00:00I say this every week, but we really did
00:02have another incredible week of AI news.
00:05Tesla has multiple AI launches. OpenAI is
00:08raking in the cash, but also struggling.
00:11AI sets records at drone racing, beating
00:13the best humans, and
00:14meta launches Code Llama,
00:16which quickly beat GPT-4 at coding tasks.
00:19We have a ton of other
00:20fantastic tech and AI stories today,
00:22so sit back, relax, and enjoy. Let's go!
00:26First, let's talk about Tesla. This week,
00:28Tesla launched a massive
00:29$300 million AI cluster,
00:31including 10,000 NVIDIA H100 Compute
00:34GPUs. By the way, this is
00:36why NVIDIA is crushing earnings
00:39lately. Their new AI cluster will be used
00:41to power several AI
00:42applications, but of course,
00:44its main use will be to continue training
00:46its full self-driving
00:47product. And they have good reason
00:49to have such a massive supercomputer.
00:51According to Tim Zaman, AI infra and AI
00:54platform engineering
00:55manager at Tesla, "Due to real-world
00:57video training, we may
00:59have the largest training
01:00datasets in the world. Hot-tier cache
01:03capability beyond 200
01:05petabytes. Orders of magnitude larger
01:08than LLMs." Also, this week, Elon Musk
01:11showed off the newest
01:12version of full self-driving,
01:14version 12, in a live stream he did while
01:17driving, violating
01:18Tesla's rules for its advanced
01:20autopilot technology. And the 45-minute
01:22demo went mainly well,
01:24except for a couple issues,
01:25including almost running a red light.
01:28Musk takes over at that point
01:29and reminds viewers that full
01:30self-driving V12 is still in beta,
01:33although V12 will be the first version to
01:35remove the beta label.
01:36But that Tesla was able to navigate many
01:39complex driving
01:40situations, including roundabouts and
01:41construction zones. Musk also mentions
01:44that V12 will be the first
01:45time that full self-driving is
01:47entirely cameras and AI, as opposed to
01:50previous versions which
01:51mixed in other sensors. Tesla has
01:53found that the best way to account for
01:55the thousands or even
01:57millions of edge cases humans
01:58experience while driving is to use a pure
02:01neural network approach, which is
02:02different from other
02:03companies like Cruise and Waymo. In true
02:05Musk Uber troll fashion, he
02:08claims in the video that he'll
02:09drive over to Mark Zuckerberg's house to
02:11initiate their much-anticipated but
02:12highly unlikely fight.
02:14Next, let's talk about OpenAI. Just a
02:16couple weeks ago, it was
02:17reported that OpenAI is burning
02:19through a tremendous amount of cash,
02:21about $700,000 a day, and costs
02:23associated with running their AI
02:24systems. But now, an article was
02:26published claiming OpenAI
02:28is on track to generate more
02:29than a billion dollars in revenue over
02:31the next 12 months, or about $80 million
02:33a month. A staggering
02:35number given they earned a total of $28
02:37million in all of last
02:38year. So it looks like burning
02:40all that cash isn't going to complete
02:41waste. It's generating a
02:43massive amount of revenue.
02:44But at the same time, another report
02:46published by Spark Toro shows visits to
02:49chat GPT are down 29%
02:51since its peak in May. And the majority
02:53of usage is for a coding
02:55assistant. I can tell you from my
02:57experience that coding assistance is my
02:59number one use case. And as we'll see in
03:01a story later in this
03:02video, that dominance may already be
03:05threatened by open source and
03:06completely free code llama by
03:08meta. And there are some other
03:10interesting findings from this report.
03:12Users are pretty split
03:13between using only one prompts during
03:15their session and using five
03:17prompts with nothing really in
03:19between. And those two ends of the
03:20spectrum accounting for
03:21nearly 70% of all visitors. And
03:24some of the most popular words used in
03:26chat GPT prompts include
03:27write, create, list and fun. This
03:30article has other awesome findings. I'll
03:32drop a link in the
03:33description below so you can check it
03:34out. So I really can't tell how open AI
03:36is doing revenue is
03:38exploding costs are exploding and
03:40usage is down. My take is that chat GPT
03:42is still settling into its
03:44baseline. Since it was such a
03:46revolutionary product, people are still
03:47figuring out how to
03:48integrate it into their lives. And to
03:50continue growing its revenue open AI has
03:52launched chat GPT enterprise.
03:55I can tell you firsthand from
03:56conversations I've had with my clients
03:58that privacy and security are
03:59the number one concern amongst
04:01businesses when considering chat GPT.
04:04Companies don't want to give
04:05sensitive data over to chat
04:07GPT to help train their models for later
04:09to be found in responses by
04:11that AI by other companies
04:12effectively leaking company secrets. Now
04:15with chat GPT enterprise,
04:17that concern has been more
04:18less quelled features from chat GPT
04:21enterprise include that
04:22customer props and customer data
04:24are not used for training open AI models
04:26data encryption at rest
04:27and in transit and their
04:30certified SOC 2 compliant. They also
04:32offer several highly
04:33requested features including an
04:35admin console, single sign on unlimited
04:38usage of GPT for increased speeds and
04:41larger context sizes.
04:43Chat GPT enterprise is a highly
04:45compelling product, a strong
04:46offering in the face of growing
04:48competition from the open source model
04:50world. With the guarantee
04:51of privacy and security,
04:53I can now recommend chat GPT as a real
04:55option amongst the open
04:56source models to companies that
04:58ask me which model they should use for
04:59their business. But this
05:00wouldn't be AI news if meta AI
05:02didn't launch something incredible and
05:05open source. At the end of last week,
05:08meta launched CodeLama,
05:09a fine tuned version of llama to
05:11explicitly trained for
05:12coding tasks. Shortly after that,
05:15multiple fine tuned versions of CodeLama
05:17were released that beat
05:18GPT for at coding problems.
05:20Yes, you heard me right, beat not just
05:24chat GPT, but GPT for
05:26also. This is an incredible
05:28accomplishment given I didn't think GPT
05:30for would have any competition in the
05:32coding realm anytime
05:33soon. GPT for has been my go to coding
05:35assistance since it was
05:36launched. But now I have a completely
05:38free and open source alternative. Not
05:40only that, but quantized versions with
05:42sizes ranging from a
05:43billion parameters all the way up to 70
05:45billion allow for pretty much any
05:47hardware to run these
05:48models. There's even a full unquantized
05:5134 billion parameter version
05:53running at over 20 tokens per
05:55second on an M2 ultra Mac. Be sure to
05:58check out the videos I made testing
06:00CodeLama versus GPT for
06:01and also the tutorial videos showing how
06:03to install CodeLama
06:04locally. Both will be in the
06:06description below. Our next story is
06:08about the constant march of AI beating
06:10humans at new things.
06:12This week, AI beat world champion drone
06:14racers. The AI system called Swift,
06:17designed by University
06:18of Zurich researchers beat the best human
06:21drone racers in the world. A feat
06:23considered impossible
06:24just a few years ago. Drone racing is a
06:26popular sport where
06:27racers navigate drones through
06:28complicated courses at speeds exceeding
06:31100 kilometers per hour,
06:32controlling them remotely
06:33through a VR like headset connected to an
06:35onboard camera. Training
06:37for this AI occurred in a
06:38simulated environment. And then the race
06:40occurred on an actual course. The AI
06:43control drone was able
06:44to beat the world record by a half of a
06:46second, which doesn't seem
06:47like much. But in the world of
06:48drone racing, everything is measured in
06:50fractions of a second. This
06:52accomplishment isn't just for
06:54fun. It actually has a lot of real world
06:56applications such as
06:58environmental monitoring,
07:00disaster reporting and rescue operations.
07:02What do you think will be
07:03the next thing that AI beats
07:04humans at? Let me know in the comments.
07:06Our next story is one that
07:08I'm very happy to be talking
07:09about. A16Z, the famed venture capital
07:12firm out of Silicon Valley
07:14seems to end up in my news
07:15videos almost every week now. This week
07:18they announced a grant
07:18program where they're giving
07:19away funding to a small group of AI
07:22developers to help open source
07:23community. Creating artificial
07:24intelligence is extremely expensive given
07:27the hardware
07:27requirements. Just look at the $300
07:30million AI cluster that Tesla just
07:32launched. The open source
07:33community gives their software away
07:35for free. So acquiring the expensive
07:36hardware to create and run
07:38open source models is nearly
07:39impossible. Now A16Z will be giving
07:42grants to some of the community's most
07:43prominent open source AI
07:45developers. Tom Jobins, also known as the
07:47bloke who I mentioned all
07:48the time, was one of the
07:50initial recipients of the grant. And I'm
07:52very happy to see this
07:53because I use his quantized models
07:55all the time. Congrats to all the grant
07:57recipients and a big thank
07:58you to A16Z for helping bolster
08:00the open source community. Next, not to
08:03be left out of the AI wave,
08:04Google made many announcements
08:05this week. First with its launch of Duet
08:07AI and Google Workspaces.
08:09This is a massive launch
08:11because Google Workspaces has 3 billion
08:13users. That number blew me
08:15away. I really didn't understand
08:16how that's possible and that's on par
08:18with Facebook. I don't
08:20know how they calculate those
08:21users but I guess it includes every Gmail
08:23user. Now Google Workspace
08:25users can access Duet AI,
08:27which Google describes as a powerful
08:29collaboration partner that
08:30can act as a coach, a source of
08:32inspiration, and productivity booster.
08:35You'll find Duet features in
08:36almost every product within the
08:38Google Workspace suite of products. And
08:40not only that, Google
08:42unveiled several new AI tools and
08:44capabilities at the Google Next
08:45conference in San Francisco.
08:47Let's take a look at some of the
08:48launches. Google's cloud service now
08:50includes 20 pre-built AI
08:52models optimized for enterprises
08:54like LAMATU and CLOD2. They also launched
08:57their new AI
08:57watermarking product Synth ID,
09:00which helps people identify AI-generated
09:02images created by
09:03their AI Gen art product,
09:05Imagine. The watermark is undetectable by
09:08the human eye but also persists even
09:10after modifications to
09:11the image are made, like filters, color
09:13changes, and brightness adjustments.
09:15Google also launched
09:16access to their new AI training cluster
09:18based on their custom-built TPU
09:20architecture, which can be
09:21used to train and fine-tune AI models.
09:24Last, Google updated its Vertex AI
09:26platform with upgrades to
09:27POM2, enhanced code generation, and new
09:30search and conversational models. Even
09:32with these launches,
09:33it still does feel like Google is playing
09:35catch-up to meta, open AI, and Microsoft.
09:37Now let's switch gears for a minute. In
09:39tech news, it's been reported that
09:41Silicon Valley Elite are
09:42building a city from scratch. According
09:44to the article in Marin Independent
09:46Journal, billionaire
09:47VC Michael Moritz and others had dreams
09:50of transforming tens
09:51of thousands of acres
09:52into a bustling metropolis that,
09:54according to the pitch,
09:55could generate thousands of jobs
09:57and be as walkable as Paris or the West
09:59Village in New York. He
10:00painted a kind of urban blank slate
10:03where everything from design to
10:05construction methods and
10:06new forms of governance can be
10:07rethought. Since the initial idea, large
10:10plots of land have been
10:11purchased and $800 million
10:13has been committed to the project from
10:16tech elites. These
10:17secretive land purchases have
10:19locals worried, though, unsure what will
10:21become of their quiet
10:22towns. Some of the investors that
10:24have been identified include Reid
10:26Hoffman, the founder of LinkedIn, Mark
10:28Andreessen of Andreessen
10:29Horowitz A16Z, Chris Dixon, Patrick and
10:32John Collison, who are
10:33the founders of Stripe,
10:35Lauren Powell Jobs, Steve Jobs' wife, and
10:37more. And this isn't the
10:38first time tech entrepreneurs
10:39have tried to affect California's
10:41significant and ongoing housing crisis.
10:43As someone who lives in
10:44California, anything to bring down the
10:46cost of living is something
10:47I'm all for. So I hope they
10:49build something incredible. Next, look
10:51out mid-journey. Another
10:53competitor is On the Horizon.
10:55ideogram this week launched in beta with
10:57a unique differentiator,
10:58being able to add text to AI
11:00generated images. Text in AI images has
11:03been a difficult problem
11:04to solve, but ideogram seems
11:06to have successfully solved it. Founded
11:08by ex-Google brain
11:09researchers, ideogram received a massive
11:12$16.5 million in funding from powerhouse
11:15investors like A16Z and Index
11:18Ventures. I don't know if just
11:20being able to add text within AI
11:22generated images is going to be enough to
11:24set them apart in such a
11:25crowded field, especially since
11:27competition is likely to add this
11:29functionality soon enough.
11:30Still, I wish them luck and the more
11:32competition, the better for consumers.
11:35Speaking of generative art, Runways Gen 2
11:37had another big release this week called
11:39Motion Slider. This feature allows you to
11:42select the number from 1 to
11:4310 to control the amount of
11:44movement in your output video. Take a
11:46look at this example. It seems like each
11:49week text to video is
11:50becoming better. Next, Apple may be well
11:53positioned to win the
11:54hardware game for AI. As it is
11:56increasingly difficult to get your hands
11:58on Nvidia GPUs, it turns
12:00out that Apple's own silicon,
12:01the M1 and the M2 are incredibly good at
12:04running AI models. In a
12:06lengthy tweet by AI pioneer Andre
12:08Carpathi, he details why the M2 chip is a
12:11great option for running
12:12large language models. And,
12:14as mentioned earlier, ex-user Georgi
12:17Gurgunov showed a video of himself
12:18running an unquantized
12:2034 billion parameter version of CodeLama
12:22at 20 tokens per second
12:24on an M2 Ultra. So, all you
12:26really need to run incredibly powerful
12:28large language models is
12:29an Apple computer. But,
12:31you may not even need a computer.
12:33According to the Stability AI
12:34founder, he believes we'll see
12:36a chat GPT level model on a mobile phone
12:39next year with a GPT4 level
12:42model the year after that.
12:43This is incredible news for the open
12:44source AI community and hints at what
12:46could be coming from
12:47the iPhone maker. Your move, Tim Apple.
12:49Now, for the AI video of the week. In
12:52what is sure to scare
12:53the pants off of Disney, ex-user Jeff
12:56Synthesize created a two and a half
12:58minute long AI-generated
13:00Pixar-like film using Mid-Journey and Gen
13:022 called Glitch. The
13:04video looks absolutely
13:05incredible and could have easily been
13:07created by Pixar. But instead it was
13:09created by one person,
13:10a very hard-working AI artist. Take a
13:12look at this 20
13:13second clip from the film.
13:37Generally, films like this take dozens if
13:41not hundreds of people to create. So,
13:43the implications for Disney are
13:44tremendous. Amid an ongoing
13:46writer strike and declining stock
13:48performance, I imagine Disney is looking
13:51very closely at AI
13:52technology to help them reduce
13:54their costs of creating amazing films. If
13:56you want to submit an
13:57entry for AI video of the week,
13:59jump in my Discord and find the AI video
14:01of the week channel. I'll
14:02link it in the description
14:03below. Our last story is about AI and
14:06copyright. As regulators
14:08race to figure out how to handle
14:09the avalanche of AI content being
14:11generated, they are now
14:13asking for input from the public
14:14in determining how to create AI copyright
14:16policy. The US Copyright
14:18Office has opened for public
14:20comment to figure out how to answer three
14:22main questions. How AI
14:23models should use copyrighted
14:25data in training, whether AI generated
14:27material can be copyrighted
14:29even without a human involved,
14:31and how copyright liability will work
14:33with AI. Just last week I
14:35reported that it was ruled
14:36AI art can't be copyrighted, but it seems
14:38that decision isn't the
14:39last word on the subject.
14:41And also last week I reported on major
14:43lawsuits filed against OpenAI for
14:45allegedly training their
14:46models on copyrighted data. It'll be
14:47interesting to see how all of
14:49these legal elements of AI play
14:50out and I'll keep you up to date all
14:52along the way. If you liked
14:54this video, please consider
14:55giving me a like and subscribe and I'll
14:57see you in the next one.
