Tech Brew Ride Home - Fri. 12/27 – A New “Definition” Of AGI

Episode Date: December 27, 2024

Microsoft and OpenAI are reportedly having problems deciding how make this whole, reclassifying as for-profit thing work. Oh, also, they secretly decided on a new definition of artificial general intelligence that is tied to tangible dollar value. At the tail end of the year, the most interesting new AI model of the year. And, of course, the Weekend Longreads Suggestions. Links: Microsoft and OpenAI Wrangle Over Terms of Their Blockbuster Partnership (The Information) Microsoft and OpenAI’s Secret AGI Definition (The Information) DeepSeek-V3, ultra-large open-source AI, outperforms Llama and Qwen on launch (VentureBeat) Weekend Longreads Suggestions: Video Games Can’t Afford to Look This Good (NYTimes) Hawk Tuah Wasn’t What It Seemed (The Atlantic) The Paper Passport Is Dying (Wired) You Need to Create a Secret Password With Your Family (Wired) Learn more about your ad choices. Visit megaphone.fm/adchoices

On April 4th, 2023, around 2 in the morning, a man was found stabbed multiple times on a sidewalk in downtown San Francisco. Hey, who did this to you? What happened next turned the story into a political firestorm. Reports have identified the victim as Bob Lee, the founder of Cash App. From Bloomberg Podcasts, this is Foundering, the Killing of Bob Lee, beginning April 16. Welcome to the Tech meme right home for Friday, December 27th, 2024. I'm Brian McCullough today. Microsoft and Open AI reportedly are having problems deciding on how to make this whole reclassifying as a for-profit thing work. Oh, and also they secretly decided on a new definition of artificial general intelligence that is tied to tangible dollar value at the tail end of the year, the most interesting new AI model of the year, and of course the weekend long read suggestions.

Here's what you missed today in the world of tech. couple of interesting open AI scoops from the information for you. First up, they're reporting that Microsoft and OpenAI are having trouble sorting out the details of Microsoft's stake in any forthcoming for-profit OpenAI entity. Also, they're trying to work out details like the use of OpenAI's IP, the continued collection of 20% of OpenAI's revenue, and more. Quote, the companies have been negotiating potential changes in Open AI structure since around October. Those talks have focused on four areas. Microsoft's equity stake in the for-profit entity, whether Microsoft will continue to be Open AI's exclusive cloud

provider, how long Microsoft will maintain rights to use Open AI's intellectual property in its products as it pleases, and whether Microsoft will continue to take 20% of Open AI's revenue, according to a person who has talked to Altman about the discussions. The negotiations also reflect the remarkable acceleration of OpenAI's business, as well as its ambitions to develop everything from a server chip to a web browser to a humanoid robot. OpenAI has projected around $4 billion in revenue this year and $100 billion in 2029, mainly off the back of chat GPT. Given such growth, the terms of its contract with Microsoft, including the 20% revenue share and its reliance on Microsoft servers, are now harder for OpenAI to swallow. It isn't clear when OpenAI and Microsoft plans,

plan to complete the process, but they are working quickly and are on the clock. If OpenAI fails to make the change in the next two years, investors in the recent capital raise can take their money back, plus 9% interest, a total of about $7.2 billion. If the company continues growing like it has been, it isn't clear they would want to get the money back. Company leaders have told employees, Open AI wants to buy some of their shares after the for-profit conversion, so they have every reason to want to make this change soon. Open AI has said it capped potential profits for investors to balance shareholder returns with achieving its ethical and social goals of developing AI to benefit humanity.

It previously talked to Microsoft, which was entitled to $93 billion in maximum profits about raising the cap 20% per year. If it raised the cap, the actual AGI profit capability target could be closer to $120 billion, including the maximum profits owed to other investors, such as Y Combinator and Coast of ventures. Those kinds of profits are not happening anytime soon, however, OpenAI currently loses billions of dollars a year, and around September this year, it told potential investors it didn't anticipate turning its first annual profit until 2029. That means it could take years beyond that point for OpenAI to generate more than $100 billion in total profits. Another growing

point of friction between OpenAI and Microsoft involves their cloud computing deal. Contractually, Microsoft is the exclusive supplier of cloud servers to OpenA. and the only company allowed to resell Open AIs models to cloud customers. But Microsoft has struggled to supply the startup with enough servers to train and run its AI, Open AI has claimed, and some Open AI leaders believe the company could increase sales if other cloud providers such as Amazon and Google were allowed to resell Open AI models. It isn't clear if Microsoft will budge on its cloud exclusivity, but Open AI has pushed the boundaries of the deal when it comes to getting access to cloud servers without Microsoft's

help. In Abilene, Texas, it negotiated directly with other providers, including Oracle for access to AI servers in the middle of next year, though Microsoft is technically still the main customer of the project. Microsoft had negotiated the right to block any arrangements OpenAI made with other cloud providers, so it likely would have had to bless the deal before it moved forward. Open AI also may be getting help for Microsoft's rivals in loosening its cloud grip. Google has asked U.S. authorities to scrutinize and seek to break apart. the Microsoft OpenAI Cloud deal on antitrust grounds, end quote. But wait, let me back up for a second and underline something that you might have missed between

the lines in the previous paragraphs there. The information kind of buried the lead about a previously undisclosed 2023 agreement between Microsoft and Open AI that defines achieving AGI artificial general intelligence as the point when OpenAI develops AI systems that generate more than $100 billion in profits. Honey, wake up, a new definition of artificial general intelligence has arrived. Quoting the information again, for OpenAI and Microsoft, its largest investor and exclusive cloud provider, AGI has a very specific definition. The point when OpenAI develops AI systems that can generate at least $100 billion in

profits. This is especially important for both parties because once OpenAI re-eastern reaches AGI, it can effectively end its arrangement with Microsoft, meaning that the tech giant won't be able to use the technology OpenAI produces after that point. In some ways, it feels wrong to place a price tag on AGI, a concept that Open AI, CEO Sam Altman and his peers, have raved about as a major milestone on our journey to a future where AI can help us cure diseases and colonize Mars. But in other ways, it's completely unsurprising. It never would have made sense for a multi-trillion dollar technology company like Microsoft to have OpenAI make such a consequential decision based on a subjective calculation. Luckily for Microsoft, the AGI pronouncement seems a long

way off since Open AI has projected it'll lose billions of dollars until it finally turns a profit in 2019. As we noted in the piece, though, there are some aspects of the AGI declaration that are open to interpretation. And as part of the current negotiations, it's possible Open AI could amend its rules so that Microsoft could keep using the tech after AGI is achieved. Still, with this new knowledge, it's funny to see Altman and other AI execs try to work AGI into every speech. And over time, Altman has toned down the importance of AGI saying things like, quote, AGI is basically the equivalent of a median human that you could hire as a co-worker. And, quote, my guess is we will hit AGI sooner than most people in the world think, and it will

matter much less, end quote. Indeed, he's emphasized super intelligence, the aforementioned advanced AI that can help us cure diseases and colonize Mars, as the real end goal to worry about. I wonder how much that level of intelligence would be worth to Microsoft, end quote. All right, I've got another new AI model to tell you about, but wait, wait, I can already hear you rolling your eyes or hitting fast forward, so let me tell you upfront why I think this one is different and notable. Deep Seek is a Chinese AI company, so that's a it different. And they've released DeepSeek V3, which they claim outperforms top models like GPT-40, though lots of people say their models are cutting edge and beat the current leaders in the

clubhouse. But also, it's an MOE or mixture of experts model, so that's different. And DeepSeek V3 is open source, so that's also notable. The idea that open source models can outperform the proprietary black box ones is interesting, but also get how much they're saying they're outperforming this state of the art. They say that their model has 671 billion total parameters with 37 billion parameters activated per token. And get this, quoting Venture Beat, notably during the training phase, deep seek used multiple hardware and algorithmic optimizations, including the FP8 mixed precision training framework and the dual pipe algorithm for pipeline parallelism to cut down on the cost of the process. Thus, overall,

claims to have completed DeepCc V3's entire training in about 2,788KH800 GPU hours or about $5.57 million, assuming a rental price of $2 per GPU hour. This is much lower than the hundreds of millions of dollars usually spent on pre-training large language models. Lama 3.1, for instance, is estimated to have been trained with an investment of over 500,000. million, end quote. So that's why this is notable. A model that is outperforming the state of the art that only took $5.5 million to train versus $500 million to train, quoting no less than Andre Carpathie on X, quote, for reference, this level of capability is supposed to require clusters of closer to 16,000 GPUs. The ones being brought up today are more around 100,000 GPUs,

EG Lama 3405B used 30.8 million GPU hours, while Deepseek V3 looks to be a stronger model at only 2.8 million GPU hours or around 11 times less compute. If the model also passes vibe checks, e.g. LLM Arena rankings are ongoing. My few quick tests went well so far. It will be a highly impressive display of research and engineering under resource constraints. Does this mean you don't need large GPU clusters for frontier LLMs? No, but you have to ensure that you're not wasteful with what you have, and this looks like a nice demonstration, that there's still a lot to get through with both data and algorithms, end quote. And here's Elad Gill, quote, good evidence that a lot of efficiency is being left on the table in US labs with massively scaled clusters, end quote. And to that end, quoting Bojan Tungu's quote,

all the export bans on high-end semiconductors might have actually been counterproductive in the worst way imaginable. They seem to have forced Chinese researchers to be far more ingenious and resource-efficient than they might have otherwise been. It also seems to confirm my own hypothesis that we are far, far away from having the best algorithms for the ML part of AI, end quote. But back to Ventrabee to sum this up, quote, The work shows that open source is closing in on closed source models, promising nearly equivalent performance across different tasks. The development of such systems is extremely good for the industry, as it potentially eliminates the chances of one big AI player ruling the game. It also gives enterprises multiple options to choose from and work with while also orchestrating their stacks.

Currently, the code for Deepseek V3 is available via GitHub under an MIT license, while the model is being provided under the company's model license. Enterprises can also test out the new model via DeepSeek chat, a chat GPT-like platform, and access the API for commercial use. DeepSeek is providing the API at the same price as DeepSeek V2 until February 8th. After that, it will charge 27 cents per million token inputs and seven cents per million tokens with cash hits and $1.10 per million output tokens, end quote. Time for the weekend long read suggestions. First up, as I've said several times recently,

one of the biggest stories of the year has been the anus horribulus the video game industry has had in 2024, but it's not just layoffs. It's not just consolidation. It's not just poor sales. It's also that the very fundamental model the entire industry has been operating under for at least 20 years seems to be fraying around the edges. For example, the New York Times takes a look at how making cinematic games. The cutting-edge graphics are getting so expensive and time-consuming that investing in graphics is seemingly providing diminishing financial returns. Quote, one way to understand the video game industry's current crisis is by looking closely at Spider-Man's spandex. For decades, companies like Sony and Microsoft have bet that realistic graphics were the key to attracting bigger

audiences. By investing in technology, they have elevated flat, pixelated worlds into experiences that often feel like stepping into a movie. Designers of last year's Marvel's Spider-Man 2 used the processing power of the PlayStation 5 so Peter Parker's outfits would be rendered with realistic textures and skyscraper windows could reflect rays of sunlight. That level of detail did not come cheap. Insomniac Games, which is owned by Sony,

spent around $300 million to develop Spider-Man 2, according to leaked documents, more than triple the budget of the first game in the series, which was released five years earlier. chasing Hollywood realism requires Hollywood budgets, and even though Spider-Man 2 sold more than 11 million copies, several members of Insomniac lost their jobs when Sony announced 900 layoffs in February. Cinematic games are getting so expensive and time-consuming to make that the video game industry has started to acknowledge that investing in graphics is providing diminishing financial

returns. It's very clear that high-fidelity visuals are only moving the needle for a vocal class of gamers in their 40s and 50s, said Jacob Navak, a former executive, at Square Inix, who left that studio known for the Final Fantasy series in 2016 to start his own media company. What does my seven-year-old son play, he asked? Minecraft, Roblox, Fortnite, end quote. In other words, games that graphics don't really matter at all for. Then, from the Atlantic, the rise and fall of the meme of the year, Hock Tua, quote, The way Haley Welsh has told the story on her podcast, she initially spent weeks hiding from the hawk to a meme, barricaded in her bedroom, streaming rom-coms. She was embarrassed by the joke,

which she had told while drunk, and horrified by the number of people who had seen it. A friend persuaded her to come out of her room only because other people were profiting off of her, selling bootleg merch and getting views on their own videos that used the clip. It would be silly not to get her own piece of the pie, especially because the pie might soon be gone. By July, she was everywhere, end quote. And then, Wired takes a look at how the paper passport is done, quote, the push to remove paper passports is happening worldwide. So far, airports in Finland, Canada, the Netherlands, the United Arab Emirates, the United Kingdom, Italy, the United States, India, and elsewhere have been trialling various levels of passport-free travel or the technology needed to

make it happen. In October, officials in Singapore announced that its residents can fly to and from the country without using their documentation, and foreign visitors can, quote, enjoy the convenience of passportless clearance when they depart Singapore. More than one and a half, million people have used the systems. Officials claim, while trials around the world are at different stages and use different technical infrastructure, they broadly work in similar ways. Information historically stored in your passport's NFC chip, including facial data, is instead stored digitally and linked to your phone. The EU is planning to build an official travel app for this. When you are at an airport, the phone can be shown, and a face recognition camera will try to match you to the passport

photo, end quote. And finally, finally, this is something that our family did just this very week. Also from Wired, what with AI voice cloning, supercharging scams where people call your loved ones using your own voice, and you got to imagine that video cloning is months away from being equally as good, you probably need to come up with a family password. That way, when someone calls grandma on FaceTime with your face or voice, asking for money, she can say, what's the password? And if the scammer doesn't have it, she knows it isn't you. Again, we did this this week because the family was all together and we could collectively decide on a password we'd all remember, but also because, you know, I have

thousands of hours of my voice and my face online. But seriously, you should consider doing this too, no matter what your situation. Quote, as with your online passwords, there are things you should and shouldn't do when it comes to creating a shared passphrase. For starters, you shouldn't make a passphrase the same as your passwords. And they shouldn't be things a scammer could easily find, such as street names, birthdays, pets, or other personal information that may be shared online. Consider anything that you or your loved ones post online as data available to scammers. England said, even if you keep all social media private, your data is available to your connections and followers who can be hacked. A good family passphrase, according to Starling Bank's

advice, could be anything that is unique, easy to remember, and can ideally be shared with friends and family in person. The guidelines give some hypothetical examples such as a short phrase like cheese puffs or rainbows and dragons or a mnemonic device like ABC, which stands for ants bake cakes. Avoid joking about your code word in your text messages, social media posts, etc. Tobeck says, we see folks make a family code word and then it's hashtagged in their Instagram post. That's not a great idea. It's got to be kept private. Tobeck says that while family passwords can be useful, it's crucial to be aware of their potential limitations. We have to remember how human beings actually act in an emergency, she says.

For instance, if someone has really been in a car accident, Tobac says it's likely adrenaline would kick in and they may not remember a passphrase. Tobac recommends using a method she calls being politely paranoid, which at its simplest is trying to verify the identity of the person who is contacting you by a second method. If you receive a call from a nephew who says they are in an accident and need money for legal fees, you can say, cool, 100%, I can help you. I just texted you a word. Go ahead and read that out to me, end quote. You might remember that my wife, the architect, works as an architect on Broadway, so she knows tons of Broadway singers and actors, because she works with them. So she recommended a voice trick that

she's seen all of that vocal talent use on Broadway, a nebulizer. I got a cheap one from VIX. over this holiday week and I've been using it. My voice isn't 100% back yet, but it's getting there. I'll talk to you on Monday, and then we'll figure out how many shows we're going to do next week, the New Year's week. I guess it'll be news dependent probably. Anyway, hope you're having fun with family as I am. See you on Monday.