- Forward Future AI
- Posts
- Forward Future Episode 21
Forward Future Episode 21
Tesla Goes Full AI, LLaMA Beats GPT4, OpenAI Biz, AI Drone Racing, Apple Silicon for AI
I say this every week, but we really did have another incredible week of AI news this week. Tesla has multiple AI launches, OpenAI is raking in the cash but is also struggling? AI sets records at drone racing, beating the best humans, and Meta launches code llama, which quickly beat GPT4 at coding tasks. We have a ton of other fantastic tech and AI stories today, so sit back, relax, and enjoy!
Tesla AI
This week, Tesla launched a massive $300 million AI cluster, including 10,000 Nvidia H100 compute GPUs. By the way, this is why Nvidia crushed earnings last week. Their new AI cluster will be used to power several AI applications, but of course, its main use will be to continue training its full self-driving product. And they have good reason to have such a massive super-computer. According to Tim Zaman, AI Infra & AI Platform Engineering Manager at Tesla: "Due to real-world video training, we may have the largest training datasets in the world, hot tier cache capacity beyond 200PB — orders of magnitudes more than LLMs.”
Tesla Full Self-Driving
Elon Musk showed off the newest version of Full Self Driving, version 12, in a live stream he did while driving, violating Tesla’s rules for its advanced autopilot technology. The 45-minute demo mainly went well, except for a few issues, including almost running a red light. Musk takes over and reminds viewers that FSD v12 is still in beta, although v12 will be the first version to remove the beta label. The Tesla could easily navigate many complex driving situations, including roundabouts and construction zones. Musk also mentions that v12 will be the first time that FSD is entirely cameras and AI, as opposed to previous versions, which mixed in other sensors. Tesla has found that the best way to account for the thousands or even millions of edge cases humans experience while driving is to use a pure neural network approach, unlike other companies like Cruise and Waymo. And in true Musk uber-troll fashion, he claims in the video that he’ll drive over to Mark Zuckerberg’s house to initiate their much anticipated but highly unlikely fight.
OpenAI
A couple of weeks ago, it was reported that OpenAI is burning through a tremendous amount of cash, about $700k/day, in costs associated with running their AI systems. But now, an article was published claiming OpenAI is on track to generate more than $1 billion in revenue over 12 months, or about $80 million per month, a staggering number given they earned a total of $28 million in all of last year. So it looks like burning all that cash isn’t going to waste. It’s generating a massive amount of revenue.
ChatGPT Usage Decline
But at the same time, another report published by Sparktoro shows visits to ChatGPT are down 29% since its peak in May, and the majority of usage is coding assistant. I can tell you from my experience that coding assistance is my #1 use case. And, as we’ll see in a story later in this video, that dominance may be threatened already by open-source and completely free Code LLaMA by Meta. And there are some other interesting findings from this report. Users are pretty split between using only one prompt during their session and 5+ prompts, with those two ends accounting for nearly 70% of all visitors. Some of the most popular words used in ChatGPT prompts include Write, Create, List, and Fun. This article has other excellent findings, and I’ll drop a link in the description so you can check them out. So, I really can’t tell how OpenAI is doing. Revenue is exploding, costs are exploding, and usage is down. My take is that ChatGPT is still settling into its baseline. Since it was such a revolutionary product, people are still figuring out how to integrate it into their lives.
ChatGPT Enterprise
And to continue growing its revenue, OpenAI has launched ChatGPT Enterprise. I can tell firsthand from conversations with my clients that privacy and security are the number 1 concern amongst businesses when considering ChatGPT. Companies don’t want to give sensitive data over to ChatGPT to help train their models for it to be later found in prompts by other companies, effectively leaking company secrets. Now, with ChatGPT Enterprise, that concern has been more or less quelled. Features from ChatGPT Enterprise include that Customer prompts and company data are not used for training OpenAI models, Data encryption at rest (AES 256) and in transit (TLS 1.2+), and they are Certified SOC 2 compliant. They also offer several highly requested features, including an admin console, single sign-on, unlimited usage of GPT-4, increased speeds, and larger context windows. ChatGPT Enterprise is a highly compelling product, a strong offering in the face of growing competition from open-source models. With the guarantee of privacy and security, I can now recommend ChatGPT as a natural option amongst the open-source models to companies that ask me which model they should use for their business.
Code LLaMA Beats GPT4
But this wouldn’t be AI news if Meta AI didn’t launch something incredible and open-source. At the end of last week, Meta launched Code LLama, a fine-tuned version of LLama 2, explicitly trained for coding tasks. Shortly after that, multiple fine-tuned versions of Code LLama were released that beat GPT4 at coding problems. Yes, you heard me right, beat not just ChatGPT, but GPT4 also. This is an incredible accomplishment, given I didn’t think GPT4 would have any competition in the coding realm soon. GPT4 has been my go-to coding assistant since it was launched, but now I have a completely free and open-source alternative. Also, quantized versions with sizes ranging from 1 billion to 70 billion parameters allow pretty much any hardware to run this model. There’s even a full, unquantized 34B version running at over 20 tokens/second on an M2 Ultra Mac. Be sure to check out the videos I made testing Code LLama vs. GPT4 and a tutorial video showing how to install Code LLama locally.
AI Drone Race
Our next story is about the constant march of AI beating humans at new things. This week, AI beat world champion drone racers. The AI system called “Swift,” designed by the University of Zurich researchers, beat the best human drone racers in the world, a feat considered impossible just a few years ago. Drone racing is a popular sport where racers navigate drones through complicated courses at speeds exceeding 100 km/hr, controlling them remotely through a VR-like headset connected to an onboard camera. Training occurred in a simulated environment, and the race occurred on an actual course. The AI-controlled drone beat the world record by half of a second, which doesn’t seem like much, but in the world of drone racing, everything is measured in fractions of a second. This accomplishment isn’t just for fun. It has many real-world applications, such as environmental monitoring, disaster reporting, and rescue operations. What do you think will be the next thing AI will beat humans at? Let me know in the comments.
AI Grant Program
Our next story is one I’m thrilled to be talking about. a16z, the famed venture capital firm out of Silicon Valley, seems to end up in my news videos almost every week now. This week, they announced a grant program where they are giving away funding to a small group of AI developers to help the open-source community. Artificial intelligence is extremely expensive to create, given the hardware requirements. Just look at the $300 million AI cluster Tesla just launched. The open-source community gives their software away for free, so acquiring the expensive hardware to create and run open-source models is nearly impossible. Now, a16z will be giving grants to some of the community's most prominent open-source AI developers. Tom Jobbins, also known as “The Bloke” was one of the initial recipients of a grant, and I’m pleased to see this because I use his quantized models all the time. Congrats to all the recipients of the grants, and a big thank you to a16z for helping bolster the open-source AI community.
Google Duet AI
Not to be left out of the AI wave, Google made many announcements this week with its launch of Duet AI in Google Workspaces. This is a massive launch because Google Workspaces has 3 billion users. That number blew me away, on par with Facebook. I don’t know how they calculate those users, but it includes every Gmail user. Now, Google Workspace users can access Duet AI, which Google describes as a powerful collaboration partner that can act as a coach, source of inspiration, and productivity booster. You’ll find Duet features in almost every product within the Google Workspace suite of products.
SynthID
Google unveiled several new AI tools and capabilities at the Google Next conference in San Francisco. Let’s take a look at some of the launches. Google’s Cloud service now includes 20 prebuilt AI models optimized for enterprises like LLama 2 and Claude 2. They also launched their new AI watermarking product SynthID, which helps people identify AI-generated images created by their AI gen art product, Imagen. The watermark is undetectable by the human eye and persists even after image modifications, such as filters, color changes, and brightness adjustments. Google also launched access to their new AI training cluster based on their custom-built TPU architecture, which can be used to train and fine-tune AI models. Last, Google updated its Vertex AI platform with upgrades to Palm 2, enhanced code generation, and new search and conversational models. Even with these launches, it still does feel like Google is playing catch up to Meta, OpenAI, and Microsoft.
Silicon Valley Elite Buy Massive Land
Let’s switch gears for a minute. In tech news, it has been reported that Silicon Valley Elite is building a city from scratch. According to the article in Marin Independent Journal, billionaire VC Michael “Moritz and others had dreams of transforming tens of thousands of acres into a bustling metropolis that, according to the pitch, could generate thousands of jobs and be as walkable as Paris or the West Village in New York…He painted a kind of urban blank slate where everything from design to construction methods and new forms of governance could be rethought.” Since the initial idea, large plots of land have been purchased, and $800 million has been committed to the project from other tech elites. The secretive land purchases have locals worried, unsure of what will become of their quiet towns. Some of the investors that have been identified include Reid Hoffman, founder of Linkedin, Marc Andreeson of Andreeson Horowitz, Chris Dixon, Patrick and John Collison, founders of Stripe, Laurene Powell Jobs, and more. And this isn’t the first time tech entrepreneurs have tried to affect California’s significant and ongoing housing crisis. As someone who lives in California, anything to bring down the cost of living is something I’m all for, so I hope they build something incredible.
Ideogram Launches
Next, look out Midjourney. Another competitor is on the horizon. Ideogram this week launched in beta with a unique differentiator: being able to add text to AI-generated images. Text in AI images has been a complex problem, but Ideogram has solved it successfully. Founded by ex-Google Brain researchers, Ideogram received a massive $16.5 million in funding from powerhouse investors like a16z and index ventures. I don’t know if just being able to add text within images is going to be enough to set them apart in such a crowded field, especially since competition is likely to add this functionality soon enough. Still, I wish them luck; the more competition, the better for consumers.
Gen2 Motion Slider
Runway’s Gen2 had another big release this week called Motion Slider. This feature allows you to select a number from 1 to 10 to control the amount of movement in your output video. Take a look at this example video. It seems like each week text to video is becoming better.
M2 Mac for AI
Apple may be well-positioned to win the hardware game for AI. As it is increasingly difficult to get your hands on Nvidia GPUs, Apple’s silicon, the M1 and M2, are incredibly good at running AI models. In a lengthy tweet by AI pioneer Andrej Karpathy, he details why the M2 chip is an excellent option for running large language models. And, as mentioned earlier, x user Georgi Gerganov showed a video of himself running an unquantized 34b version of code llama at 20 tokens per second on an M2 ultra. So, all you need to run compelling large language models is an Apple computer! But you may not even need a computer. According to the StabilityAI founder, he believes we’ll see a ChatGPT-level model on a mobile phone next year with a GPT4-level model the year after that. This is incredible news for the open-source AI community and hints at what could be coming from the iPhone maker. Your move, Tim Apple.
AI Video of the Week
Now for the AI video of the week! In what will scare the pants off Disney, X user Jeff Synthesized created a two-and-a-half-minute-long AI-generated Pixar-like film using Midjourney and Gen 2 called Glitch. The video looks incredible and could have easily been created by Pixar, but instead, it was created by one very hard-working AI artist. Generally, films like this take dozens, if not hundreds, of people to create, so the implications for Disney are tremendous. Amid an ongoing writer’s strike and declining stock performance, I imagine Disney is looking very closely at AI technology to help them reduce their costs of creating notable films. If you want to submit an entry for AI video of the week, jump into my discord and find the “AI video of the week” channel.
AI Copyright
Our last story is about AI and copyright. As regulators race to figure out how to handle the avalanche of AI content being created, they are now asking for input from the public in determining how to create AI copyright policy. The US Copyright Office has opened for public comment to figure out how to answer three main questions: how AI models should use copyrighted data in training, whether AI-generated material can be copyrighted even without a human involved, and how copyright liability would work with AI. Just last week, I reported that it was ruled AI art cannot be copyrighted, but it seems that decision isn’t the last word on the subject. And also last week, I reported on significant lawsuits filed against OpenAI for allegedly training their models on copyrighted data. It’ll be interesting to see how all of the legal elements of AI play out, and I’ll keep you up to date all along the way.
Reply