What a Difference Three Years Makes
AI Inside for Wednesday, December 2, 2025
Lots of interesting stuff this week on the AI Inside podcast, including ChatGPT turning three amid a code red at OpenAI, Apple AI chief John Giannandrea retiring next spring, and AWS announcing Trainium3 at re:Invent.
But first, a shoutout to our Patron of the Week, Daniel Kroft (who helps me immensely behind the scenes), Daniel Patterson, and Brenda Muir! Check out patreon.com/aiinsideshow to support the show.
ChatGPT Turns Three
Believe it or not, ChatGPT is now three years old. It launched on November 30, 2022 as a research demo, and it’s interesting to note that initially OpenAI staff was unimpressed and it was possible that it be shelved for specialized tools. But its release netted 1 million users within five days, and now that sits around 800 million weekly active users btw.
Three years later, the technology landscape is drastically altered. No one likes the emdash anymore. And Will Smith eating spaghetti doesn’t look like an alternate alien universe anymore.
That doesn’t mean OpenAI is on an easy path. CEO Sam Altman sent a memo to staff, according to the Information, where he calls a “Code Red,” says that Gemini is a threat, and also that Anthropic is gaining serious momentum in the enterprise. That puts a pause on efforts like Ads, Pulse, and shopping and health agents, refocusing on ChatGPT’s speed, personalization, and reliability.
OpenAI is facing a lot right now: slowdown in engagement, less benchmark dominance, and investor concerns around its valuation and investments, along with questions about whether its focus on infrastructure buildout can pay off. Also remember that a few years back, Google had its own code red thanks to the advancements of ChatGPT, and Anthropic recently shared it has 300k+ business customers, up from less than 1k two years ago.
Apple AI Leadership Shift
Apple AI chief John Giannandrea will retire next spring, shifting to an advisory role until then. Former MS corporate VP for AI and longtime Google Gemini engineering leader Amar Subramanya will step in as VP of AI, and he now reports to Craig Federighi. He takes charge of Apple’s foundation models, ML research, and AI safety. The remaining pieces of Giannandrea’s organization will be reassigned to Sabih Khan and Eddy Cue.
All of this comes on the heels of a few key AI misses, including the failure to ship the heavily marketed “Apple Intelligence” Siri upgrade for iOS 18 and iPhone 16, and a slew of Apple AI team defections. Apple is also reportedly working on a partnership with Google to integrate Gemini into its Siri offering next year.
At the same time, Giannandrea has had a fruitful career and is getting older, so retirement just might be timed right for his life. It’s hard to know how much of this is a direct action when faced with Apple’s recent AI misses, but it’s at least a notable data point.
AWS AI Chip Announcements
This week at AWS re:Invent, Amazon made some big announcements, including Trainium3, a 3nm AI training chip packaged in a new Trainium3 UltraServer system. It combines custom compute and networking for training and inference, and Amazon says it’ll be 4x faster than the current system.
AWS says Anthropic, LLM Karakuri, Splash Music, and Decart have been testing Trainium3 systems, and they have cut inference costs up to 50 percent. AWS also revealed Trainium4, now in development, which will support Nvidia’s NVLink Fusion, meaning it interoperates with Nvidia GPUs. That combines Amazon’s lower cost server racks with Nvidia’s accelerators, and it means companies already invested in Nvidia can more easily shift their workloads onto Amazon’s cloud.
Amazon also announced Nova Forge, a new AWS service for enterprises who want to build their own “frontier-class” models. They can start from internal Nova checkpoints and mix in their own proprietary data across pre, mid, and post training stages. It’s a more powerful alternative to fine-tuning or RAG, at least that’s how it’s being positioned, and customers can bake knowledge directly into the base model.
Google TPUv7 Push
Then there’s Google and its plan to widen its TPU business that we discussed on the show last week. Google’s TPUv7 is looking to be a serious competitor to Nvidia, particularly for hyperscalers and frontier labs. This according to a report from Semi-Analysis that looks at Google’s TPUs vs Nvidia’s GPUs. They say large buyers could lower cost by 20-50% per useful flop with Google TPUs. Anthropic has a 1 million TPUv7 commitment. Around 400k TPUs bought directly and 600k rented. Their estimated training cost per PFLOP is around 50-60% lower than Nvidia GB300. In other words, this move could seriously restructure the landscape that Nvidia has enjoyed to date.
AI in Prisons
MIT Technology Review has an interesting look at how AI systems are being used inside prisons and monitoring phone calls to and from inmates. Specifically, telecom provider Securus mines years of recorded calls, video chats, texts, and emails to flag “contemplated” crimes in real time. When detected, it feeds the information to human reviewers for investigation. They are building state-specific models trained on these massive data sets to do this, and they say they’ve disrupted human trafficking, gang coordination, and contraband smuggling.
Yes, inmates are notified that their calls are being recorded, but they might not know their calls are being used to train AI models. Some advocates say this is coercive consent, because without these calls they have no contact with the world outside the prison walls.
James Cameron on GenAI
James Cameron says genAI can be a useful way to cut VFX costs in film, but he says the use of AI to replace actors is “horrifying.” He argues that while the market will be flooded with sub-par derivative AI-generated mashups, human creativity will feel more “sacred” and carry with it a premium. He says the fact that AI is a mere remix of what came before it means human work like his will never be obsolete.
The NYT has an article that looks at a different kind of human replacement. Silicon Valley startups are cloning major sites like United.com, Amazon, Airbnb, and Gmail into legally gray “shadow” platforms so agents can learn to click, type, and basically transact at scale without being blocked by bot protections or limited by a reduction in public training data.
Deepseek’s New Models
Deepseek, the Chinese “thorn in my side” startup for Silicon Valley AI startups, released two open-weight 685B parameter models, V3.2 and V3.20 Speciale. They supposedly match or beat GPT-5 and Gemini 3.0 Pro on elite math, coding, and reasoning benchmarks. They also cut long-context inference cost by around 70 percent, add “thinking in tool-use” agent capabilities, and make it all free under an MIT license. Take that, US competitors!
ChatGPT Shopping Research
ChatGPT is introducing Shopping Research, a new guided shopping mode. It integrates a conversational approach to clarify questions, pull in and compare current prices, specs, reviews, and availability from retail sites. It then takes all that and packages it into a buyers guide with tradeoffs, top options, and recommendations tailored to budget. Eventually, instant checkout too.
Runway Gen-4.5
Runway announced its new Gen-4.5 video model, which is aimed squarely at high-end creators and studios. It’s doing very well on the Artificial Analysis leaderboard, outperforming all competing models. It offers improved visual fidelity (of course), creative control, and physical accuracy, and it has been sped up by fresh training and inference tricks on NVIDIA Hopper and Blackwell GPUs. It will be rolling out broadly across its subscription tiers in the coming days.
Rise of AI Degrees
And I bet you couldn’t have guessed by now, but AI degrees are kind of a big deal. According to the NYT, colleges are rapidly rebranding and reshaping computer science into specialized AI degrees. Students are, of course, looking for ways to future proof themselves and show that their time in college isn’t all for naught if the AI comes along and replaces their hard work in the marketplace. At MIT, “artificial intelligence and decision-making” is now the second most popular undergraduate major.
A Deep Dive into 360 Reviews with My360.ai
Finally, in case you missed it, I hung out with Jared Goralnick on the AI Inside YouTube channel, the Founder and CEO of My360.ai, an AI service that’s all about 360 performance reviews and a cool system to bring them into the chatbot age. We demo the service and talk about it at length.
HUGE thank you to Executive Producers on the Patreon: DrDew, Jeffrey Marraccini, Radio Asheville 103.7, Dante St James, Bono De Rick, Jason Neiffer, Jason Brady, Anthony Downs, Mark Starcher, AND Karsten Samaschke!!
Thank you for watching and reading! See you next week.