Latest / The AI Edition / Meta Killed Its Own AI in 72 Hours - And Apple Sued OpenAI
Transcript
- Hello again, you wonderful people... we've had another big few days in AI, and there's a real doozy up top. Apple just filed a lawsuit accusing OpenAI of stealing its hardware secrets... through a former Apple engineer who, according to the filing, texted a colleague that breaking into old company files was, quote, "so funny." Apple's lawyers are not laughing. That's coming up at number one. But first, here's what else we've got coming up... Higgsfield can now build you an entire working app from a single sentence, ByteDance's newest model generates thirty straight seconds of 4K video in one shot, a survey of six thousand tech workers says AI is now reshaping how people feel about their jobs more than their own managers do, Gemini just hopped into your Waze app, and a well-known AI safety researcher just walked back his own doomsday number in a very big way. I'm Parker Gate, and this is The AI Edition from Collata Media... the AI stories that matter, ranked and all in one place. Here are the ten stories that actually matter this week... let's get into the countdown. Kicking off the countdown at number ten... this one's for the creators in the room. Higgsfield — the AI video and image studio a lot of you already use — shipped something called Apps on July ninth. Here's the pitch... you describe the tool you want in plain chat, and it builds you a working generative app end to end: the design, the code, the database, and Higgsfield's own image and video models, all wired in automatically. No coding. So instead of generating one image at a time, you could spin up, say, a little app that turns your product photos into ad variations on demand. To launch it, they're running an App Contest with a hundred-thousand-dollar prize pool, open until July twenty-first. Why it matters... this is the "generate a whole tool, not just a single output" idea showing up in a platform creators actually touch every day. If you're a marketer or a solo operator who's been copy-pasting between five different AI tools, this is a bet that you can just describe the workflow and get an app that runs it. Coming in at number nine, and staying in creative tools... ByteDance's Doubao team officially rolled out Seedance two-point-five in early July, and it moves the bar on AI video length. The headline... it generates a native thirty-second clip in 4K, in one shot, without stitching separate segments together — which has been the awkward workaround for basically every AI video tool until now. It also takes up to fifty reference images to keep your characters and products consistent, and it generates its own audio, no separate sound pass. Why it matters... thirty continuous seconds is about the length of an actual ad or a social spot, so this is AI video creeping from three-second novelty clips toward something you could genuinely cut a campaign from. If you make short-form video for a brand, this is the model to test next. At number eight... something that sounds like science fiction, but is a real lab result. On July thirteenth, researchers demonstrated a new way to use quantum computing... machines that process information in a fundamentally different way than your laptop... to help design peptides, which are basically tiny protein fragments, for new drugs. They used thirty-two optical channels and a technique called Gaussian boson sampling, essentially a specialized way of using particles of light to solve tricky math problems, to seed the search for promising molecules. This builds on work from back in March, when Cleveland Clinic and IBM used a quantum computer to simulate a miniature protein made of three hundred and three atoms. And in a separate project published in Nature Biotechnology, a similar quantum approach helped propose cancer drug candidates... fifteen molecules were actually synthesized, and two showed real promise. Now, to be clear, this is a demonstration, not an approved drug, and the whole quantum drug discovery market is still small, projected to grow from around one hundred twenty six million dollars this year to about six hundred thirty eight million by twenty thirty five. But if you're wondering where quantum computing's first real payoff might show up in ordinary life... drug discovery for rare diseases is looking like a pretty good bet. Number seven, and this one's for anyone who leans on ChatGPT to actually get things done. OpenAI rolled out two new modes... Work mode and Codex mode... built on the new GPT-5.6 model, which came out of a limited preview with about twenty government-approved partners before going fully public on July ninth. Work mode is built for multi-step tasks, with access to your files, your apps, and a built-in browser. Codex mode is aimed squarely at building websites. And here's the number that matters if you're the one paying the bill... the most efficient version, called Sol, is fifty-four percent more token-efficient for coding tasks, according to Sam Altman, which basically means it gets more coding work done for less money per request. If you're a marketer or a small operator who's used ChatGPT for one-off questions, this is OpenAI pushing it toward handling entire workflows instead. Number six is a genuinely striking one. A researcher named Davidad, who used to run a fifty-nine million pound UK government program called ARIA Safeguarded AI, has spent years publicly estimating the odds that AI leads to catastrophe for humanity. Back in twenty twenty-two, that number was north of seventy percent. This week, on a podcast appearance, he said it's now under five percent. His reasoning... he's been running what he calls wisdom probes on newer models, and he says systems like Gemini two point five Pro and Claude's Opus four have started answering in ways that read, to him, as more aligned and less deceptive than earlier models. That's one respected researcher's personal reassessment based on his own testing, not a scientific consensus... but going from seventy percent down to under five is about as big a swing in tone as you'll hear anywhere in AI safety circles right now. Number five moved fast. On July seventh, Meta launched what it called its most advanced image model yet, Muse Image... and buried in that announcement was a feature that let you, right inside the Meta AI app, @-mention any public Instagram account and pull that person's actual likeness straight into an AI-generated image. The catch... it was opt-out, not opt-in. If you didn't want strangers generating images of you, you had to go turn it off yourself, or lock your profile. That did not go over well. The talent agency CAA said no one's name, image, or likeness should be used by an AI model without clear, documented consent. SAG-AFTRA went further, calling anything short of a clear opt-in unacceptable, and told its members to protect their likeness. And by Friday the tenth... roughly seventy-two hours after launch... Meta pulled it, saying, quote, "we've heard the feedback that this feature missed the mark, so it's no longer available." Why it matters... a company the size of Meta reversed a flagship AI feature in three days because creators and unions pushed back. Worth remembering the next time a new AI feature quietly turns itself on for you. Number four is quieter, but it tells you where OpenAI's head is at. The company posted a job listing for a product manager in San Francisco, specifically to build ChatGPT experiences for families, caregivers, and older adults. Look at the numbers behind it. ChatGPT users thirty-five and older made up thirty-one percent of its global user base last quarter, up from twenty-six percent a year earlier, while the eighteen to twenty-four crowd actually shrank. Almost one in four American parents with smartphones used ChatGPT last quarter, nearly double what it was a year before. Google's Gemini is actually ahead with parents right now, thirty-two percent reach versus ChatGPT's twenty-four. And here's a number worth sitting with... a survey of over four thousand families found thirty-eight percent of kids used generative AI in the past week, but only twenty-seven percent of their parents realized it. ChatGPT is quietly becoming a family tool, whether or not every parent in the house knows it yet. Number three takes us out of the house and into the car. Google is expanding Gemini inside Waze, and if you drive, this one's actually going to show up in your life. You can now just talk to the app... ask it to find a coffee shop that's open right now, or the gas station nearby with the lowest prices... and it figures that out conversationally instead of you typing an address. There's also an upgraded voice-based way to report traffic incidents, a new Less Chatty mode for people who want fewer interruptions, and a Motorcycle Mode launching first in places like Brazil, Mexico, and the Philippines, with the rest rolling out globally on both Android and iOS starting with Waze's beta community. It's AI showing up in one of the most-used apps on your phone, doing something pretty unglamorous... just making the thing you already do every day slightly less annoying. Number two is one a lot of you are going to feel personally. Lenny's Podcast ran its second annual tech worker sentiment survey, close to six thousand people across product, engineering, design, research, and marketing roles. The headline number... AI is now having three times the effect on how people feel about their professional identity than something as basic as whether their manager is actually good at the job. Fifty percent of people say AI makes them feel amplified... more capable, more effective. The other half describe feeling destabilized, or like their role is being redefined out from under them, or just plain diminished. Separately, a Conference Board survey of U.S. workers this June found about thirty-nine percent said AI tools improved their job satisfaction, and only around seven percent said it made things worse, even as overall job satisfaction hit the highest point in the Conference Board's thirty-nine years of tracking it. So this isn't a simple story of AI good or AI bad at work... it's that AI is now shaping how people feel about their jobs more than almost anything else in the workplace. Worth sitting with, whichever half you're in. And here's number one. Apple has sued OpenAI. Filed this week in federal court in Northern California, and it's about as direct as these things get. Apple's case centers on a former iPhone engineer named Chang Liu, who joined OpenAI back in January. Apple alleges he used a bug to get into confidential Apple files after he'd already left the company... and, according to the filing, texted a colleague that it was, quote, "so funny." But Apple isn't just going after one engineer. The lawsuit also names OpenAI's chief hardware officer, Tang Tan, a nearly twenty-four-year Apple veteran who left to join Jony Ive's hardware startup, io, before OpenAI bought that startup for six point four billion dollars back in May of last year. Apple says over four hundred of its former employees have now joined OpenAI. Apple is asking the court to stop OpenAI from using any of what it allegedly took, plus damages. OpenAI's response so far is that it has, quote, "no interest in other companies' trade secrets." All of this is unfolding while OpenAI builds toward its own Ive-designed hardware device, expected around twenty twenty-seven... so this lawsuit isn't just corporate friction. It's Apple putting a legal speed bump in front of one of the biggest hardware bets in the industry right now. That's the countdown. A lawsuit over hardware secrets, a safety researcher rewriting his own odds, and Gemini riding shotgun on your commute. See you next time. I'm Parker Gate. Stay informed. Stay sharp.