isn't about money, it's about expressing yourself creatively; no matter how self conscious or creative AI gets, you can still make your own Art.

As for professional art, or selling your work... that's a job. AIs are coming for ALL of our jobs; including the creative ones. And that's awesome. Depending on how things land, it could leave us in a state where all we have left to do is express ourselves creatively (with or without AI helping us).

[kimber]

Correction: Neurologists are divided about whether they can actually READ thought patterns like code written for a computer. In fact that's a hint.... like computer code. Anyway some neurologists are promoting that fMRI can, and other neurologists (James Giordano Georgetown University, from the Pentagon's 25 on Havana Syndrome, a Neurologist as well as part of the board of Ethics) says we can't. The truth is our brain works like a Plasma Ball. It NEVER produces the same signals or neural synapses for a thought. Never. I could visualize hammer 100x and it would be different every single time. What they are doing is using Behavior Analysis Software along with Massive Big Database Collections (NSA, Google, Microsoft, FaceBook, Governmental & private) for a big CON, a Lie. On top of that they are combining the lies with Neurocognitive Augmented Brain Warfare to assist us humans in our "perception" of reality they way they want, not necessarily the way it is or how we would want. This is all part of brain warfare - you've been warned.

[headmetwall]

Been playing around with whisper for the last few days, while it does have bugs appear from time to time, when it works it's really good. The fact I can just run this on my pc and create transcripts as good or at times better than what companies like google have is almost unbelievable.

[GarethDavidson]

I actually laughed out loud at "I'm kidding, it's a transformer"

[avi]

"Meta can read your brainwaves" if knowing exactly how to ad-target you and predicting your behavior weren't creepy enough

[mikenashtech]

Wow amazing ML news. Really loaded with great stuff. Nice work, thank you Yannic.

[Mutual_Information]

35:53 that textbook is Probabilistic Graphical Models.. excellent book 👌

[NeoShameMan]

I notice there is a pattern to Yannic video, there us an intro, a content, a sponsor, a content, and an outro, I find that interesting he too falls into that pattern 🤣

[amoynunso.]

Leave it to a Swiss to not understand the purpose of taxation 🤣
Sorry, but that one was too tempting.

This is not a problem for average people wanting to cool off in the summer. These are people who can afford a house with a large enough property and builders building a pool in it. Maybe they don't even live in the house but just want to increase the property value to sell it more expensive at a time when the south of France had it's worst drought ever. Nobody is keeping rich folks from having a pool but if their investment pays off as a hefty increase of their wealth due to the raised property value, then it gets taxed. It's about having everybody play by democtratic rules and not allowing loopholes for rich people with their pools.

[khatharrmalkavian]

3:27 - If the stats are going all over the place but the average is improving steadily, that just means that it hasn't converged yet. It's a signal that it has plenty of potential to keep improving.

[Theeoldmann]

@27:41 potential for Futurama's smell-o-vision 🤣 nice

[Parisneo]

I like the idea of a game that is fully generated by AI. I can imagine a game fully produced by a stable diffusion model where the model produces images out of a text that changes over time according to the user inputs (pressing arrows and stuf). With temporal coherence, we could have consistant frames and pretty cool original graphics. The twist here is that two people playing the game may end up having completely different graphics which is cool.
AI world is just getting more and more exciting every day !!

[JLeonSarmiento]

This is the only news outlet I care about. Thanks.

[merkwur]

A good challenge; using the tiling of the bottom of the pool to create a one-pixel attack.

[btno]

Like an eeg you need that hack the firmware to replace cerebral reads with infrared so it can go further down to brain stem cerrabellum basil ganglia amigdyla basically main cortex and sub cortex and stem

[texasfossilguy]

I'm ready for us to eliminate governments... speaking of taxing swimming pools UGH disgusting...

[sinity]

Abt building AGI: "It doesn't work" isn't "the worst that can happen". Opposite really.

As for other orgs building AGI, well, they're somewhat more apprehensive of the risks. That's why OpenAI got, uh, closed.