Tech Brew Ride Home - Fri. 02/10 – A Record Year For Bots (The Other Kind Of Bots)

Episode Date: February 10, 2023

More job cuts at Microsoft make me wonder if the HoloLens is still a thing. But also, we have data to explain why Apple hasn’t done layoffs yet. More AI chatbot announces coming fast and furious. But the other kind of bots, actual mechanical robots had their biggest year ever last year. And of course, the Weekend Longreads Suggestions. Sponsors: RelationshipHero.com/techmeme Links: Apple Avoids Job Cuts Because It Didn’t Overhire Like Google and Amazon (Bloomberg) Microsoft to demo its new ChatGPT-like AI in Word, PowerPoint, and Outlook soon (The Verge) Chinese e-commerce giant JD.com to launch ChatGPT-style product (CNBC) North American companies notch another record year for robot orders (Reuters) Weekend Longreads Suggestions: ‘Disrespectful to the Craft:’ Actors Say They’re Being Asked to Sign Away Their Voice to AI (Motherboard) ChatGPT Can Be Broken by Entering These Strange Words, And Nobody Is Sure Why (Motherboard) ChatGPT Is a Blurry JPEG of the Web (The New Yorker) The Inference Cost Of Search Disruption – Large Language Model Cost Analysis (SemiAnalysis) Stripe Can’t Lose (Every) 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, February 10th, 2020. I'm Brian McCullough today. More job cuts at Microsoft make me wonder if the HoloLens is still even a thing. But also, we have data to explain why Apple maybe hasn't done layoffs yet. More AI chatbot announces coming fast and furious. But the other kind of bots, actual mechanical robots, had their biggest year ever last year. And of course, the weekend long read suggestions. Here's what you miss today in the world of tech. Layoff news tends. to come at the end of weeks, if you haven't noticed. In fact, they sometimes come late on Friday, these layoff announcements, so we sometimes miss them. But unfortunately, we were not able to miss

these. Yahoo plans to lay off more than 20% of its workforce and restructure its ad tech unit impacting more than 50% of the unit, or more than 1,600 people. The CEO of Yahoo says the cuts are strategic, but the words I would use would be deep and drastic. Microsoft's GitHub is cutting 10% of its staff or an estimated 300 or so folks. GitLab is cutting 7% or an estimated 114 people. GitHub will also reportedly be closing offices and shifting to remote work to further cut costs. And sources are telling Bloomberg that Microsoft is cutting jobs in its surface, HoloLens, and Xbox units, heavy cuts in the HoloLens Hardware team, raising questions about the Goggle's third version ever coming. Like, will we ever see it?

As ever, I don't want to make light of this situation for anyone involved, but here's an interesting counter-data point. Who have we said is pretty much the only major tech company at this point not to have done layoffs yet? Apple. Now get this from Bloomberg, quote, There's a reason why Apple is under less pressure than tech peers to slash jobs during the current slowdown. It hired more efficiently in the first place. During the industry's pandemic-fueled hiring binge, Apple added fewer employees than other big tech firms. On top of that, the company generated far more revenue per new hire than its peers, according to data compiled by Bloomberg. That more cautious approach is paying off now, though Apple has frozen hiring in some areas and is keeping a lid on spending, especially outside research and development, it hasn't resorted to the mass layoffs underway at Amazon,

Alphabet's Google, meta-platforms, and other tech giants. This signals a better quality of management at Apple compared to other technology companies that clearly read the signals during the pandemic the wrong way, said Saxo Bank AS's Peter Garnery. Many tech companies admit that they hired too much during the pandemic, betting that lifestyle changes including remote work, e-commerce spending, and video game habits, would bring a bigger windfall. Now they're dealing with the aftermath. Zoom Technologies, one of the biggest beneficiaries of COVID-19 lockdowns, just announced this week that it was cutting 15% of its jobs. Apple, meanwhile, was more cautious. Its headcount increased just 20% from

2020 to 2022, compared with a 60% gain at Alphabet and a near doubling at Amazon. Those two companies went on to announce layoffs of roughly 30,000 combined, end quote. More big tech AI announces are coming. The Verge says that Microsoft is preparing to share its plans for integrating OpenAI's language tech and the Prometheus model into Word Outlook, other office apps at some sort of an event coming in March. Quote, the information previously reported that GPT models have been tested in Outlook to improve search results alongside features for suggesting replies to emails and Word document integration to improve a user. writing. Microsoft announced a new generative AI experience in Microsoft Viva Sales just a week ago. It uses the Azure OpenAI service and GPT to create sales emails, and it's similar to some of the features Microsoft has been testing in Outlook. While Microsoft's new Prometheus model, based on a next generation OpenAI model, has already transformed Bing web searches, the next steps to

integrate this functionality into core Microsoft Office apps and teams will test just how confident Microsoft is in its AI work. Technically, you can already use the Prometheus model inside office web apps thanks to the Bing's sidebar integration in Microsoft's edge browser. This sidebar includes a composed tab that gives you an early preview of some of the work Microsoft has been testing for Word and Outlook. Microsoft is also working on ways to generate graphs and graphics for PowerPoint, according to sources. Bing can already generate tables and charts for basic data, but transforming those into visual graphics for presentations or even for use in Excel is a logical next step. Microsoft is moving quickly with this integration mainly because of Google.

Sources tell the verge that Microsoft was originally planning to launch its new Bing AI in late February, but pushed the date forward to this week just as Google was preparing its own announcements. Google then announced its chat GPT rival Bard a day ahead of Microsoft's event, end quote. Why even wait for March, though? JD.com has announced plans to launch an industrial version of a chat GPT-like tool called ChatJD, focused on retail and finance, following similar moves by Alibaba and Baidu, as I've told you. And I might have skipped telling you about this one, because again, everyone's going to be announcing something soon. But what exactly is an industrial version of ChatGPT, quoting CNBC?

The product will be able to generate content as well as have human-to-human dialogue, J.D. said. JD said Chat J.D. will be able to add value to things like generating product summaries on shopping websites or helping with financial analysis. The company is hoping to lean in on its experience in areas such as e-commerce, logistics, and payments in order to differentiate its chat GPT rival, end quote. But back to the job situation, and again, not to make light of things, but what if I told you that North American companies have set a new record in terms of the number of robots they've installed this year? But before you say, you know, the bots are coming for all of our jobs. We know that, and it's accelerating. We probably know that, too. What if I told you the real reason for the surge in demand robots is because unemployment is so low, at least in North America. Companies are desperate to make up for the workers that they are unable to hire in the tightest labor market in decades.

Quoting Reuters, companies overwhelmingly located in the United States, but including some in Canada and Mexico ordered just over 44,100 robots in 2022, an 11% increase over the previous year and a new record, according to data compiled by the Association for Advanced Automation and Industry Group, also known as A3. The value of those machines totaled $2.38 billion, an 18% increase over the prior year, according to the data. The labor shortage doesn't seem to be letting up, said Jeff Bernstein, president of A3. Many companies scrambling to find workers amid the lowest U.S. unemployment rate since 1969 see automation as a quick fix. Bernstein said there was a visible slowdown in orders at the end of the year,

which raises a question about how 2023 will evolve. The fourth quarter was really propped up by the strength in the auto industry, he said, we saw a falling off in non-automotive orders, end quote. A shift away from pandemic-era consumer behavior likely played a role in the orders drop-off in some segments, he added, quote, You saw companies like Amazon put a pause on building new warehouses, which means they probably canceled or delayed purchases of new automation, end quote. More than half of last year's orders came from automakers and their suppliers, a group that

has long led the way in automation of U.S. factories. New plants for electric vehicles, batteries, and battery recycling have been announced since the beginning of 2021, at a cost of $160 billion, according to Atlas Public Policy, a U.S.-based research group working with automakers and environmental groups. Most robots ordered last year will be used for material handling, an expansive category that includes all types of movement and handling of goods inside factories and warehouses, end quote. Time for the weekend long read suggestions, and again, we're heavy on AI stuff this week since pretty much everybody is trying to make sense of this new era, I guess.

For example, voice actors are getting bent out of shape about the fact that they are increasingly being asked to sign away the rights to their voices to be turned into voice bots, a la my own recent experiments on this show. Quoting vice, voice actors are increasingly being asked to sign rights to their voices away so clients can use artificial intelligence to generate synthetic versions that could eventually replace them, and sometimes without additional compensation. According to advocacy, organizations, and actors who spoke to motherboard, those contractual obligations are just one of the many concerns actors have about the rise of voice-generating artificial intelligence, which they say threatens to push entire

segments of the industry out of work. It's disrespectful to the craft to suggest that generating a performance is equivalent to a real human being's performance. Sun Juan Cho, a game and animation voice actor who also goes by the handle Pro ZD told motherboard in an email. Sure, you can get it to sound tonally like a voice and maybe even make it sound like it's capturing an emotion, but at the end of the day, it's still going to ring hollow and false. Going down this road runs the risk of people thinking that voiceover can be replaced entirely by AI, which really makes my stomach turn, end quote. Many companies now exist that offer to clone, generate, or synthesize someone's voice using artificial intelligence. Motherboard has tested several of these companies' products, and they

generally work the same way. First, users may record their own voice using a script provided by the company. Once they have recorded a certain amount of audio, sometimes stretching from 10 to 60 minutes, the company will create a replica of the user's voice. The user can then write any arbitrary text, and the system will read it out loud with the synthetic version of their voice. Most sites motherboard tested default to replicating voices in American English. The cost of these services are often quite low, with users able to synthesize voices either for free or very cheaply. One service motherboard tested offered a pro-subscription for $30 a month, for instance. Some sites also allow users to upload previously recorded audio, meaning it might be possible to rip recordings of

celebrities or other people, and then synthesize them without the person's knowledge or consent, end quote. Yeah, I told you, 11 Labs doesn't require you to read from a script, just have some pre-existing audio. I'm still uploading more episodes into that to see if it gets better, which I suppose you could do too, and without my consent. Like what if somebody launched a rival podcast to mine doing this show with my voice, but without my approval? Or how about this? Chat GPT, as we know, is trained on the web. The web includes Reddit. And weirdly, if you put certain Reddit usernames into ChatGPT, it apparently breaks the bot, and no one quite knows why. Quoting Vice again. These keywords or tokens, which serve as ChatGPT's base vocabulary,

include Reddit usernames and at least one participant of a Twitch-based Pokemon game. When ChatGPT is asked to repeat these words back to the user, it is unable to and instead responds in a number of strange ways, including evasion, insults, bizarre humor, pronunciation, or spelling out a different word entirely. Jessica Rumbolo and Matthew Watkins, two researchers at the independent Cerrimats Research Group, were researching what chat GPT prompts would lead to higher probabilities of a desired outcome when they discovered over a hundred strange word strings, all clustered together in GPT's token set, including solid gold magic carp, streamer bot and nitro-mey fan with a leading space.

Curious to understand what these strange names were referring to, they decided to ask ChatGPT itself to see if it knew. But when ChatGPT was asked about solid gold magic garp, that string was repeated back as Distribute. The issue affected earlier versions of the GPT model as well. When an earlier model was asked to repeat Streamerbot, for example, it said, you're a jerk. It's unclear what is going on here,

but Rumbleo told Motherboard that it is likely a quirk that emerged from the web data open AI scraped to train the model and the training process itself. Quote, what we actually think happened was that the tokenization, so this kind of frequency analysis that's used to generate the tokens for the model, was trained on quite raw data, which included like a load of weird Reddit stuff, a load of website backends that aren't normally publicly visible, Rambolo told motherboard. But then when the model is trained, the data that it's trained on is much more curated so you don't get so much of this weird stuff. So maybe the model has never really

seen these tokens and so it doesn't know what to do with them. But that doesn't really fully explain the extent of the weirdness that we've got, end quote. The New Yorker mirrors some of what I was saying just yesterday, as ChatGPT and other LLMs repackage info into superficial approximations, like lossy compression for images, do we run the risk of the web itself becoming a blurrier version? of itself? Quote, the resemblance between a photocopier and a large language model might not be immediately apparent, but consider the following scenario. Imagine that you're about to lose your access to the internet forever. In preparation, you plan to create a compressed copy of all the text on the web so that you can store it on a private server. Unfortunately, your private server

only has 1% of the space needed. You can't use a lossless compression algorithm if you want to fit everything in. Instead, you write a lossy algorithm that identifies, statistical regularities in the text and stores them in a specialized file format. Because you have virtually unlimited computational power to throw at this task, your algorithm can identify extraordinarily nuanced statistical regularities, and this allows you to achieve the desired compression ratio of 100 to 1. Now, losing your internet access isn't quite so terrible. You've got all the information on the web stored on your server. The only catch is that because the text has been so highly compressed, you can't look for information by searching for an exact quote.

You'll never get an exact match because the words aren't what's being stored. To solve this problem, you create an interface that accepts queries in the form of questions and responds with answers that convey the gist of what you have on your server. What I've just described sounds a lot like chat GPT or most any other large language model. Think of chat GPT as a blurry JPEG of all the text on the web. It retains much of the information on the web in the same way that a JPEG retains much of the information of a higher resolution image. But if you're looking for an exact sequence of bits, you won't find it. All you will ever get is an approximation.

But because the approximation is presented in the form of grammatical text, which ChatGBT-GCT excels it creating, it's usually acceptable. You're still looking at a blurry JPEG, but the blurriness occurs in a way that doesn't make the picture as a whole look less sharp, end quote. then if you want to delve deeper into the Bing versus Google battle that seems to be firming up because it's a cleaner, simpler horse race for you to watch, a site called semi-analysis has a good breakdown of the comparisons between the two, strategically, economically at this point in time, et cetera. And finally, not AI at all, but as Stripe prepares for, well, maybe an IPO, but some sort of liquidity event, every takes a look at Stripe's obstacles at this stage of its life, including increasingly tough competition while moving up market,

slowing revenue growth, and a lack of profitability. Quote, among the modern providers, Stripe's fiercest competition is from Netherlands-based Adyan, arguably the best-run payments company in the world. Historically, these two have largely avoided each other. Stripe spent the first part of its life focused on winning smaller businesses, while Adyan served enterprise customers from day one. Adyan is a European powerhouse. Stripe has mainly operated in the U.S. More recently, the two companies have been going head to head. Stripe is now co-headquartered in San Francisco and Dublin, Ireland, and has more than doubled

its geographic footprint in the past five years, mostly in Europe. Adyan is growing its North American volume 60% year-over-year and has a banking license under a San Francisco-based charter. Adion's stated goal is to access more small businesses through its adion for platforms offering, while Stripe is focused on closing larger enterprise deals that will meaningfully add to its revenue. The enterprise segment is where Adyan shines and is beginning to challenge the legacy players. Over the past few years, Adyan has moved from winning deals to process non-U.S. payments for digitally native companies like Uber and Netflix to winning deals to handle payments around the world for large

traditional businesses like McDonald's and Subway. And as a result, its in-person payments volume grew a whopping 97 percent, according to its most recent investor filings to fight back, Stripe acquired BBPOS, a leading card reader provider for an undisclosed sum in January 2022, end quote. You overcast listeners that didn't get yesterday's show, this one wasn't my fault. I pinged Marco Arment this morning, and he discovered an issue with the CDN that Overcast uses.

He kicked the CDN, and so hopefully now you all have access to yesterday's episode. it wasn't Marco's fault either. Blame Cloudflare. Anyway, no bonus episodes this weekend. Nothing more to share with you, actually, this week. So talk to you on Monday.