Tech Brew Ride Home - Thu. 03/02 – And Right On Schedule, Here Come The APIs

Starting point is 00:00:00 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 Technification. me right home for Thursday, March 2nd, 2023. I'm Brian McCullough today. A chat GPT API for businesses here.

Starting point is 00:00:42 Microsoft gives Bing those knobs and dials I've been talking about. What are multimodal LLMs? New turmoil in crypto, this time around one of the big crypto-friendly banks. How is it going in terms of social platforms diversifying into subscription revenue? And why the FDA has rejected Neurrelinks applications to begin human testing of brain implants? Here's what you missed today in the world of tech. Something, something, time for the APIs. OpenAI has launched a chat GPT API for businesses with dedicated capacity plans priced at 0.002 cents for every 750 words and says that Snap and Shopify are among the early adopters.

Starting point is 00:01:29 I guess we could have seen that coming after Snap's recent bot announcement, quoting TechCrunch. OpenAI being a business, albeit a capped profit one, had to monetize chat GPT somehow less than investors get antsy. It took a step toward this with the launch of a premium service chat GPT Plus in February, and it made a bigger move today introducing an API that will allow any business to build chat GPT tech into their apps, websites, products, and services. An API was always the plan, that's according to Greg Brockman, the president and chairman of OpenAI, and also one of the co-founders. He spoke with me yesterday afternoon via a video call ahead of

Starting point is 00:02:08 launch of the chat GPT API. It takes us a while to get these APIs to a certain quality level, Brockman said. I think it's kind of this like being able to meet the demand and the scale, end quote. Brockman says the chat GPT API is powered by the same AI model behind OpenAI's wildly popular chat GPT, dubbed GPT 3.5 Turbo. Chat GPT 3.5 is the most powerful text-generating model open AI offers today through its API suite. The Turbo, the turbo Moniker refers to an optimized, more responsive version of chat GPT 3.5 that OpenAI has been quietly testing for chat GPT. Priced at 0.002 per 1,000 tokens or about 750 words. Brockman claims that the API can drive a range of experiences including non-chat applications.

Starting point is 00:02:59 Snap, Quizlet, Instacart, and Shopify are among the early adopters. The initial motivation behind developing GPT3.5 Turbo might have been, to cut down on chat GPT's gargantuan compute costs. OpenAI CEO Sam Altman once called chat GPT's expenses eye-watering, estimating them at a few cents per chat in compute costs. With over a million users, that presumably adds up quickly. But Brockman says that GPT 3.5 turbo is improved in other ways. If you're building an AI-powered tutor, you never want the tutor to just give an answer to the student. You want it to always explain it to help them learn. That's an example of the kind of system you should be able to build with the API, Brockman said.

Starting point is 00:03:41 We think this is going to be something that will just make the API much more usable and accessible, end quote. The chat GPT API underpins My AI, Snaps recently announced chatbot for Snapchat plus subscribers, and Quizlitz, new QChat virtual tutor feature. Shopify used the chat GPT API to build a personalized assistant for shopping recommendations, while Instacart leveraged it to create Ask Instacart an upcoming tool that'll allow Instacart customers to ask about food and get shoppable answers informed by product data from the company's retail partners, end quote. And what did I say after my own AI bot experiments back in January, if we're going to be prompt engineers soon, we're going to need the knobs and dials necessary to modify the outputs that we want?

Starting point is 00:04:33 Well, Microsoft updated its Bing chatbot to let users toggle between creative, balanced, and precise tones, though this is likely largely an effort to counter Bing's recent wild outbursts, quoting the verge. There are three options for the AI-powered chatbot's responses, creative, balanced, and precise. The creative mode includes responses that are, quote, original and imaginative, whereas the precise mode favors accuracy and relevancy for more factual and concise answers. Microsoft has set the default for the Bing chatbot to the balanced mode, which it hopes will strike a balance between accuracy and creativity. These new chat modes are rolling out to all Bing AI users right now, and around 90% of users should be seeing them already. The update includes a, quote, significant reduction in cases where Bing refuses to reply for no apparent reason.

Starting point is 00:05:24 According to Mikhail Parrakeen, the head of web services at Microsoft, it should also mean, quote, reduced instances of hallucination in answers. So hopefully the Bing chatbot produces less wild responses, end quote. So 11 Labs, which is what I have been using for my voice bot, has similar settings. Like if I dial up the creativity, that's how you get more of those natural breath breaks and some of the personality behind some of the intonations on certain words. But if you go too far in that direction, it starts to get overly dramatic, overacting, if you will, so you have to find a balance with the precision setting. I found that roughly a 70-30 split seems to work well. Microsoft researchers have unveiled Cosmos 1, a multimodal LLM. They claim can

Starting point is 00:06:18 understand image content, pass visual IQ tests, and accepts a variety of input formats, which apparently is key here, quoting R's Technica. The researchers believe multimodal AI, which integrates different modes of input such as text, audio, images, and video, is a key step to building artificial general intelligence, or AGI, that can perform general tasks at the level of a human. Visual examples from the Cosmos One paper show the model analyzing images and answering questions about them, reading texts from an image, writing captions for images,

Starting point is 00:06:52 and taking a visual IQ test with 22 to 26 percent accuracy. While the media buzzes with news about large language models, LLMs, Some AI experts point to multimodal AI as a potential path toward general artificial intelligence, a hypothetical technology that will ostensibly be able to replace humans at any intellectual task and any intellectual job. AGI is the stated goal of OpenAI, a key business partner of Microsoft in the AI space. In this case, Cosmos 1 appears to be purely a Microsoft project without open AI's involvement. The researchers call their creation a multimodal large language model, or MLLM, because its roots lie in natural language processing, like a text-only LLM, such as ChatGBTBT.

Starting point is 00:07:37 And it shows. For Cosmos 1 to accept image input, the researchers must first translate the image into a special series of tokens, basically text, that the LLM can understand. Microsoft trained Cosmos 1 using data from the web, including excerpts from the pile, an 800-gigabyte English text resource, and Common Crawl. After training, they evaluated Cosmos 1's abilities on several tests, including language understanding, language generation, optical character recognition, free text classification, image captioning, visual question answering, web page question answering, and zero shot image classification. In many of those tests, Cosmos 1 outperformed current state-of-the-art models according to Microsoft, end quote. Ah, but there's another potential roadblock

Starting point is 00:08:27 on the way to all of this AI stuff taking over, and it's another one that I should have thought of, but didn't. Email app BlueMails co-founder says Apple blocked an update to the Blue Mail app that added ChatGPT-powered features over inappropriate content concerns requiring a 17-plus age limit, quoting the Wall Street Journal. Apple took steps last week to block an update of email app BlueMail because of concerns that a new AI feature in the app could show inappropriate content. According to Ben Volok, co-founder of Blue Mail developer Blix, and documents viewed by the Wall Street Journal. Your app includes AI-generated content but does not appear to include content filtering at this time, Apple's app review team said last week in a message to the

Starting point is 00:09:11 developer reviewed by the journal. The app review team said that because the app could produce content not appropriate for all audiences, Blue Mail should move up its age restriction to 17 and older or include content filtering the documents show. Mr. Vlach says it has content filtering capabilities. The app's restriction is currently set for users four years old and older. Apple's age restriction for 17 and older is for categories of apps that may include everything from offensive language to sexual content and references to drugs. Mr. Volutch says that this request is unfair and that other apps with similar AI functions without age restrictions are already allowed for Apple users. Apple's attempt to set an age restriction to help moderate content

Starting point is 00:09:52 from a language model-based AI is an indication the tech giant is closely watching the new technology and the risks it poses. The company has long said it must carefully curate and review what software can be accessed on its iPhone and iPad through its app store to keep its products private and secure, end quote. Non-AI news, the big news in the world of crypto right now is that crypto-friendly US bank Silvergate has delayed its annual report citing investigations from regulators. Silvergate shares have plunged over 40 percent, and this morning, Coinbase announced that it was switching banking partners for its prime accounts from Silvergate to signature, after Silvergate reported being, quote, less than well capitalized. Quoting CoinDesk. In light of recent developments and out of

Starting point is 00:10:45 an abundance of caution, Coinbase is no longer accepting or initiating payments to or from Silvergate and will be facilitating institutional client cash transactions with our other banking partners, the company said. The exchange adopted signature bank for its prime customers, according to an email sent to clients. The changes come after Silvergate, a large bank for FinTech and crypto, said it would not be able to file its annual financial report on time. The bank said it may be, quote, less than well capitalized, end quote, and that it was, quote, re-evaluating its business, end quote, and noted that it expected to, quote, record further losses related to the other than temporary impairment on the securities portfolio, end quote. Silvergate shares hit an all-time low in pre-market trading this

Starting point is 00:11:28 morning, end quote. Snap plans to let Snapchat users pause streaks by buying streak restores for 99 cents each time and will soon let Snapchat plus subscribers pause streaks indefinitely. So in a way, Snap is getting into the insurance business, quoting TechCrunch. Maintaining Snap Streaks, where you send a snap to your friend once every 24 hours, is really important to a lot of Snapchat users. To bank on that fact, the company is now allowing users to pause their snap streaks so they don't have to worry about breaking them if they decide to not access the app for a while. Snapchat will let users pause only one streak for free during this test phase. A lost snap streak doesn't mean your friendship has gone cold, so starting today we're making

Starting point is 00:12:18 it easier to take a break with a new feature we're testing to let you reignite the spark and restore one streak for free with just one tap. The company said in an announcement, the company said that users will have, quote, the option to add more streak restores right from the app, and that will cost 99 cents per pause in the U.S. Snap also mentioned that it will soon roll out the ability for Snapchat plus users to freeze streaks completely. So essentially, people will have to pay either way to maintain streaks. During its Investor Day event in February, Snap said it now has more than 750 million

Starting point is 00:12:52 monthly active users with over 2.5 million paid users, end quote. So more on that, actually, if diversifying away from just ads to also, subscription products is the hotness now for everyone from Snap to Twitter. Who's doing it best? Well, according to Aptopia, so far in 2023, TikTok has generated $205 million more in-app purchase revenue than Facebook, Instagram, Snapchat, and Twitter combined. Quoting Forbes. Before iOS 14.5, which introduced Apple's privacy protections, big social networks made almost all their money via advertising. More user data meant better targeted ads and better targeted ads. generated higher revenue multiples, especially for Facebook. With less data available, ad relevance has

Starting point is 00:13:43 suffered and ad revenue has taken a hit. Collectively, Big Social appears to be trying to replace that loss revenue via direct payments from its users. Now, Snap offers Snapchat Plus for $40 a year and Twitter offers Twitter Blue for $115 a year. Both programs offer premium access to the platforms and exclusive features for subscribers. Facebook and Instagram primarily offer payments for followers and fans to reward their favorite creators or to boost the visibility of posts. TikTok primarily focuses on creator rewards, and it's doing better than all the other social networks combined. TikTok's app revenue has grown for seven consecutive quarters. Blacker says TikTok has generated $205 million more than Facebook, Instagram, Snapchat, and Twitter combined via IAP revenue.

Starting point is 00:14:27 So far in 2023, end quote. Microsoft is fighting back. Just last month, Mark Zuckerberg announced a paid verification program via his Instagram account. Users will pay $12 a month via web payments or $15 a month if they subscribe in-app to get a meta-verified badge on their accounts, along with some safety improvements and additional visibility or reach on the platform. The company took in $56 million in-app purchases, according to Apptopia, down from highs in previous years, but showing improvement, especially on Instagram. Instagram itself is hitting monthly in-app revenue of nearly $1 million in February. Twitter pulled in almost $900,000 in February, small potatoes perhaps, but also likely the beginning of something bigger. Snap has been on this journey longer and is making much more. Daily revenue from Snapchat is hitting about $125,000 a day and growing steadily in-2020.

Starting point is 00:15:18 TikTok, though, is the real giant here. In Q4 of 2022, TikTok generated over $350 million in-app revenue. In 2020, the same quarter hit just $150 million. Both numbers pale before its full year. revenue from last year. But, quote, TikTok has had IAPs in-app purchases since its very beginning, and its app revenue last year was a whopping $1.5 billion, Blacker says. Its IAPs are similar to Facebook's in that users pay for coins, which can be used to tip and pay for things from their favorite creators, end quote. It's important to note that in-app revenue is really only material for TikTok. Essentially, the small sums that Facebook, Instagram, Twitter, and Snap are bringing in via in-app purchases are rounding errors for those companies overall revenues. However, the example

Starting point is 00:16:03 that TikTok is setting shows Facebook and the other social networks that in-app payments and purchases are a very real and very significant revenue opportunity if they can make them work out on their platforms, end quote. And finally today, I don't really know what this means, or if it means anything material long term, but apparently U.S. regulators have rejected Elon Musk's bid to test brain chips from Neurrelink in humans, citing dozens of issues that Neurlink would need to resolve before human testing could begin, including the device's battery and removal. You know, you want to be able to get the thing back out at some point, quoting Reuters. On at least four occasions since 2019, Elon Musk has predicted that his medical device company Neurrelink would soon

Starting point is 00:16:54 start human trials of a revolutionary brain implant to treat intractable conditions such as paralysis and blindness. Yet the company founded in 2016 didn't seek permission from the U.S. Food and Drug Administration or FDA until early 2022, and the agency rejected the application, seven current and former employees told Reuters. The rejection has not been previously reported. In explaining the decision to Neurlink, the agency outlined dozens of issues the company must address before human testing a critical milestone on the path to final product approval, the staffer said. The agency's major safety concerns involved the device's lithium battery, the potential for the implant's tiny wires to migrate to other areas of the brain, and questions over whether and how the device can be

Starting point is 00:17:37 removed without damaging brain tissue, the employees said. A year after the rejection, Neurrelink is still working through the agency's concerns. Three staffers said they were skeptical the company could quickly resolve the issues, despite Musk's latest prediction at a November 30th presentation, that the company would secure FDA human trial approval this spring. Such FDA rejections do not mean a company will ultimately fail to gain the agency's human testing approval, but the agency's pushback signals substantial concerns, according to more than a dozen experts in FDA device approval processes. The rejection also raises the stakes and the difficulty of the company's subsequent requests for trial approval, the experts said. The FDA says it has approved

Starting point is 00:18:17 about two-thirds of all human trial applications for devices on the first attempt over the past three years. That total rose to 85% of all requests after the second review. But firms often give up after three attempts to resolve FDA concerns rather than invest more time and money in expensive research, several of the experts said. Companies that do secure human testing approval typically conduct at least two rounds of trials before applying for FDA approval to commercially market a device, end quote. So a quick little update on the ads situation. Ads are back, people. It's almost like the switch has been flipped back into the on position almost overnight. Now, why am I telling you this? It's not like you have any great love for this show having

Starting point is 00:19:06 advertisements, and the cupboards being bare really only affected my peace of mind, not yours, but I'm sharing this for two reasons. Number one, just a small bit of single anecdotal data that the tech recession slash winter might be thawing. And number two, if you heard the ads today, hopefully you heard some of the ads from listeners who stepped in last month to help me keep the lights on. Since the days where I have unsold inventory seems to be coming to an end, I did a thing where since today and the rest of this week had some open inventory, I digitally inserted some of those reads from listener sponsors from the last few weeks to give them a few thousand additional impressions as a way of thanking them for helping me out.

Starting point is 00:19:51 Hopefully it rotated evenly between the ads and the ads that you heard, We won't necessarily match what others heard, but everyone got some extra exposure somewhat evenly. Anyway, again, just doing this to thank those who stepped in to help us out. So step up and patronize them if you're so inclined. Talk to you tomorrow.

Tech Brew Ride Home - Thu. 03/02 – And Right On Schedule, Here Come The APIs

There aren't comments yet for this episode. Click on any sentence in the transcript to leave a comment.