Major Changes in AI and Mac Technology  🔗 **Source:** [Major Changes in AI and Mac Technology](https://share.snipd.com/episode/cd2f25f5-4fb7-459e-bf63-47b9ae0c78e5) ## Highlights - **Claude Unifies Memory Across Devices** - Anthropic merged Claude Chat and Claude Cowork memories into one system across web, desktop, and mobile. - Jaeden Schafer says the change fixes disconnected workflows when switching devices and preserves continuity across Claude products. Transcript: Jaeden Schafer Let's talk about what's going on with Anthropic. They just announced that Claude Chat and Claude Cowork now share one single memory system, and this is across the web, desktop, and mobile. Hallelujah. I think for anyone that is a heavy user of Claude, you'll know that they have Claude Chat, Claude Cowork, Claude Code, and they have desktop, they have mobile. And I, for one, so many times will be working on a project over on Claude, you know, co-work on my computer, for example. I hop on my phone because I'm away from the office to get something done and I'm trying to remember what I was talking about. Or I get a new piece of information I want to put into a chat thread. And it's a completely different system over on my phone than it was on my desktop. So everything feels so disconnected. They're doing some work to help bring that together and to get rid of that issue. So you're now going to be able to read, edit, or delete anything Claude has stored on any topic. There's going to be a new memory inspector that lets you do that. ([Time 0:01:56](https://share.snipd.com/snip/59fafbe6-2470-45b5-a1fc-48d93e10e05c)) - **Use The Memory Inspector And Toggle Sensitive Topics Off** - Memory is on by default but users can read, edit, or delete stored items with a new memory inspector and toggle sensitive categories off. - Jaeden recommends updating the iOS/Android app and toggling sensitive topics like health or religion off to avoid unwanted recall. Transcript: Jaeden Schafer So you're now going to be able to read, edit, or delete anything Claude has stored on any topic. There's going to be a new memory inspector that lets you do that. Memory is on by default. So any free Pro Max Plan users, it's for iOS and Android, and they need the latest. So you've got to go actually update your app if you want to get that, by the way, just pro tip. Any sensitive categories. So they've had things like health, religion, politics, gender identity, race. All of those are excluded unless you go and toggle on include sensitive topics in the memory. I think this is actually kind of a cool feature. Something that my wife always complained about with chat GPT is that, you know, she's like always worried about asking about some sort of health problem. Because in the future, when she asked it about something, it'll be like, remember when you said you had a rash on your elbow? Well, blah, blah, blah. And it like uses that as context. And she's like, no, like forget this thing ever happened. She never had a rash on her elbow. That's a good example. So I just think at the end of the day, you have certain topics that you don't want it to always remember and refer back to. And in fact, I know people that have, they'll specifically use something like perplexity or a completely different service that's not attached to anything else you work on day to day Whenever they have like sensitive questions that they want to go ask. So anyways, I think this is also another great step towards keeping people on platform for as much as possible. Claude will never store your government ID, your social security number, your criminal history, or your immigration status. ([Time 0:02:44](https://share.snipd.com/snip/9e29ac60-78b0-40fa-a7fa-4f7010893a35)) - **Personal Fear Of Persistent Health Memory** - Jaeden shares his wife's concern about chat history: she fears an AI will later recall a medical question like a rash and reference it unexpectedly. - This personal example shows why people prefer not to have health or private topics remembered automatically. Transcript: Jaeden Schafer Any sensitive categories. So they've had things like health, religion, politics, gender identity, race. All of those are excluded unless you go and toggle on include sensitive topics in the memory. I think this is actually kind of a cool feature. Something that my wife always complained about with chat GPT is that, you know, she's like always worried about asking about some sort of health problem. Because in the future, when she asked it about something, it'll be like, remember when you said you had a rash on your elbow? Well, blah, blah, blah. And it like uses that as context. And she's like, no, like forget this thing ever happened. She never had a rash on her elbow. ([Time 0:03:03](https://share.snipd.com/snip/14578bae-5310-460e-bb7b-f126d56fcc00)) - **Apple Pushes On-Device AI With New M6 And M5 Ultra** - Apple launched two-nanometer M6 Mac mini and a dual-die M5 Ultra Mac Studio aimed at on-device AI with up to 512 GB unified memory and 1.2 TB/s bandwidth. - Jaeden highlights specs: M6 12-core CPU, M5 Ultra equals two M5 maxes on one die with 36 CPU and 80 GPU cores. Transcript: Jaeden Schafer And I think all of this is helping make Claude a much more useful tool. And the goal here is that this is going to be better for local AI happening you know local AI generation happening right on your computer Apple's two nanometer m6 and dual die m5 ultra Push up to 512 gigabytes of unified memory and 1.2 terabytes of bandwidth and They're really just hoping developers running local LLMs are the ones that are going to be using this the Mac the m6 mac mini it's going to be 899 it's going to start with 16 gigs of memory the m5 ultra mac studio is 5 500 and it will scale to 512 gigabytes of unified memory the m5 ultra hits 1.2 Terabytes per second of unified memory bandwidth with 36 CPU cores and 80 GPU cores. So basically it's two M5 maxes and it's all happening on one die, which is really impressive. M6 is Apple's first two nanometer chip. It has a 12 core CPU and Apple says that it's up to 40% faster in multi-thread workloads than the M4. ([Time 0:04:10](https://share.snipd.com/snip/e1be1ffb-f0fb-40cf-a455-0c902d2db1ea)) - **macOS 26.2 Enables Distributed Local Model Inference** - macOS 26.2 enabled distributed MLX inference over Thunderbolt 5, which pairs with Apple's new chips for scalable local model workloads. - Jaeden notes this OS feature is what makes multi-device local ML acceleration feasible now. Transcript: Jaeden Schafer M6 is Apple's first two nanometer chip. It has a 12 core CPU and Apple says that it's up to 40% faster in multi-thread workloads than the M4. So they're saying this is a pretty big boost from the M4. Pre-orders are just opened up for this. So I'll be curious to see what their sales look like in the future. This is all happening. So they did this whole kind of refresh. This is all after macOS 26.2 came out in December of last year, which basically enabled distributed MLX inference over Thunderbolt 5. So this is something that's only really possible with some of the recent updates they have made. ([Time 0:05:14](https://share.snipd.com/snip/b8416e0a-7dd1-4a10-8faa-ce5649338b71)) - **Stability AI Raises Strategic Funding From Media Giants** - Stability AI raised $76M in a Series B from Universal, Sony, Warner, and EA and is turning investors into strategic co-development partners. - Jaeden explains deals include content licensing co-development with Universal Music and EA to expand creative tools and services. Transcript: Jaeden Schafer All right, Stability AI has raised $76 million in a Series B from Universal, Sony, Warner, and EA, and their cap table now doubles as a content licensing map, in my opinion. I mean, that's why they're raising a lot of this money. Now that they have EA and a bunch of other people on this, it's, you know, it's beyond just getting money. They're actually trying to get money from people that are strategically going to help their company. And with all of these different investors, they signed co-development deals with Universal Music and with EA. And they actually did that back in October and Warner Music in November. So they already have these kind of like co-development deals where they're working on new models with those companies. ([Time 0:05:54](https://share.snipd.com/snip/a86d72c5-86f5-40d1-8216-9cf5b3a5c72d)) - **Radiologists Are Growing Not Shrinking** - Radiologist headcount is growing about 26% despite predictions they would be obsolete, and three quarters of FDA-cleared AI medical devices target radiology. - Jaeden cites 1,400 FDA-cleared AI devices and rising salaries plus projected headcount growth over decades. Transcript: Jaeden Schafer There's more and more radiologists. I think the growth in that particular field is up 26%. And this is interesting because I believe in 2016, the quote unquote, godfather of AI, you've probably heard of him throughout the years, he got on stage at a conference and said that, You know, I think it was Jeffrey Hinton, he predicted 10 years ago, that radiologists would be obsolete by 2021. So I mean, he was obviously ahead of his time. But three quarters of the FDA's 1400 cleared AI medical devices right now are in are targetingologist field, and yet they're up 26%, and I was looking at something that was showing their Salaries were also up by a ton, and it's a really healthy career path, and a lot of people are in there. So like I mentioned, the FDA has cleared about 1,400 AI-enabled medical devices as of early this year, three quarters of those are in radiology. So that's like the most popular place that these the AI enabled medical devices are inside of radiologist red headcount is also projected to grow over the next three decades. ([Time 0:07:48](https://share.snipd.com/snip/523fe271-93c2-4df2-9a22-05200379e364)) - **$2.5B Supermicro AI Server Smuggling Case** - Taiwan indicted nine people including an NVIDIA manager over a $2.5B Supermicro scheme that forged documents to ship restricted B300 AI servers to China. - Jaeden details 74 servers reached Chinese customers, 56 were blocked, and arrests included a Supermicro co-founder. Transcript: Jaeden Schafer Taiwan has just indicted an NVIDIA manager in a $2.5 billion super micro AI server smuggling scheme. It sounds like something straight out of a James Bond movie, but there was nine people charged with forging documents to send 74 restricted B300 servers to Chinese customers. This is, of course, as US export controls were in place, and so they were kind of going in defiance of all of those. In the whole scheme, they had about 74 of the servers that actually got to Chinese customers, and there was 56 more that were blocked when Taiwanese customs flagged some of the irregularities That were happening. All of this is a massive deal. The scheme was $2.5 billion. And it was, of course, all of these restricted AI servers to China. And it's been happening since 2024. So this was first exposed in March of this year, they had some arrests of a Supermicro co-founder, Wally Liao, and the probe has targeted at least 12 different locations. Supermicro says that it has fired several employees connected to the scheme. ([Time 0:10:52](https://share.snipd.com/snip/75a9970d-8677-495e-80be-900df7ea65dc)) - **Data Center Expansion Creates Export Control Loopholes** - U.S. lawmakers are considering the Remote Access Security Act as Southeast Asia scales data centers from 2 to 31, creating a cloud loophole for restricted hardware access. - Jaeden warns rented data centers could let firms run models on NVIDIA hardware despite export controls, pressuring China to build domestic chips. Transcript: Jaeden Schafer U.S. Lawmakers are weighing the Remote Access Security Act to close a cloud loophole as Southeast Asia data centers scale from two today to 31 planned data centers. So, I mean, if they just build a bunch of data centers and Chinese companies can just go rent the data centers and run all of the models. I mean, they basically are getting access to all of the NVIDIA hardware there as well. So at the end of the day, I think a lot of these Chinese firms are quite savvy and they'll figure out a way to China's, you know, directly trying to figure out a way to build their own chips As competitors, their own GPUs and services. And they have a lot. Of course, NVIDIA, I mean, it's hard to beat and catch up with them. But I think China's putting a lot of pressure on because they know this is a really important race to win. And so there's a lot of loopholes they're looking at getting around. ([Time 0:12:06](https://share.snipd.com/snip/6565815f-b169-466b-ac4d-2f41cbcd9409))