Loading summary
Transcript1 lines
- [00:05]
Brian McCullough
Welcome to the Tech Meme Ride home for Wednesday, July 30, 2025. I'm Brian McCullough. Today, age verification for web users is Sweeping the globe. ChatGPT debuts a study mode for students. What exactly is Zuckerberg trying to achieve by hiring everyone in AI? Maybe cohere is the real dark horse in the AI model race. And what if the real money in AI video is in training robots? Here's what you missed today in the world of tech. YouTube is rolling out age estimation technology in the US to identify teen users and serve more age appropriate content, regardless of the birthday those users might have given AT signup, quoting TechCrunch. When YouTube identifies a user as a teen, it introduces new protections and experiences which include disabling personalized advertising, safeguards that limit repetitive viewing of certain types of content, and enabling digital well being tools such as screen time and bedtime reminders, among others. These protections already exist on YouTube, but have only been applied to those who's verified themselves as teens, not those who may have withheld their real age. For instance, in 2023, YouTube began limiting repeated viewing of videos that could trigger body image issues or those that display social aggression. The company has also been developing digital well being tools since 2018. If the new system incorrectly identifies a user as under 18 when they're not, YouTube says the user will be given the opportunity to verify their age with a credit card, government ID or selfie. Only users who have been directly verified through this method or whose age has been inferred to be over 18 will be able to view the age restricted content on the platform. The machine learning powered technology will begin to roll out over the next few weeks to a small set of US Users and will then be monitored before rolling out more widely, the company says. The plans to introduce age inference technology were announced in February as part of YouTube's 2025 roadmap. The plans are also step in attempting to make YouTube safer for younger users following the 2015 launch of the YouTube Kids app and the 2024 rollout of supervised accounts. The features also arrive as social media more broadly is coming under increased government scrutiny in the United States, where platform makers including Apple and Google have pitted their lobbyists against those from big tech companies like Meta over who's responsible for age verification and children's safety. In the meantime, a handful of US States have taken matters into their own hands as over a dozen states have passed or proposed laws to regulate minors use of social Many of these require age verification or parental consent, including those in Louisiana, Arkansas, Florida, Georgia, Utah, Texas, Maryland, Tennessee and Connecticut, among others. However, some laws like those in Utah and Arkansas are blocked by litigation at this time and are not enforceable, while others are still pending implementation. End quote. Yeah, the whole age verification thing is a trend sort of sweeping the globe right now. Australia now says it will include YouTube in its ban on social media for children under 16 years of age, reversing an earlier decision to exempt the platform. And of course, you might have heard you have all of that recent news about the UK enforcing those new online child safety laws requiring websites that host porn, self harm, suicide and eating disorder content to verify users ages. Speaking of younger users, OpenAI has debuted a ChatGPT study mode designed to guide students through learning material at their level. Available to logged in free +Pro and team users. Quoting Wired, the mode is designed around the Socratic method, so when activated, OpenAI's generative AI chatbot rejects direct requests for answers, instead guiding the user with open ended questions. OpenAI has significantly disrupted the education system over the past few years, with students becoming some of the earliest adopters of ChatGPT. Even so, OpenAI claims the bot is currently an overall boon to learners if asked to roleplay as a synthetic tutor. When ChatGPT is prompted to teach or tutor, it can significantly improve academic performance, says Leah Belsky, a vice president of education at OpenAI. But when it's just used as an answer engine, it it can hinder learning. The problem is, no matter how engaging ChatGPT's study mode becomes, as OpenAI iterates on this feature, it exists just as a toggle click away from ChatGPT itself, with direct answers and potential fabrications about whatever class you're working on. That could be quite hard to resist for younger users still developing their frontal lobe. It's true that students on the hunt for easy ways to avoid engaging with the substance of a course have always had resources available to them, like the CliffNotes series of literature summaries. Still, the immediacy and personalized nature of chatbots feels like an escalation. Multiple AI focused smartphone apps that can solve homework problems with just a snapshot. Like ByteDance's gouth rocket in popularity whenever the school year gets back into session, many educators have recently raised concerns about the continued and often secretive use of AI by students. OpenAI CEO Sam Altman doesn't buy it. I remember when I was in school, junior high, Google first came out and all the teachers freaked out, altman said on a recent podcast. Similar to the Internet and the calculator Altman sees AI as a tool capable of helping you think better. ChatGPT's study mode is an attempt to foster more thorough engagement on a topic with users by throwing questions back at them and asking for more context about their learning goals. Instead of giving a very long, drawn out answer up front, it's first asking you, hey, what are you trying to optimize for? What's your current level? Says Abhi Muchal, who works on the product team at OpenAI. While the launch of OpenAI's study mode is more focused on universities, with college students giving their best beta tester testimonials during the press briefing, the company has its attention on an even broader swath of learners, which includes younger students. OpenAI is currently partnering with learning experts from Stanford to quote, study and share how AI tools, including Study Mode, influence learning outcomes in areas like K12 education, reads the company's announcement blog. This release comes not too long after the Trump administration's executive order focused on getting more AI usage into classrooms in the United States. Even if the initial research that OpenAI participates in about education and AI supports the company's claims that students learn better with a bot by their side when it's used as a tutor rather than an answer generator, I'm still concerned. What are the long term impacts of turning to an AI tool for guidance with increasing regularity? It's still unclear if young people who may grow up constantly asking ChatGPT for help develop an over reliance on the software that impedes critical thinking. So while ChatGPT study mode is designed to guide students through learning material at their level of understanding, the onus remains on users to engage with the software in a specific way, ensuring that they truly understand the material. In the AI age, the hardest challenge of all for students might just be resisting the urge to swap out of Study Mode. Snap a photo of the homework question and have ChatGPT tell them exactly what you want to here, which is the answer, end quote Wait, we've gone entire days now without discussing Mark Zuckerberg's big AI hiring spree? Well, let's fix that toot suite first. Mark Gurman says that Apple has lost its fourth AI researcher in a month to Meta's new Superintelligence team. He says the Apple foundation model team is in flux and Apple marginally increased AFM staff pay in an attempt to compensate. Quote Bowen Zhang, a key multimodal AI researcher at Apple, left the company on Friday and is set to join Meta's recently formed Superintelligence team. According to people familiar with the matter. Zhang was part of the Apple Foundation Model Group, or afm, which built the core technology behind the company's AI platform. Meta previously lured away the leader of the team, roming pang with a compensation package valued at more than $200 million, Bloomberg News has reported. Two other researchers from that group, Tom Guntner and Mark Lee, also recently joined Meta. AFM is made up of several dozen engineers and researchers across Cupertino and New York. In response to the job offers from Meta and others, Apple has been marginally increasing the pay of its AFM staffers, whether or not they've threatened to leave, said the people, who asked not to be identified because the moves are private. Still, the pay levels pale in comparison with those of rivals. The departures have thrown Apple's models team into flux. Paying played a central role in defining the department's roadmap and research direction, and multiple people within AFM now say its future is unclear. Additional engineers are actively interviewing for jobs elsewhere, according to the people. Another team member, Floris Weirs, left for a startup in recent weeks. The AFM team is critical to Apple's broader AI strategy. The group's work underpins the Apple intelligence platform, which launched last year, but now the company is considering a shift toward using more third party models. Some Apple executives see its homegrown models as a stumbling block to catching up with AI rivals, the people said. And the uncertainty over whether to outsource the technology has hurt morale at the company and helped fuel the attrition. End quote. Geez, can Apple do anything right when it comes to AI? Then a source tells Wired that Meta approached over a dozen people at Thinking Machines Lab, that startup that Mira Moradi started, and one offer was for more than a billion dollars over a multi year span. You heard that right, A billion dollars. But interestingly, apparently not a single person has left Thinking Machines for Meta, quoting Wired. The rest were offered between 200 and 500 million dollars over a four year span, multiple sources confirm. In the first year alone, some staffers were guaranteed to make between 50 and $100 million, sources say. Zuckerberg's initial outreach is low key, according to Messages View. In some cases, he sent recruits a direct message on WhatsApp asking to talk. From there, the interviews move fast. A long call with the CEO himself, followed by conversations with Chief Technology Officer Andrew Boz Bosworth and other Meta executives. Here's a pre Meta Superintelligence Labs recruiting message Zuckerberg sent to a potential recruit the tone hasn't changed much today. Apparently quote We've been following your work on advanced technology and the benefits of AI for everyone over the years. We're making some important investments across research, products and our infrastructure in order to build the most valuable AI products and services for people. We're optimistic that everyone who uses our services will have a world class AI assistant to help get things done. Every creator will have an AI their community can engage with, every business will have an AI their customers can interact with to buy things and get support and every developer will have a state of the art open source model to build with. We want to bring the best people to Meta and we would love to share more about what we are building end quote during these conversations, Baz has been upfront about his vision for how Meta will compete with OpenAI. While the tech giant has lagged behind its smaller competitor and building cutting edge models, it is willing to use its open source strategy to undercut OpenAI. Sources say the idea is that Meta can commoditize the technology by releasing open source models that directly Compete with the ChatGPT maker. End Quote and finally, what is the whole end game here exactly? Quoting the FT Mark Zuckerberg's aim, according to people familiar with his thinking, is to create a startup like unit within Meta focused on developing advanced AI technology that is unencumbered by the bureaucracy of the larger $1.8 trillion company but supported by its well capitalized business. Zuckerberg has set up the unit in a separate workspace at the company's Menlo Park, California headquarters, isolated from the rest of its AI efforts, and has met the team members as they joined in recent weeks. Heading up the lab are Alexander Wang, the former Scale AI chief executive and ex GitBoss hub Nat Friedman. The lab's work has been closely guarded from the wider company. Investors and even insiders have little insight into its vision and budget as Meta is expected to post its slowest profit growth in two years when it reports its second quarter earnings. On Wednesday, a person close to the lab said no clear strategy had been shared with staff as the team was still being built and remained walled off within Meta. It's like the Manhattan Project. They are throwing all their cash at AI and trying to work out what to do, said Uday Sharuvu, portfolio manager and analyst at Asset Manager Harding Lovner, which invests in Meta. What are the products you're going to build? It can't just be a moonshot. What is the business model? He added. If there isn't a commercial plan that we as investors can see then investors will get impatient. Developing superintelligence is now in sight, zuckerberg said in a memo on Wednesday, hours after this article was first published. He added that the company was focused on personal superintelligence, through which people will have greater agency to improve the world in the directions they choose, rather than directing the technology solely toward automation and productivity gains, End quote In the AI horse race, the argument could be made that the real dark horse is Cohere, the Canadian AI startup. Apparently, Cohere has told investors it expects to generate more than $200 million in annualized revenue by the end of this year, up from just 70 million in annualized revenue as of February. Quoting the information, Cohere is in talks with investors to raise hundreds of millions of dollars in equity financing at a $6.3 billion valuation before the funding, according to three people with knowledge of the fundraise. That would be a roughly 25% increase in its valuation compared to its last equity financing a year ago. The company aims to raise between 300 and $500 million of equity, said a fourth person familiar with the talks. Cohere sells AI models and applications to companies and government agencies. While it has struggled to keep up with Anthropic and OpenAI and fell far short of its earlier financial goals, it is continuing to grow, reflecting the AI market's fast global expansion. Cohere has told investors it expects to generate more than $200 million in annualized revenue, calculated by multiplying the most recent month's revenue by 12 by the end of this year, according to Two. That pace would represent a near threefold rise from $70 million in annualized revenue as of February. Most companies would envy such growth, but Anthropic's fourfold growth to $4 billion in annualized revenue over the last six months looks even more impressive. However, Anthropic is likely losing a lot more money than Cohere, though it isn't clear which company is operating more efficiently, Cohere's cash burn rate couldn't be learned. Cohere this year said it had clinched deals with enterprise customers in Canada such as Telecom, Bell Canada and Royal bank of Canada. It hit its $200 million annualized revenue goal sooner than year end due to recent eight figure contracts it has been signing with new customers, said one of the people. The company has also told investors IT projects hitting 4 to $5 billion in annualized revenue by 2029, according to one of the people. End quote. Finally, today, the video AI companies are having a moment quoting the information. Imax on Monday announced that it would screen the winning films from Runway's AI Generated Film Festival at some of its theaters. Netflix recently said it had used AI to generate a scene in one of its shows for the first time, a building collapsing in the eternaught, an Argentine sci fi show. Both announcements show that video generating AI software startups such as Runway and Luma AI are making inroads in Hollywood. Now those startups are eyeing a more futuristic source of revenue, training robots. Both Runway and Luma AI are in talks with robotics and self driving car companies that are interested in using the video models to improve their robots, according to executives for both companies, Luma AI expects robotics to eventually be one of the largest sources of revenue for the company, CEO Amit Jain told me. Likewise, Runway expects that generating video clips in real time for robotics as well as video games will eventually dwarf revenue from generating footage for a studio to later incorporate into a film, said the company's co founder and cto, Anastasis Jaramandius. We previously reported that Runway hopes to hit $300 million in annualized revenue this year. You might think of Runway as a company that's mostly working with media, with Hollywood, with videos, said Cristobal Valenzuela, CEO and co founder of Runway, in an interview with the Informations TI tv. But the thing is that many of the underlying capabilities and features of the models themselves are also very useful and applicable in a wide domain of industries. Roboticists have long turned to computer simulations to train the AI models that serve as the brains for robots. But these programmed simulations have never been realistic enough to prepare robots for many scenarios they might encounter in the real world. For example, these simulations have often treated boxes as though they're rigid, so a robot trained in such a simulation would struggle to pick up the flexible envelopes that e commerce companies use. The video AI models, in contrast, can generate clips with complicated physics such as water sloshing around or clothes bending, making them better teachers for robots, say the company's founders. For instance, Runway's AI model could generate video clips of what a self driving car might see if it turns left at an intersection versus what it would see if it turns right. A self driving car company could use those simulations to figure out which turn is less likely to result in a crash than train its cars to take that action. End quote. Reminder again to listen to the end of tomorrow's episode for a major, major podcast announcement. Talk to you then.